The roadmap concluded with a definitive DNA of firms positioned to win: an exceptional founding team with domain expertise, AI-driven platforms that integrate client infra, pricing closer to usage and outcomes, ability to show business impact quickly, and margins upwards of 60%.

On a different timeline, Sidu Ponnappa was already living in the eye of this ambition — with all the elements of that DNA. Sidu’s founding of realfast.ai, an AI-native IT services firm that clocked $3M revenue in H1 of 2026, happened right before the drop of ChatGPT.

He had built and sold an IT services firm (C42) once before, gathered a brilliant founding team, their AI-platform (Exo) had cracked scaled delivery… operating with a ROI-first pricing, delivering results under a quarter, at close to a 60% margin.

All based on a belief that GPT 3.5 was going to bring in an impending 2.0 for the IT industry; and as we have since seen, he was on to something.

We sat down with Sidu to dig into the details. What emerged is a list of lessons and guiding principles for anyone building or backing an IT challenger firm.

Table of contents:

• The wedge: picking up pieces and smallest valuable projects

• Gathering a team with 3 specific skillsets

• Why pricing must be molded around the customer

• The new, not absent, form of productizing at a neofirm

• 5 notes for challenger firm founders

#1: The wedge: picking up pieces and smallest valuable projects

At an IT firm, every org served has a different shape of work, often built around a customer’s org chart, processes, and complexities. A wedge, then, can’t be something standard like an NDA or an audit report. Here are Sidu’s heuristics for a wedge in IT, from what worked for them:

The first one is pain that is already felt. “What we do is we pick up the pieces. We will not touch a deal where the customer has not tried and failed.” Sidu says.

He shared two examples to illustrate this: first, where they picked up a module that a previous vendor had missed the deadline on, and second, where they were asked to rip out a legacy SaaS tool everyone was sick of. Both cases were easier sells because the pain was already felt.

Another wedge to leverage is efficiency, passing on the gains of AI to their customers. This worked because, in an industry that is 95% snakeoil, as Sidu put it: “no relationship [with existing vendors] survives someone coming in offering twice as fast, twice as good.”

This focus on efficiency is also an internal mandate. “For every dollar they spend with you, can they visibly and obviously see that they’re getting $2 back?” Sidu said, with an emphasis on visibly and obviously.

Beyond these, another heuristic they followed was: “starting with the smallest valuable project, one that moves the needle.” It has since proven to be a better approach than negotiating a big scope upfront.

Sidu and team treat it as an entry point, one that if they deliver on, inevitably leads to expansion because of the sheer scale of demand for IT work. It has also worked for them.

Sidu shared how one account started as a single manager’s discretionary spend, and soon had several managers stack budgets making it a mid-six-figure deal — entirely through word of mouth inside the org.

He calls it the “customers feeling the $2”.

#2: Gathering a team with 3 very specific skillsets

Delivering at this bar of quality and speed needed a team Sidu hadn’t seen fully assembled before. Our neofirms piece had surfaced a composition of SMEs, builders, and amplifiers. Sidu makes a few more additions specific to selling IT services:

• Management consulting skillset — navigating top management, customer politics, aligning stakeholders,

• Consumer big-tech product execution, and

• Services delivery management.

The bend for consumer product execution is curious. Sidu shares that he has seen consumer product folks bring a sense of metrics-obsession he hasn’t seen anywhere else, which isn’t just a good-to-have but a necessity for delivering great services. The phrase he used was they “live or die by a number they see every day — usage, retention, whether anyone came back”.

On the builder front, Sidu has a high bar for what makes someone great: “Two weeks of velocity is easy. Sustaining it past two weeks is the brutal part.” At the base of these filters is Sidu’s search for people who operate like kaarigars, or artisans, and carry a strong sense of pride in their work.

It’s a little nostalgic almost, and reminded us of Sidu’s interview with the Gojek leadership where they discussed 10X engineers and Sidu’s bar for talent as the then Head of Global Tech Recruitment at Gojek.

#3: Why pricing needs to be molded around the customer

realfast.ai doesn’t operate on one standard pricing model. Outcome-based, fixed-price, or hourly — Sidu calls them implementation details that only matter if they align with how the customer wants to pay and what you must charge to cover your costs.

In a recent X article, Sidu called it a fitness function for anyone running a services company: revenue, immediately followed by margins.

He closed mincing no words:

“[In services] Your revenue is a percentage of your customer’s revenue. How does it matter whether the customer pays you that percentage through billable hours or through fixed price fixed scope or through ownership of metrics, input metrics or ownership of output metrics or the ownership of outcomes?”

He’s candid that this is a fairly new motion. So, to keep it from becoming endlessly malleable, Sidu maintains a steady north star: “My objective is I have to run my projects at 60% gross margin.” He goes on to add that services have two core cost units: cost of tokens and avg. cost per billable hour, and “everything else is packaging”.

“The stress test is: you go to a customer with a model, the customer says no, I don’t do this. Are you going to tell the customer to get lost? So who gives a sh*t what I think? We have to give a sh*t what they think.”

Which is why, at realfast.ai they sell how the buyer wants to buy. The messy details of the math behind the margins is the founders’ job.

#4: The new, not absent, form of productizing at a neofirm

If you go to realfast’s website you will find mentions of their Exo product, an interesting juxtaposition for a services company. So I asked, what does it do? The story wasn’t what I was expecting.

Exo is entirely internal. Sidu calls it their alpha.

It’s on the website because one, customers want to know how a new kind of services firm will deliver on the tall promises it is making. And two, the industries they sell to have stern compliance issues and the buyers want a high level of confidence in their vendor’s operations.

Exo is a product, but in an entirely new outfit. It is the internal tooling carrying the agent harness, logging, and compliance controls that allows a team of 5 engineers on a pod to deliver at superhuman speed and capabilities. Constantly made better atop new model capabilities.

If it’s so good, why not sell it? We asked, nursing a SaaS hangover. “Why would I sell my alpha? Why wouldn’t Uber sell its fleet management solution? Same question. It’s tailor-made alpha. Why would Google not sell its search algorithm?” Sidu replied.

Beyond wanting to protect the alpha, the fit doesn’t work either: “Remember when all our clothes were tailor-made? One person’s clothes wouldn’t fit another. That’s software. SaaS is ready-made clothes. We’re talking about something tailor-made — nobody else can wear it unless they’re the same size and shape.”

A reminder of how services, by its fundamental gravitational force, is custom, and that fighting it is a fool’s errand.

This conviction and clarity on what to productise, what not to, and whether that product stays internal or gets exposed wasn’t easy to come by. One of the most painful experiences, Sidu recalls, is spending two and a half years trying to productize in Salesforce.

“In 2023 and 2024, realfast built a unit-test-writing harness for Salesforce Apex code. We spent 4 months on solving accurate diffs, but then GPT-4o shipped and solved the problem outright.” It taught them one core lesson: a services firm should never build a product ahead of demand, no billable project = no platform work.

Sidu’s another guiding principle for services is: “When the models get better, your TAM should go up.” You’ll often find him tweeting news and updates with a cheeky “TAM just went up.”

#5: Five lessons to other challenger firm founders

“Agentic,” “harness,” “product” — Sidu calls them useful shorthands that often get mishandled. His advice to founders building in services is to use the buzzwords carefully, and focus on the real thing underneath.

“Strip the labels”, he argues, “and things get simpler, customers have a problem and a budget, and they want a dollar to come back as two, visibly, fast.” The ROI obsession, reiterated.

He shared five more concrete pieces of advice in our chat and in this X article:

First: beware sales-market fit without product-market fit. An IT services deal is easy to close on a good relationship and a confident pitch, but whether the client’s team actually adopts what you’ve built and keeps running it the material problem.

Second: sell speed and quality only to the people inside a client org for whom it’s urgent right now. Most of an enterprise doesn’t want humanly-impossible delivery — in IT specifically, plenty of stakeholders below the top read speed as risk, not as a benefit, and will resist a vendor moving faster than their own governance.

Third, and this is the one he returns to most: don’t lock into a vertical or a thesis ahead of demand. “Run a greedy algo: take the best-paying work in front of you, extract what compounds, repeat.”

Fourth: whether you are VC-funded or bootstrapped, run the numbers before you take a customer’s money. His formula is straightforward: blended build rate across your team, plus average cost per token, is your cost base. Mark it up 50 to 60 percent and sell at that. “There is no way a services company makes a loss unless you’re screwing it up.”

Fifth: pivot, pivot, pivot.

“Pivot intentionally every time there is a step function change in AI capabilities. A step change in the underlying tech requires a step change in the strategy to take advantage of it.

#6: Looking ahead

RealFast has grown 26x since the start of this year and closed $3M in revenue in H1 of 2026 alone. Still early numbers, he says, but promising.

In parallel, the pressure on the incumbents is real. The NIFTY IT saw one of its sharpest corrections after the announcement of the OpenAI Deployment Company, a new business that embeds engineers directly inside client organisations to redesign workflows and deploy AI at scale.

The market read it as a signal that AI companies were coming for the transformation work and did a selloff; it has partially recovered since.

Whether this pressure gives realfast more room, or the IT giants adapt fast, or the labs swoop in, or all those worlds end up blazing in a competitive rhythm, is yet to be seen. As far as Sidu is concerned, he is building, unfearing and focused: “There’s so much software to be built.”

After all, he didn’t start realfast on a back of a bold thesis, but a simpler and sturdier belief in the step change: that the oldest tradeoff in the IT, “cheaper, better, faster, pick two”, wasn’t true anymore, and those who find a way to give their customers this superpower will win.

Visibly. Obviously.

Thanks a bunch to Sidu for taking out the time to talk to us! P.S. realfast.ai is backed by PeakXV and RTP Global, Bessemer’s thesis stands independent of it.

A narrative violation. This was typically the moment where you would build a tool to sell to vendors; not become a vendor yourself. But somehow the latter is exactly what is happening today, giving birth to firms that are part industry experts and part AI, or as they have now been named: neofirms.

Julien Bek of Sequoia has a thesis for this turning tide. He captured it in his March 2026 essay, Services: The New Software, where he wrote that the next $1T company will be “a software company masquerading as a services firm.”

Implying that it would sell work rather than tools, restructure all kinds of service industries, and do so while breaking away from the poor-margin, high-headcount past of services and the limited-TAM, limited-agency past of SaaS.

He also argued that this is one of the better, fortified ways to build today, while ensuring you don’t become a casualty in the next vertical expansion maneuver from the big labs: “If you sell the tool, you’re in a race against the model. But if you sell the work, every improvement in the model makes your service faster, cheaper, and harder to compete with.”

Crosby is another such neofirm, in legal. And so is Quanta, in accounting. But that’s not all, the recent cohorts of YC alone feature startups building neofirms for something as broad as finance audits to something as niche as government affairs.

While the theses abound, there isn’t a synthesis of the founders’ thinking and hunches, hard-won and tested in practice. In this edition we will collate one, surfacing useful patterns and mechanics, and end with one big open question.

To ground these explorations, we will refer to Crosby, Quanta, and Harper, 3 neofirms with great ambitions and traction. We will also take notes from Emergence Capital’s incredibly insightful playbook on the matter along with other VC theses such as Bessemer, Foundation Capital, and of course Sequoia.

P.S. Even if you aren’t building a neofirm, there’s a lot to be learnt here because the central expectation from all enterprise software is increasingly shifting towards delivering end-to-end work.

Table of contents:

• Work spend : tool spend — the gap and the opportunity

• Zeroing-in on an industry

• Finding a wedge

• Ways neofirms need to operate differently

• Proving your PMF

• Where’s the defensibility?

• Will this work?

#1: Work spend : tool spend — the gap and the opportunity

The central premise underlining the services thesis is that AI lets you sell work into a much larger TAM, given it now includes the TAM of labour and services. Julien illustrated this using the 1:6 ratio in his essay — stating that organizations spend $6 on services for every $1 on tools.

In the categories we are looking at, the gap is even more extreme.

Legal:
US legal tech spend sits in the $7–8B range on the narrower end, most of it going to Westlaw, Lexis, Relativity, and Clio. Crosby’s co-founder, Ryan Daniels, puts US legal services at roughly $400 billion. So, per back-of-the-napkin math, about $1 of software for every $50 of legal work. Ryan’s estimate of the legal tech market is higher, closer to $20B, which would still leave a 1:20 gap.

Accounting:
US accounting software spend was around $6.4B in 2025, most of it on QuickBooks, Xero, NetSuite, and Sage Intacct. US accounting services the same year was roughly $164B. So, about $1 of software for every $25 of accounting work.

Insurance brokerage:
US agency management software was around $1.7B in 2025, split mostly between Vertafore, Applied Systems, and EZLynx. US brokerage revenue the same year was $261.7B. About $1 of software for every $150 of brokered work.

In practice, all 3 of these industries beat the 1:6 tool:work estimate.

A fair question that needs addressing before we go any further is why wouldn’t the traditional services disrupt themselves?

Ryan has a theory: “When you build a really great labour and services business that’s working, there’s not a lot of incentive to innovate and disrupt yourself, especially when your margins are pretty thin.”

That said, Kirkland & Ellis, the largest law firm in the world by revenue, is already spending $500mn building its own AI technology but there are no public wins yet.

As far as these upstarts and their new model is concerned, early demand signals seem encouraging. Today:

Crosby has a hundred+ clients including Cursor, Ramp, and Clay, and has reviewed about 13,000 contracts; where a single, unassisted NDA review takes 58 mins, Crosby claims to do it in 4 min 53 sec. Their revenue is up ~400% since October, with $85.8 million raised.

Quanta has hundreds of customers, with annual contracts of $15,000 to $25,000, and their close time is under a day against the industry standard of four to eight weeks. They have raised $20 million so far.

Harper has six to seven thousand customers, is past $5 million in ARR with 32 employees and $46.8 million raised. While a traditional broker takes weeks for a complex insurance, Harper does it in 24-48 hrs.

#2: Zeroing-in on an industry

Start with a list of service industries and you’ll drown; all of them are large, most are badly served, and they have a way of looking close to identical on paper. That’s not how the founders of Crosby, Quanta, and Harper made their respective picks.

Instead, each of them started with a space they could already get inside, and then went and looked at the work being done by hand. Crosby’s Ryan Daniels has a name for it: Genchi Genbutsu. From the Toyota production system, meaning ‘go see for yourself’ as documented in this brilliant New Ontologies profile.

Across his story and others, a heuristic for zeroing-in emerges:

1. Leverage your proximity and credibility:

Services run on trust, and a neofirm is no different. Given that, the founding team’s background can be especially critical.

Emergence Capital puts it succinctly: “In traditional SaaS, you’re selling a product. In AI-native services, you’re selling yourself. Domain credibility isn’t just important; it’s existential.”

Ryan trained as a lawyer at Stanford, practiced at Cooley then went in-house as the only legal person at a startup growing from 10 to 100 people, he had the credentials. Further, as someone who spent most of his time on contracts and MSAs, he knew the pains, the processes, and the bottlenecks incredibly well.

Quanta’s founder, Helen Hastings, too, was able to find early leads for research and demos because of warm introductions. She says that her experience at Affirm was why people said yes.

"Legal services, I think economists call it a credence good, which means that you only know how good it is after you’ve experienced it and consumed it, and you need an expert to actually tell you the quality. A lay person, or even not a lay person — a sophisticated CEO — often doesn’t know, by definition, how good their legal work is.” — Ryan on why trust is paramount in services.

A lack of credibility and proximity can be fixed, though. Jake Saper from Emergence adds two conditions that break this constraint: the founders go as deep as humanly possible on the service, and/or they hire senior domain experts (with credibility and network).

2. Go where the demand is overwhelming:

Dakotah and Tushar didn’t have an insurance background per se, but Dakotah’s childhood gave him exposure to insurance because of his family business. The domain fell on his radar per chance.

What made him commit to it was a strong signal of unfulfilled demand. During research, they had applied to hundreds of insurance sales jobs on Indeed and heard back within 10 minutes. They saw that there was a dearth of people who could sell insurances; all while many small and medium businesses were retiring.

#3: Finding a wedge

Much like software, once you have picked an industry, you’ll need a wedge to get inside — that or you’ll be boiling the ocean.

Here are the markers of a strong wedge in services:

1. Start with what is outsourced:

Julien’s base characteristic of a good wedge in services is outsourced work:

“If a task is already outsourced, it tells you three things. One, the company has accepted that this work can be done externally. Two, there’s an existing budget line that can be substituted cleanly. Three, the buyer is already purchasing an outcome. Replacing an outsourcing contract with an AI-native services provider is a vendor swap. Replacing headcount is a reorg.”

It is indisputable that the TAM for an autopilot is all of the labour spend in a category, but the easiest place to start is where outsourcing already exists. This, then, isn’t a weakening of the ambition, but a strategic entry point. Julien urges founders to think of the outsourced task as the wedge, and the insourced work as the long-term TAM.

And it checks out with the approaches of the three startups we are focusing on. Contract review sat with outside counsel, bookkeeping sat with an outsourced firm or offshore team, and commercial insurance was always brokered.

Helen distills the why: “The service already existed and customers already disliked it, so there was no category to create and no user education to fund.”.

Go in. Rip and replace.

2. Pick work that is more intelligence-heavy than judgement-heavy:

Julien’s second criterion is to pick work that’s more intelligence-heavy than judgement-heavy, where intelligence refers to work with defined rules. Ryan also felt this in legal: “An enormous amount of legal work is formulaic, because contracts are full of if-then conditional statements.”

Ryan and Dakotah share two other interesting details to notice:

• Slightly too complex for AI alone. Ryan says that AI is already good at NDAs and offer letters, which is why they aren’t a wedge anymore. Master Services Agreements (MSAs) at 12 or 13 pages is. An NDA is usually one negotiated term, but an MSA is a full commercial relationship with details like payment terms, liability caps, IP, data handling, etc. It’s complex enough that AI alone gets it wrong, and simple enough in structure that AI plus a lawyer is genuinely faster than a lawyer alone.
  

• Look beyond the commoditized level of service. Dakotah quickly ruled out commoditized insurance where people just wanted the cheapest price and no intermediation was needed. What he wanted was regulation that forced an intermediary, and low standardization to warrant a service.

“There’s a lot of things that happen in insurance that result in a $10,000 payday for revenue that may have 50 steps. If you’re able to reduce 50 steps to five steps, times all of the customers that you have in that type of revenue, you can build a business that’s a massive revenue business with a crazy margin.” — Dakotah Rice, Harper

3. Judgment does not disappear, it moves above the abstraction

“Most lawyers would just say it’s impossible [to automate legal work]. Most technologists would say this is all automatable. And the truth is somewhere in between.” — Ryan Daniels, Crosby

If you go looking for a purely intelligence-shaped wedge, you may not find one.

Dakotah is of the opinion that it is impossible to disconnect judgement and intelligence entirely. Knowing this, at Harper, agents handle collection and the back-and-forth, and human judgment concentrates at the point of actually selling the policy.

Ryan also shares that there’s a paint-by-numbers element to law that is hard to even identify. He describes the difficulty of taking a hundred-step workflow and working out that steps 17 to 30 can be automated while 36 to 41 can’t.

Given this, the more grounded question on wedges isn’t what is purely intelligence but where in the workflow the judgment still remains difficult to codify. If it sits at the end, near the close, you can automate the long middle and keep a human on the last mile. (We will get into the team structure needed to enable this below.)

#4: How neofirms operate differently

A neofirm has quirks of its own. Some tensions that resemble how services get sold and built, some that exert pressures on software’s way of roadmapping and operating, and some that are entirely unique to it; here are a few, emerging mechanics:

1. Take the liability seriously

“We have malpractice insurance. We take liability for all the work we do. I don’t think we exist as a business if we don’t do that.” — Ryan Daniels, Crosby

Some say that in addition to the service, organizations pay for “a throat to choke”. Underneath that universally deployed, nonetheless disturbing analogy is a simpler truth: if you sell a service, you also have to take on liability, often of a magnitude that’s directly proportional to the criticality of your service.

Crosby knew legal is high stakes, so they went and incorporated a law firm, employing barred attorneys and carrying its own malpractice insurance from the beginning.

Harper had to comply because brokerage is licensed state by state. Quanta made a different call, offering back-office bookkeeping and controller work while stopping short of anything requiring a public accountancy license.

2. Own the implementation and the outcome

“In SaaS, a customer buys your product and implements it. In AINS, you ARE the implementation. Delivery isn’t a support function; it’s the core of what you sell.” — Emergence Capital

This dramatically changes what “good enough” means. A SaaS company can ship an excellent product and deliver a mediocre implementation, and the customer will mostly blame themselves.

A neofirm can’t.

Harper, Crosby, and Quanta treat daily delivery as their primary engineering signal. And structure teams in ways that allow for that signal to reach the people writing code, fast. Crosby seats lawyers next to engineers, offset one by one, where evals land every few hours.

This emphasis on delivery makes the case for features that might otherwise seem unnecessary, but give users the controls that they need to feel confident in a new way of doing things. For instance, Braintrust’s (a Crosby customer) Chief of Staff, named cross-checking features, internal notes, and ETA metrics as what earned the migration, as shared here.

Most of these orgs also have specialized agents and processes for delivery.

Crosby has an internal paralegal agent that quickly routes every query to the best available lawyer so a customer doesn’t have to wait.

Harper’s equivalent is matching — a system that synthesizes across its entire network of wholesalers, MGAs, and carriers at once, weighing appetite and pricing patterns in a way no single broker’s personal relationships could match, so a business doesn’t wait on one person’s memory.

3. The roadmap tradeoff

“If a customer has a new feature request, we can’t say I got it on the roadmap for 6 months from now. If we hit an edge case, what should be a new feature, we are on the hook to handle it because we committed to doing all of the accounting.” — Helen Hastings, Quanta, at Beelieve’26

That’s what being an AI-enabled service is — you are on the hook for getting things done, and a second-order effect of being-on-the-hook is a new kind of pressure on your roadmap.

Every neofirm user comes in with custom requests, cumulatively it can pull you into all directions, without giving you a chance to say no. There are two reasons for this: 1) you promised the outcome, 2) the edge cases help you build coverage.

That considered, Ryan has a number for when you can start being more mindful:

“As a rule of thumb, we started to see patterns when we got to about 80 clients. Before that, we didn’t know what client preferences were rules or exceptions. That’s when we felt comfortable starting to say no to customer requests.”

4. A genuinely new team composition

Crosby’s early team can be split into 3 kinds:

• SMEs: lawyers and domain experts;

• builders: researchers and engineers;

• amplifiers: operationally minded technical people.

The third group is especially interesting.

Crosby’s ops team came partly from Scale — people with CS backgrounds who were not necessarily engineers. They would watch lawyers take 6 clicks where 1 would do, then ship or spec a quick fix.

Ryan calls this out as one of the things he wishes he had known earlier: ”For tech-enabled services businesses, having a cracked hit squad that’s technical but really operationally minded changes everything”, he adds.

This resembles neither a standard SaaS org structure nor services.

Emergence also shares two notes on the customer-facing team: “Industry experience is especially helpful in customer-facing team members; so much of this is about trust. You want the buyer to recognize the places your team has worked, ideally trusted brands from legacy service providers.”

They then specify that it is important to include someone who has a deep understanding of the service right in your sales process. This stops you from making false promises or committing to poor, unachievable timelines, which can impact trust.

5. A new kind of pricing model, following the work

AI has brought forth a new pricing philosophy that attempts at tying work, as closely as possible, to the price you pay — per outcome, per credit, per unit of work, etc. The neofirms are particularly well-suited for this evolution.

Emergence’s guidance is that discrete, well-bounded work should be priced per piece of work, while continuous, variable work suits a credits model.

Crosby charges per document, with Forbes reporting prices from roughly $250 to $1,000 per contract, depending on length. Ryan says that killing the billable hour was an intentional, early decision they made. They knew they might risk being margin-negative on some work and took it anyway, because the incentive to find a more productive way to work would then run through the whole company.

Harper follows the brokerage model, where revenue generally comes through carrier commissions rather than a separate SaaS-style fee. That means the customer’s buying motion does not change much, they still pay for insurance coverage, while Harper earns like a brokerage.

Quanta appears to price more like an accounting service than a SaaS product. Public pricing is not clearly listed, and Helen is candid that they are still working it out. Her current position is to price roughly in line with service incumbents and be the obvious choice on a side-by-side comparison.

Emergence thinks that most neofirms might need to start with the market norm.

Bessemer suggests anchoring to the legacy provider and undercutting it, which Ryan disagrees with, saying that undercutting might be read as low quality.

Helen also shares a challenge you might face with a neofirm’s work-based pricing: “People say, well with your competitors I’m paying for a team of humans, and for you I’m not. So why should you be charging me more?”

The honest advice is to use these as references and work out your market and your customers while expecting to land on something that is entirely different.

#5: Proving your PMF

“Real PMF requires proving you can scale non-linearly relative to your costs. To get there, your AI must drive measurable improvements in cost, quality, or speed, or ideally, all three.” — Emergence Capital

A given with software PMF that now applies to services.

Every neofirm has its own north-stars that it treats as leading indicators for PMF. From what we saw, three kinds surfaced: a service metric, an engineering metric, and a quality metric.

Crosby has two service metrics. First is TAT (Total Turnaround Time), measured from the moment a contract lands in the inbox to the moment it leaves, across every round of a negotiation.

The second metric is HURT or Human Review Time, tracked from the moment a lawyer opens a document to the moment they close it. Ryan believes that as HURT approaches zero, their margins will approach software margins.

Harper uses Turnaround Days, against a status quo where a customer spends a month calling broker after broker. Quanta’s is Close Time, against books arriving four to eight weeks late.

Under the engineering metrics, Tushar from Harper shared two that he looks at: throughput and reasoning. Throughput is a bottleneck hunt — whatever currently limits how many deals move through the system. Reasoning is capability, where the same throughput system handles progressively more complicated deals as reasoning models commoditize.

Quality is the third and the hardest to measure. Ryan says there’s a line of academic thinking going back decades that legal quality is simply unmeasurable: “Why do you pay that lawyer $2,000 an hour instead of that lawyer $1,000 an hour? Is he really two times better? It’s probably impossible to explain why.”

John Sarihan, co-founder and CTO, Crosby shares why quality matters so much: “One of the dangerous traps of language models today is they get to 90% for basically free... getting them to 99 or 99.99 is actually extremely difficult.” and that last mile improvement in the quality is what you get paid for.

Then there are more traditional metrics that matter from a purely business POV:

• Revenue per employee, ideally meaningfully above legacy providers.

• Per-customer margin breakdown, essential for identifying customers that are on a path to AI leverage vs those stuck in labor-intensive delivery.

• Gross margin with honest COGS; inference, API spend, and human-in-the-loop labour all inside it.

“You only truly have it when AI is doing a material share of the work at a high gross margin and delivering superior customer outcomes. Otherwise, you’ve built a good services firm financed with the wrong kind of capital.” — Emergence Capital

#6: Where’s the defensibility?

With the number of moving parts in AI, defensibility for anyone, including AI-native neofirms is an open question. Fortunately, we’re seeing some strong ones emerge:

1. Become the system of record.

This is one of the strongest structural moats available, and one that Quanta believes in as well, influencing their decision to build their own general ledger rather than sitting on top of QuickBooks or NetSuite.

Crosby, too, has an SoR called Bailiff — a central system that every contract, template, policy, and guideline flows through. What makes it a moat is what it does with volume: Crosby pools anonymized data across its client contracts to predict how a specific counterparty will react to specific terms, so redlining gets more proactive and negotiation rounds shrink as the dataset grows.

Emergence’s framing is that when you’re the infrastructure layer and the service layer at once, switching becomes doubly hard.

2. Build the proprietary data flywheel.

As Sequoia framed it recently, paraphrased, building as a law firm rather than pure software allows access to confidential contract data that foundation models never see, creating a competitive moat through specialized training datasets.

Whatever data is in distribution, the labs will get to. What would remain untouched is the context that’s unique to each client.

Which is why Crosby is bullish on per-customer fine-tuning and evals. Context like which liability cap their GC has conceded three times, which indemnity language the counterparty always strikes, what they actually settled at last quarter — the stuff that a good firm knows and operates atop.

There are a lot of clashing opinions on whether data is a moat at all, primarily re: pure-play software. What’s under-discussed here is a core difference in the nature of the data that a software product gathers vs. what a services/auto-pilot product does. Software is traditionally limited to data on actions and states, not on the quality of work. A neofirm or auto-pilot has access to both.

Note: Make sure your MSA permits you to use service data to improve the service.

#7: The open question

The only deduction we can make from the limited data we have so far is that neofirm margins sit below software and above traditional firms.

Bessemer’s vertical AI portfolio averaged roughly 56% gross margin on a 1.6x burn ratio in 2025. The incumbents aren’t measured the same way — so what we could get to was EBITDAC. Public insurance brokers sit at 30–35%, and accounting firms at 20–30%.

“There are types of human labour that you can replace today, but the amount that your clients will pay you may not be more than the inference costs of serving those clients.” — Ryan Daniels, Crosby

The bet, then, is not that services will become software overnight, but that the best services firms will start behaving more like software companies, with higher margins, proprietary data, and non-linear revenue per employee.

Some of these companies might only end up becoming excellent, modern services firms, and that would still be a large outcome. But the more interesting possibility is the one Julien Bek pointed to in his thesis, that:

”The next $1T company will be a software company masquerading as a services firm.”

A services firm to the customer. Software economics underneath.

Three years in, there’s no clean answer to the head-scratcher: “What are the paths (known and unknown) to AI-native PMF?”

What margins look like, where value accrues, whether vertical integration still works — while a whole set of critical questions have been made impossibly murkier.

We recently brought together three people (whom you’ve likely heard of) to talk from their own first-hand collisions with this question.

ft. Nikunj Kothari, partner at FPV Ventures and a sharp commentator on the idea maze, GTM, and product who has caught the moment with his prolific Substack,

& Helen Hastings, the founder of Quanta, an accounting services company living out the “Services as Software” thesis,

& and Chun Jiang, CoS at Miro, and an ex-founder who also previously led product at Reforge and helped ship 5 new AI-native launches in a year.

➤ Three threads you should dive in for:

• Why wedges are more fraught (as well as more important) than ever before,

• how Quanta and Reforge guarded against an unraveling while scaling their PMF,

• and why pricing has become doubly critical, and Nikunj’s updated take on why his popular blog “Your ‘Per-Seat’ Margin is My Opportunity”.

Plus some hot takes! ♨️

#1➤ Finding wedges that have an AI future

“In the era of AI, each company and each person is being asked to do more. So either you have to pick a wedge that is extremely deep and you build something end-to-end or if you are picking a horizontal platform, you have to be multi-product from day one.”

Having advised plenty of early-stage founders, Nikunj has seen the wedge question — how you enter a market, and through which door — become increasingly harder to answer. He mentions two traps that he scans for when investing.

“Either the product is so narrow that the market is not large enough and it’s not solving a hard pain point or it is so horizontal that they lose their identity. The best companies are able to go both deep and horizontal at the same time — and that’s extremely challenging, especially when there’s so much competition with every vertical and every space.”

What he’s saying is that if you are too narrow, you can build something genuinely useful but still hit a ceiling at $5M ARR. If you are too horizontal, you can confuse buyers.

Nikunj calls them traps because they are deceptively fundable in the early rounds, but neither is venture-scale. There’s a simpler, clarifying heuristic he has since developed. He recommends going super deep on the problem but also keeping an eye on the adjacencies you can get into in the next 6 to 18 months.

Helen is building a new kind of company, an AI-enabled services startup for an established problem, which has given her a unique PoV on the question of wedge.

“Our wedge was to start by selling to tiny companies even though I ultimately knew that our main buyer would be at a bigger company.” She thinks of wedges as what needs to be true in the DNA of a company from day one. For Quanta, that DNA-level-necessity was that they needed to do all of the accounting and be the full financial source of truth.

“That was my vision for how we’d actually be able to capture a very, very large market in the long term. But that would be a lot to build, and have a lot of problems if I started by selling to bigger companies. So if I wanted to hold that constant, what are the other dials or variables I can turn? That was company size and stage, and so that’s how we came up with tiny companies as our wedge.”

She could adjust who she was selling to, what industries she prioritized, what geographies she operated in. But the full-service, full-ownership model was non-negotiable, what Helen calls “the natural stepping stone to a very large market.”

Helen is extra cautious of mistaking something for a wedge, especially right now where companies are choosing wedges just because they happen to be making a lot of money now without being able to paint a picture of how they’ll pass some initial milestones to get to a very, very large market.

Chun got to this lesson through a harder path.

Before Reforge, she had built Monterey AI in 2022 as an AI document generator for product teams — using GPT 2.5 to help PMs draft specs, even before ChatGPT existed. The wedge they had picked was “doing one thing really, really well and going to a market we had connections in.”

Once they had users and early revenue, they started looking for the next thing to build to make the product more sticky. That’s how they pivoted to a Voice of Customer platform before exiting to Reforge, taking a lesson along:

“If I would do it again, I’d probably double down on the market or distribution as a wedge instead of really focusing on the product itself. Because when I was selling the document generator, I would just try to sell to the people. But if I had asked more questions about their technical stack, what other pain points were in their workflows — maybe that would have given me even more ideas about where I could expand to.”

#2➤ Building for the $6 TAM

For every $1 spent on software, $6 is spent on services. That’s the market Julien Bek at Sequoia pointed at in March when he argued that the next trillion-dollar company won’t look like a SaaS business — it will look like a services firm.

Helen has been building toward that for years. Quanta doesn’t sell accounting software, it does your accounting. Month-end close, tax filings, all of it. We asked what changes when you’re on the hook for the outcome, not just the tool:

“When you are selling software, you have your one deliverable which is your product roadmap that is shared by all of your customers. When you are a services provider and on the hook for the outcome, you are on the hook for many deliverables. A deliverable per customer that all might look a little bit different and have a different messy reality that maybe you didn’t expect when you first started working with them… We have to close the books at the end of the month no matter what. No matter what weird thing they did that month.”

Helen shares some instances to help visualise: a company might adopt a buggy beta payroll feature from a vendor that creates accounting chaos for Quanta, but Helen and team would still have to do the accounting for it, or a customer’s employee might decide to live in four different states in one month and trigger complex tax implications.

“We’ll need to handle it because it happened that month and we need to close the books that month, we don’t get to say, ‘We’ll do this in Q3, we’ll put it on our product roadmap.’”

Quanta’s model does come with 4 genuine upsides, in Helen’s words:

1. Quanta is always keeping up with what the market needs because they are forced to.

2. They get to build with real data, not hypothetical test cases.

3. They have a forcing function for using AI for what it is good at and handle different kinds of edge cases that would otherwise break traditional software.

4. The $6 of services for every $1 of software is a big one, of course.

On the cons, Helen says: “We do not have a lot of control over our roadmap and often we have to build things in a hacky way and then come back and build them later.”

#3➤ Humans + ownership = the real AI leverage

Chun noted how a lot of companies go to Reforge wanting to know how the team shipped five AI-native products in a year — especially for a company that didn’t sell software until recently. They assume there’s a secret AI leverage point. There isn’t.

She goes on to talk about the real leverage: people.

“We still attribute that speed to the human foundations that we protect — collapsing the role between engineering, product, and design, pushing engineers to do customer success, pushing PMs to do sales. Once you push people to do these external-facing tasks, they start realizing, ‘Okay, why are we building these features? Is this feature really worth keeping? Should we just kill the whole product?’ As long as you give people enough protection and context and empower them to make decisions, they are the ones that will help you find PMF really fast.”

Finding PMF is only half the problem, though. The moment it works, a different kind of discipline needs to take over, one that Helen has seen and grappled with up-close.

There was a moment in Quanta’s journey that made the reality of selling work visceral: quality is everything and it takes discipline to protect.

Helen’s company was doing incredibly well, bringing on more customers per week than ever, and yet she had to make the decision to stop taking customers.

“We knew that the market wanted the product, they were banging on our door, but we couldn’t reliably handle all the customers that wanted to work with us. And when you have this service layer, it becomes really easy to say yes to a prospect that maybe you can’t fully handle. Once we started bringing on many customers per week, we couldn’t keep up — and we had to take a pause and catch up and build more.”

The obvious pressure-relief valve would have been to hire offshore bookkeepers who were available, affordable, and could handle the volume while engineering caught up. She decided against it. The quality that made Quanta’s early customers stay and refer others was accuracy, consistency, and a level of reliability that made leaders feel like they could trust the numbers

“I did not feel like that was the right decision. It would dilute what made us get product market fit in the first place — our level of quality.”

Helen’s long view on services as software is worth sitting with. The question for AI services companies isn’t whether you can preserve quality as you scale, it’s whether quality can actually compound… something traditional services businesses have never been able to do. And she believes the opportunity is now.

#4➤ The era of “yes, and” pricing models

Nikunj wrote a hot piece almost two years ago titled “Your “Per-Seat” Margin is My Opportunity” arguing that when the primary users of a platform are AI agents rather than humans, per-seat pricing becomes harder to justify and sustain.

He says that it hasn’t happened yet because a lot of buyers are too used to buying software on a per-seat basis. “It’s predictable, your OpenAI bills don’t go up by massive amounts every week or every month, it’s still a transformation, but I think slowly people are getting around to some version of a hybrid scenario.”

Like a “yes, and” thing. “Where you are paying maybe per-seat for humans from the platform perspective but a lot of workflows that run on top of it are being billed as usage or services revenue depending on who is doing it.” he adds, before pointing to the inevitability of the shift.

“Humans are not going to be the primary users of these platforms in the future. How do you price per agent? It doesn’t make any sense. Going back to the $6 to $1 — it’s end-to-end outcomes. What are the outcomes that you’re trying to solve?”

Going back to the inevitability, he points at the classic innovator’s dilemma at play here, insisting that somebody will come bottoms-up, and just do usage-based pricing and it would be far cheaper because they won’t be keeping humans in the center of what they are doing.

#5➤ Hot takes, worth taking seriously

Nikunj shares that he wishes someone had named the technology “amplification intelligence” instead. The word says what AI actually does — it multiplies what’s already there instead of replacing it.

He sees that the bottleneck for most organizations isn’t the technology, it’s the people at the top not using it themselves. The ones who do model a different pace of work, and teams follow.

Chun’s hot take is her emphasizing on two things AI can’t do: building relationships and having fun. She makes a call-back to her feelings about distribution, sharing that your first customers will come from people who trust you and to invest there.

And to have fun, striking a chord with Helen. “Have fun. My favorite thing about working at a startup is everyone’s rule breakers and innovators — find these people you really enjoy working with.”

Helen ends on margins, something we’ve all been thinking about. “The part of the founder playbook that used to work but no longer works is the belief that you don’t have to pay attention to your margins because they’re going to magically figure themselves out later on. You’re going to wake up one day and have 90 plus percent margins and everything will be okay.”

She goes on to say that she works with a lot of companies on their P&L and a lot of companies have not picked up on that yet. “They are ignoring their margins because the toolkits weren’t made for early stage companies.”

In AI businesses, costs scale with revenue by default. Infrastructure spend, model costs, and compute often rise alongside customer growth.

That was three keen, sharp thinkers on the shape of AI-native PMF, caught at Chargebee’s annual conference Beelieve at Fort Mason, SF. You can watch the full conversation here.

Also, if you’re in the UK, join us at Beelieve London on June 4 for more insights and hot takes on AI monetization from Miro, Hostinger, Fin (Intercom), Artlist, and more! Get your ticket here.

Follow this space for more.

Until then, keep on updating your priors.

Thanks to Helen, Chun, and Nikunj for talking to us so candidly. 🙌

To not pick a problem so easy that it’s just trivial, and not so hard that every day feels like pushing a boulder uphill. But just at the edge.

It takes most people a few rodeos to find it, Stefan Bader had a hunch after his first.

Stefan is a second-time founder with Cello. Prior to this, he built and sold Parkpocket, an automotive intelligence startup, to Continental AG. But lessons from Parkpocket stayed with him: what a difficult market looks like, what it feels to sell in a space that demands in-person pitches, and most importantly, how the nature of a market — even when you have all the resources — can constrain your speed of distribution.

So when Stefan started again, he had a harder checklist. It had to be a category that wasn’t crowded, with distribution baked into the product, and a problem that could birth a generational company.

B2B user referrals cleared all three. That plus the fact that he couldn’t find a single good solution for B2B referrals while looking for a tool as a CRO at Shore, right before starting Cello.

For this edition of PMF /evals we sat down with Stefan to understand the makings of this journey. Here’s what surfaced:

• What Stefan learned about B2B referrals and why paying a private person from a business is a complex FinTech problem,

• What charging on two value metrics taught Stefan about psychology, why SaaS-looking tiers beat take-rate pricing, and lessons from 5 pricing iterations,

• Stacking up moats for the future, and

• Stefan’s long-term vision for when Cello becomes the ‘revenue distribution infrastructure’ for the AI-era.

#1 ➤

The thorns: why B2B referrals had remained unsolved

Cello’s origin predates AI, but they have seen tremendous tailwinds because of the AI shift, or more specifically what they call the “AI-product-channel-fit collapse”. Their curated research makes it even more impossible to ignore: sites have faced a ~79% drop in search traffic, cold call connections have gone from 54% to 20%, and outbound replies have dried up as AI emails have flooded inboxes.

This shift has made teams as well as users value and trust WoM more.

All of this is recent, though. Even before AI made the referral opportunity clear, Stefan had stumbled upon it and identified the gap. It was at his last organisation before Cello, after hunting for months and being unable to find a good tool, that he went builder mode on the problem. He quickly realized that the market only had affiliate or consumer referral tools.

While they worked for the B2C context, the B2B context was wildly different.

“In B2C, the person using the product is the person paying the bill. In B2B, the person using the product isn’t the person paying for it, they are usually an employee and it’s the business that is paying the bill. So they are two very different contexts. You can give them all the product perks for free and it would still not incentivize them.”

One of the most effective incentives that had worked in the B2C context, and had a lot of potential in the B2B context, was cash benefits, beyond just features in the product, so Stefan started building for that. But constructing cash benefits as referrals was a complex problem:

“Paying cash from a business to an individual isn’t a trivial transaction, it’s a FinTech challenge in its own right, especially when you’re dealing with users across different geographies.” Stefan said.

For B2B referrals to do cash payouts, they need an elaborate compliance infrastructure and global coverage which B2C referral tools had never accounted for. A gap wide open, albeit technically complex.

Today, Cello is simultaneously an AI-enabled SaaS, FinTech, and infrastructure — deeply integrated with an ecosystem of gateways and internal tools — tackling issues associated with the digital movement of money, such as fraud, global compliance, and payouts.

#2 ➤

178% NRR, 5 pricing iterations, and lessons in buyer psychology

Cello’s exceptional NRR of 178% didn’t happen by chance, but rather through a deliberate set of decisions based on a deeply value-first pricing philosophy.

Today, Cello’s pricing model is fundamentally outcome-based with an underlying take rate, but their pricing looks more like SaaS tiers at first glance.

Stefan’s core pricing belief behind choosing this model has been to create a low-risk entry, let users derive value from the product, and then monetize alongside the revenue and ROI unlocked through Cello.

This hunch also came from Stefan’s experience as a CRO. He asked himself, “What pricing would I be willing to run with immediately, without a long budget approval conversation?” and had an answer.

Since then, they’ve prioritized walking in lockstep with customers instead of intimidating them with upfront commitment or complexity.

This philosophy also extended to the value metric they operate around: Active Referral ARR i.e. the active recurring revenue a customer gets from Cello-driven referrals. It’s a win-win; as that number grows, Cello makes more, and customers don’t mind it because the tool effectively pays for itself.

Cello current prices using SaaS-like tiers with outcome/value-based pricing on top of it.This took Stefan and team 4 different pricing iterations, each with its own pros/cons. Here are the key lessons that emerged from them:

1. Charging on two value metrics can be perceived as more expensive than it is: Early on, Cello charged on both Referral Signups and Active Referral ARR. Even if the math worked out similarly, showing two dimensions of billing created psychological friction. They decided to drop the volume metric and keep it simple.
   

2. SaaS-style tiers outperform take-rate pricing on perceived value: Cello is fundamentally an outcome-based business with an underlying take rate, but their pricing looks more like SaaS tiers at first glance. Stefan saw that presenting an outcome-based business made customers uncomfortable. A percentage-based fee required math, while tiers reduced cognitive load, and so Stefan went with tiers. Alexa Grabell (Pocus) ran into the same wall in her experience.
   

3. Freemium can crowd your funnel with unserious users: Cello ran a freemium model early-on before deciding against it. It drove signups but attracted a lot of “tourists” in Stefan’s words and eventually took over their onboarding capacity. They saw that a small entry price — even $200-300/month — meaningfully increased the seriousness of implementation and weeded out unserious users.
   

4. If no one complains about your price, you’re charging too little: Selling a new kind of tool meant that Stefan had to find the “right price point” almost from first principles. As much as he wanted a low-risk number, he also didn’t want to leave money on the table. Stefan looked at public benchmarks for different tools, experimented with different versions, and zeroed in a number that had some friction. “If your numbers are too good, you’ve left money on the table,” he adds.

We asked Stefan about the implications of AI inference costs on their margins, and whether they’ve had to revisit pricing since adding AI features. Stefan shared that it had no effect, before laying out why:

“Embedding AI into our product actually makes it more scalable and higher-performing. Yes, there are inference costs, but the gains in efficiency more than offset them — we need less headcount per customer. On top of that, AI improves the performance of our programs, driving higher NRR, which far outweighs the cost of inference.”

#3 ➤

Stacking up moats

As someone aimed at building a generational company, Stefan has deeply thought about moats and defensibility. He agrees with the sentiment that software isn’t a moat anymore, and that you need distribution, brand, trust, workflow integration, and network effects to survive and scale.

Not one of them, but multiple.

On the distribution moat, Cello’s baked-in Casual Contact Loop has been working tremendously well. Cello’s referral widget is embedded inside its customers’ products, so every time people see it, Cello gains mindshare.

Stefan is also a big believer in the power of network effects as one of the strongest moats. Today Cello is building multiple of them simultaneously — an unusually thorough strategy for a B2B company at seed stage:

1. Performance network effects:
   Every customer running a referral program through Cello generates data on what works. Cello monitors performance across all of them, runs A/B tests, spots patterns, and productizes what wins. “We’re running dozens of A/B tests with customers out there. Whenever we see that a certain pattern works better, we’re productizing it and rolling it out.” A new entrant would need to run the same experiments from scratch.
   

2. Fraud/compliance network effects:
   Stefan shares that any system that pays out monetary rewards tends to get gamed: self-referrals, stolen credit cards, brand keyword arbitrage all become likely. “There are always new patterns popping up and new approaches to basically hack the system in order to extract money out of it. And therefore, this is also a learning curve we’re going through and we’re patching it, productizing it.” More customers means more attack surface, which means faster detection making their pattern library self-reinforcing.

Bessemer’s 2025 State of AI (and many others) also argues that “memory and context are the new moats — the most defensible products will remember, adapt, and personalize.”

Cello’s integration depth across payment gateways, CRM, and product analytics is effectively a memory layer — it knows more about what makes a user likely to refer, what incentive structure converts, and what fraud looks like in a given market than any new entrant could reconstruct.

#4 ➤

GTM as infrastructure — the long game

Cello started with B2B referrals when there was no product for the problem, and then gained tailwind as AI made most channels collapse and reaffirmed the weight of WoM. Stefan thinks there’s a stronger shift to ride and build for, and has a thesis for how it will play out.

He envisions Revenue Growth Infrastructure sitting alongside cloud, models, and databases and becoming a core layer for any online business that wants to manage and accelerate GTM for human and/or agent-led transactions. It’s an ambitious vision, but one that derives its conviction from a simple observation: “Whatever humans do in front of a computer, agents will eventually do.” Stefan adds.

He believes that agents and humans both would need the right data to learn and act. When that happens, the infrastructure layer enabling go-to-market will become as foundational as the storage layer or model APIs.

Stefan is actively building for that world and their roadmap backs it up, with an eye on adjacencies like reviews, UGC, and GEO. Because they are already embedded in their users’ systems, Stefan’s hunch is that expansion and stickiness will be much simpler:

“We occupy a unique piece of real estate within our customers’ products — and that gives us the opportunity to expand horizontally within a space they already trust us to own.”

With 250+ customers like Typeform, Descript, and Miro, that trust is already there.

P.S. A big thanks to Stefan for taking the time to talk to us. And all the best to the Cello team for what’s ahead.

PMF /evals ◎ is just getting started. Tell us about how you’re approaching AI-native building in the wild! Or what you’d like us to cover. Hit reply.

We’ve all now learned that the truth is much murkier.

SaaS is, as it happens, simultaneously under threat — and it would be irresponsible to think otherwise. But also still deeply entrenched, and it would be equally irresponsible to “project a 5-person startup behavior onto a Fortune 500” in the words of Sarah Guo from Conviction.

A moment that renders itself perfectly to Schrödinger’s uncanny experiment.

We wanted to dedicate this edition to the nuanced takes from the interwebz that carry the rigor this moment, uncanny as it is, warrants.

Let’s start with what has fundamentally shifted, the thing that led to all of this speculation and chaos: AI has made the SaaS business model unstable.

SaaS margins were anomalously high, in most part because of three factors — zero marginal cost of reproduction, non-ephemeral value, and high switching costs. AI degrades all of them simultaneously: inference adds real cost, faster model cycles accelerate depreciation, and agents don’t particularly care which tool they’re using.

On top of that, the abundance of code can threaten pricing power, a pure-play seat-based model can have serious limitations when charging for work, and the model companies themselves can double-up as very capable competition, almost out of the blue.

The extent of the selloff was dramatic, yes, but the market pressure on public SaaS does have some feet. Avenir’s January 2026 report shows that public SaaS companies have not been performing as well as they used to. At this point 2020-21 seems like an aberration:

Public SaaS median revenue growth, peak (Q2 2021): 44%

Public SaaS median revenue growth, Q3 2025: 18%

SaaS equity performance since Nov 2021: -27%

Horizontal SaaS index since Nov 2021: -49%

Interestingly, a cool-off post the much hotly debated sell-off is already afoot.

The public markets have shown that they were, and will continue to be, swayed by incentivized narratives, poorly-informed fears, and short-term hedging. The builders need to focus on an entirely different set of questions. Let’s dive in:

#1. The threat from the model makers

It’s no coincidence that the selloff came right after Claude launched plugins for 11 use cases like marketing, legal, and sales. When the model maker shipped something that could be even a bit of the interface, every company that owned it was asked to defend itself.

Elad Gil on No Priors points to a historical pattern here: Microsoft forward-integrated into Office, Google forward-integrated into vertical search — travel, local, maps. The AI labs will likely do the same.

Elad shared that every major platform shift has seen the platform move into the most valuable applications sitting on top of it. The question he raises for every founder is: “which verticals are durable and defensible, and which ones will get eaten?”

There are three popular ways companies are navigating this — each with a different logic:

1. Own institutional judgment, not just the interface

On February 3, Anthropic released the Claude Cowork legal plugin and Thomson Reuters dropped 16–18%. Three weeks later, at a separate Anthropic enterprise event, Thomson Reuters’ CoCounsel AI tool was featured on stage and CEO Steve Hasker presented.

The stock surged 11–14% that day, its biggest intraday gain in 26 years. Hasker’s argument was that Anthropic builds good general-purpose automation, but replicating what Thomson Reuters offers would require massive investment to acquire decades of curated legal content and 2,700+ attorney editors.

Rajan frames this as the real moat: the density of institutional knowledge encoded in how your humans and agents work together. That can’t be made a feature or a plugin.

2. Specialize into what’s too specific/niche for a foundation model to bother

Spellbook, a contract drafting tool for lawyers, found that the Claude legal plugin launch actually increased their usage. Their take is that a foundation model will automate the generic user cases. It won’t go deep enough into the focused workflows, edge cases, and jurisdictional nuances that a specialized tool has spent years encoding.

As the analogy goes — every startup used to be asked why Google wouldn’t kill them and the answer has simply been the fact that those areas are just too niche or too specific to be worth it for them.

3. Build in a way where every model improvement makes you better, not obsolete

Crosby, an AI legal services company, started with lawyers in the loop — deliberately. Their reasoning: “You want to do end-to-end work, what you’d expect from a law firm. AI can’t do this alone.” Their bet is that by encoding legal judgment alongside AI capabilities, each improvement in the underlying model makes their product better rather than threatening it.

#2. Worse than SaaS. Better than most everything else.

“Software is a business tool, not a business model.” — Sam Lessin

Software was always a tool. It was SaaS that was the business model, of a certain outperforming shape. And now that shape is changing.

Dave Kellogg of Balderton Capital predicts the Rule of 40 becomes the Rule of 60 in this podcast conversation, as margins compress and AI raises the cost of delivery, the threshold for a compelling SaaS business rises.​​

This transition is yet to be realized fully, and one of the levers that will play into it the most will be the monetization model, which has dominantly been seat-based.

For instance, GitLab is one of the companies that has benefited tremendously from AI wrt an uptick in usage/adoption, but their pricing is making them leave money on the table. The GitLab CEO Bill Staples and CFO Jessica Ross hinted at the importance and urgency for a need for pricing innovation during the Morgan Stanley TMT conference. They shared:

“Their current business model of seat-based pricing didn’t anticipate the surge in AI-driven surge in platform usage. Customers get unlimited access per seat which means the business has delivered more value without capturing more revenue.”

The pricing shakeup will be uncomfortable in the short-run but will ultimately be the only way to keep and grow margins in proportion to the value delivered. And based on what we’ve seen work and break while enabling companies make this transition, we’d add a +1 to the urgency that pricing innovation warrants.

Margins will still exist, they will just be compressed. A pure software business that earned 25% income margins might earn 10–11% going forward. Yoni Rechtman calls this in his post “Creating margins after SaaS.”

“The income statements will look ‘worse’ by SaaS standards and better by every other standard. The software companies that succeed in this brave new world won’t be pure application software and they certainly won’t be traditional SaaS. But they will be bigger and throw off huge amounts of cash to their shareholders — at permanently lower margins.” — Yoni Rechtman

Lower margins don’t automatically mean bad businesses. The question is whether the company can still generate durable value. Rechtman considers three potential paths:

1. Network effects — Value that compounds with scale, cost to compete that rises with each new participant. Agents interacting with your network accelerate the compounding.

2. Sin eater — Absorbing high-stakes, operationally intensive work that businesses want off their plate. Security, fintech, compliance, legal. Too high-volume to do manually, too high-stakes to do generically. That’s durable.

3. Hybrid models — Software plus services, hardware plus software. Physical switching costs, proprietary data from sensors and devices, forward-deployed humans. This is why Samsara is holding up when pure horizontal SaaS isn’t.

“Connecting the foundation models to high value complex real-world use cases is the biggest opportunity for wealth creation in recent memory. It will happen through new (often hybrid) business models and products rather than re-running the cloud playbook.” — Yoni Rechtman

#3. The next battle is for context

Another useful idea from Avenir’s January 2026 deck is the fast-intensifying battle for becoming a system of context which is playing out across enterprise categories.

Intelligence is commoditizing fast, GPT-4 level capability dropped from $60 to $0.40 per million tokens, a 99% reduction. What doesn’t commoditize is context: situational awareness of a problem, knowledge of the humans involved, institutional operating norms.

Some incumbents are winning the early agentic revenue: Intercom’s Fin at $100M+ ARR, Salesforce Agentforce at $540M ARR, ServiceNow Now Assist at $500M+ ARR. What’s worth noting is that these wins are concentrated in the support and service workflow category. Re: the AI-natives, they are building fast, but without the context advantage yet.

Emphasis on yet.

Further, only 2–7% of enterprise buyers put more than 80% odds on replacing major vendors like Atlassian, Workday, or Salesforce in the next five years. Systems of record aren’t being ripped out en masse. But the interface layer above those systems, and the workflow layer between them, is actively being contested.

Atlan’s co-founder, Prukalpa Sankar, shares why things are going to get complicated for both sides of the battle: “context is cross-system, multilayered, and heterogeneous”. Every enterprise has a different combination of tools; one runs Salesforce + Zendesk + Snowflake, another runs something else entirely. She goes on to illustrate why:

A renewal agent, for example, doesn’t just pull from your CRM. It needs PagerDuty, Zendesk, Slack, Salesforce, and Snowflake to do its job. To truly capture the context graph, a vertical AI startup would need to integrate 50–100+ systems just to cover common cases. Incumbents represent one part of the stack. AI natives have to figure out the whole thing.

Theory Ventures’ GP (and backer of 9 enterprise unicorns!), Tomasz Tunguz, adds another dimension to the context battle: incumbents are already changing the rules of the game. Salesforce restricted Slack’s API, Datadog deactivated a competitor’s account, and Epic faced a lawsuit for gatekeeping patient records. If your architecture depends on clean programmatic access to incumbent data, plan for that access to get more expensive or disappear.

He foresees that some AI products will make such defensive attempts futile by changing the very internal processes that SORs serve today:

“When AI products are sold as services, they replace in-house labor. This changes internal processes. When the internal processes change, the opportunity to replace the system of record arises because the existing workflows are no longer relevant.”

#4. In the age of abundance, who wins?

“If code, compute, and intelligence are all getting cheaper — who actually wins?”

Packy McCormick grounds all the different facets, opportunities, hype, et al. with this one question. He then turns to history to gather what has worked in domains that have seen similar abundance of inputs.

Looking at Rockefeller, Carnegie, Swift, and Ford, he highlights the playbook they all ran: identifying the binding constraint holding an industry back, breaking it using new technology, seizing the resulting High Ground, and integrating outward from there.

“…what matters is becoming the leader in your industry in a way that is incredibly specific to that industry and in such a way that your business benefits from, instead of being threatened by, abundant improvements in general purpose technologies like AI and batteries.”

Packy adds a unique, helpful nomenclature: “The High Ground is the scarce and valuable position you win by breaking it. Moats are what keep others from taking it.”

Ramp vs. Brex makes it concrete for our world of SaaS. Ramp is worth $32B. Brex sold to Capital One for $5.15B. Not because of a 6x revenue difference, but because of something else. In Packy’s words here’s what happened:

“Ramp used a card, software, and counter-positioning to attack what it viewed as the Schwerpunkt in corporate spend (the fact that everyone was selling money, and no one was selling time) and win the transaction layer, the High Ground from which it is now expanding to eat every point solution a finance team touches.”

In this same post, Packy shared Flexport’s Ryan Petersen tweet putting it plainly: “We’re so lucky we didn’t listen to all the people who told us to be SaaS — almost everyone ‘smart,’ by the way.” He was advised to sell software to freight forwarders rather than actually becoming a freight forwarder and competing. Flexport became the freight forwarder.

He shares an equally pointed implication for investors:

“…getting comfortable with a wider range of business models to accommodate whichever is the right one for the industry in which a company operates. This does not mean that they should treat all business models equally now. Instead, they need to stop blind pattern matching altogether.”

He goes on to add that services businesses might be terrible for most companies but exactly right for some. Stripe shouldn’t be a services business. An AI-native law firm probably should be.

#5. Vibe coding as a threat, or not

The loudest version of the bear case goes like this: AI makes software so easy to build that companies will just vibe-code their own tools internally, collapsing demand for SaaS entirely.

It’s a good story. It’s not what’s actually happening.

Geoffrey Moore — three decades advising Salesforce, Microsoft, and SAP — makes the structural argument. He says that every enterprise activity is either Core (what differentiates the company) or Context (everything else — regulatory compliance, standard operations). Mission-critical context is where SaaS lives. And it should never be differentiated. You’d want payroll to be reliable, auditable, and maintained by specialists whose entire business is exactly this problem.

He adds:

“Could you use AI to write your own system of record? Yes, you could. But why? Advocates will say it will save you money — is that really true? In year one, it is. But before you declare victory, ask yourself how realistic your DIY accounting is, especially if you have neglected to incorporate annual maintenance, not to mention a discount for risk, and another discount for the opportunity cost of distracting your AI talent from pursuing core.”

The heart of Geoffrey’s statement is in the examples he states — a system of record, an accounting platform… the complex stuff — leaving a very specific tail-end of simple SaaS apps exposed, and a recent analysis by Anton Osika of Lovable validates it:

Jason Lemkin from SaaStr shared his perspective on what’s happening:

“People aren’t vibe coding their own Salesforce. They’re not rebuilding Workday or replacing Snowflake. They’re replacing the $49/month SaaS tool that does 80% of what they need and 40% of what they don’t.

Simple B2B apps. Simple ones…This is meaningful for B2B founders to pay attention to. The “simple tool” end of the market is going to get compressed.”

For the platforms and the complex tools, Avenir’s data closes the loop with a throwback insight on price compression. Cheaper software has always existed. Zoho has undercut Salesforce for two decades without meaningfully compressing SaaS pricing.

SaaS is already inexpensive relative to the labor it replaces, a Salesforce seat at $1,200/year versus an $81K sales rep salary. And R&D as a percentage of revenue is often modest, meaning there’s less cost advantage to building in-house than it appears.

Further, brand-name SaaS offers more than code. They win on feature completeness, ecosystem partnerships, security, predictability, and herd familiarity. Vibe-coded alternatives win on customization and cost, two dimensions that matter less for mission-critical workflows.

#6. The timeline and the horizon

The honest version of the bear case isn’t “everything disappears overnight.” It’s that the unit economics of pure application software have structurally changed, and the market won’t pay 2021 multiples for that profile again. Less dramatic than the discourse, but more accurate.

Rory O'Driscoll captures this shift succinctly:

“There are growth investors and there are value investors, and there is a chasm in between. The SaaSpocalypse is really about the SaaS story shifting from the former to the latter.

From 2004 to 2019, public SaaS companies grew at an average of 30% and were valued at an average of 6x revenues. It was the best, most predictable growth story in the market. Then two years of COVID gave us 40% growth at 20x revenues.

…

Everyone is finally realizing that the growth story is all AI. SaaS companies without AI-driven growth will be valued like every other company in the market, on a growth adjusted multiple of free cash flow. Amazing companies with real value but no pixie dust premium.”

That’s what it is, the pixie dust premium has settled.

Sources: MT Rajan · Yoni Rechtman · Dan Hockenmaier · Avenir · Geoffrey Moore · Packy McCormick · No Priors · Jamin Ball · Jason Lemkin · Prukalpa Sankar · Tomasz Tunguz · Rory O’Driscoll

PMF /evals ◎ is just getting started. Tell us about how you’re approaching AI-native building in the wild! Or what you’d like us to cover. Hit reply.

Pre-AI, there were two kinds of software for small businesses: a system of record (ex: a CRM) and growth tooling (ex: website builders), as distilled astutely by Primary VC.

SoRs failed at adoption because they needed someone to maintain them, and SMBs didn’t have the time to extract ROI. Growth tools saw adoption owing to their promise of topline growth but often couldn’t deliver. The market penetration for most SMB SaaS stalled at <20%.

At the same time, businesses faced pressures to grow and so there was a long tail of SMBs left behind, as SaaS moved upmarket to SMEs.

AI is shifting this, opening the market that legacy tools never reached.

One, ROI-conscious SMBs find the work-as-software model appealing; they now pay for the work not the software. Two, as Tomasz Tunguz shared in our 2025 State of Recurring Revenue & Monetization Report: “AI has enabled sales development-led experiences in market segments that historically have not been economically viable.”

Jérémy Goillot, the co-founder of The Mobile-First Company (TMFC) is surfing this newly-defined opportunity with a thesis that is working remarkably well. Today, they have 20K+ SMB customers, $17M raised, 50% MoM growth in revenue and usage, and customers exclaiming that “Moving to Allo (one of TMFC’s three products) felt like upgrading from Windows 98.”

AI-SaaS for SMBs is taking off and Jérémy clearly knows something. For this edition of PMF /evals, we sat down with him to find out. We covered:

• Why he thinks AI has made “boring” software categories an even bigger greenfield opportunity with newly-made-available budgets,

• How they’ve scaled past millions in organic views, 20K+ customers, and 500 tickets/week without any sales or support,

• How they are building a unique multi-product strategy with a plan to solve every problem that slows SMB teams down,

• Why they are focusing on pricing simply, and

• Why distribution and design >>> tech as a moat.

Jérémy spent 7 years at Spendesk, as the employee #1, before leaving to start something of his own in 2023. At Spendesk, he saw first-hand that most SMBs found traditional SaaS “too clunky, too expensive, and too difficult to use.”

When ChatGPT blew up, Jérémy knew that it was going to change SMB software in meaningful ways. “I really think that most small and medium businesses are going to want the ChatGPT experience for every software.”

#1 ➤

The “electroshock” moment of SMB excitement

“We have established companies in boring industries like renovation and energy consulting coming to us with a desire to replace legacy tools. They’ve used it for 10 years but now they are done.”

He calls it the “electroshock” of AI. For years, the industry tried to convince SMBs to digitalize. And now thanks to AI, they’re not just digitalizing, they’re open to rebuilding their entire stack, because they have a fresh set of expectations. “When people try a solution, they expect to not pay for it. They expect to try it without any friction and they expect to see the result in less than five minutes.”

This new set of expectations, and behavior, is what Jérémy placed his bet on. They started with the most boring problem they could find: phone systems.

“Selling phone numbers, and helping businesses to send texts and receive calls is a very boring problem. Our competitor is RingCentral and they are a public company. The software hasn’t changed in 20 years. If you want to buy it, you have to literally email the sales team.”

One important element of their thesis is being mobile-first. Jérémy wanted to build for the small businesses who run their business on their phones, in the field. It’s also what sets them apart, in the words of their investor Antoine Moyroud: “What sets them apart is their intuitive, AI-driven approach that seamlessly integrates into existing workflows, making it an invisible force that just works.”

Their phone system product, Allo, has a mobile-first design with an AI companion that handles post-call tasks like updating CRM, sending summaries, and automated follow-ups, going beyond what legacy systems have done.

Jérémy always wanted to build a whole suite for SMBs — something that Primary VC also presents as a strong thesis — so they asked their customer base what else hurt, and tackled adjacent problems one by one. You could think building 3 products together is stretching yourself too thin, Jérémy thinks otherwise.

“If you start to have a product for engineers, a product for restaurants, a product for laundry, growth can be super hard. But if your customer base is the same and you build for the same audience, your team can be the same, your acquisition can be the same too, and you just need to upsell.”

Today they have 3 solutions all going against legacy incumbents:

• Allo – Phone system (replacing RingCentral)

• Due – Quotes/invoicing (replacing Bill.com)

• Claim – Expenses/invoices (replacing Expensify)

They have followed the same playbook for each of these products: take a legacy software, rebuild it with AI-native workflows and consumer-first UX.

#2 ➤

Leveraging AI to scale — without a sales or support team

“Small business owners decide to implement tools outside of business hours... because they’re solving problems for their business during the day… The night is the time where the product is most active.”

If your customers are implementing tools at 10pm on a Saturday, your product needs to work without support. And it so happens, that’s how SMB owners work.

Jérémy has been very serious about using AI to be more efficient as a business. With that in mind, they have invested heavily in self-serve design and AI support (an agent handles ~500 tickets/week), and spent two years training ChatGPT on their product so users can troubleshoot via LLMs.

“We publish a lot of content to make sure the user can support themselves.” That’s how Allo got to 5,000 customers with no sales team and no support staff.

Their acquisition channels also match how SMB owners actually spend their off-hours. Here are the 3 main strategies that have helped them scale their GTM:

• SEO for replacement intent. Their highest-performing keyword is ‘RingCentral login.’ → “The keyword ‘RingCentral login’ brings me the most customers because they are active users on RingCentral and I remind them that they’re paying too much.” shares Jérémy.

• Creator-led content. Their Head of Brand, Alexandre Calvez has 1M+ YouTube followers. In addition, they also partner with content creators, restaurant owners, and AI educators, who show how Allo works for them. One of these collab videos featuring a restaurant hit 1M views.

• Organic everywhere. They are making serious investment in social, SEO articles, and consumer-style content that meets people on the channels they use nights and weekends.

#3 ➤

Taking the scary out of AI pricing with a seat-based model

Allo sells with seat-based pricing. Everything included, no usage tiers, no add-ons. We asked why. Jérémy shared that he took inspiration from how ChatGPT handles pricing psychology and how people react to it:

“ChatGPT is losing money with some users and making money with others and it makes sense. Imagine convincing your mom that she can pay $3 this month, but next month if she asks too much, she might have to pay $20. She would be so scared she wouldn’t use it.”

This is the old dilemma with pricing when selling to an audience that’s used to simplicity and predictability. Something that Wispr Flow’s Tanay Kothari and Super’s (and Slite’s) Christophe Pasquier have both nodded to in their interviews as well.

Beyond tending to the psychology, they of course want to stay cost-conscious as well. So they analyzed their best customers and designed a pricing model that remains profitable for 80% of users, deliberately losing money on the 20% high-usage segment. “You need a price serious enough to make hesitant people not purchase... but you don’t want pricing to lose you deals because usage is everything when it comes to adoption.”

For now, he is keeping it simple, and advising other founders to do the same: “Just keep the simplest pricing you can before your Series A.”

#4 ➤

Distribution and design >>> tech as a moat

“It will never be the best product that wins. It’s the most easy to use product that will win.”

AI moats are debatable and precarious. Some startups have leaned into the model as a differentiator, building their own LLMs, which makes sense for a product like Wispr Flow where the model performance can create critical variance in the output/product quality.

Jérémy has decided to not go there. Instead, they use foundation models such as ElevenLabs (for voice within Allo) and reserve innovation for the application and experience layer.

Exactly why designers make up 20% of their team.

“Everyone believes you have to build your own LLM. A year ago, everyone agreed too. Now it looks like Google will win everything and in the end I’d rather just focus on building the layer on top.”

His hunch is that the easiest-to-use tool wins, especially when that’s your wedge. Plus the one who has distribution, of course.

Looking ahead: Covering the alphabet with a suite of products

Jérémy is modeling their multi-product expansion after successes like Adobe, Atlassian, Zoho, and Freshworks. He has one customer base in mind and every solution is fair game, as long as it serves SMBs that run on their mobiles.

The way he thinks of it it just has to work. “If Allo is our A, we plan to work through the alphabet of problems that slow teams down.” They are also precise with how they define SMB: 2-3 people, with the business owner involved, as soon as there is an IT team or finance team, you’re out of scope.

If Jérémy’s right about the ChatGPT bar resetting SMB expectations, there are dozens of legacy software categories waiting to be rebuilt, AI-embedded, mobile-first. They are starting with three.

P.S. A big thanks to Jérémy for taking the time to talk to us. And all the best to The Mobile-First Company team for what’s ahead.

PMF /evals ◎ is just getting started. Tell us about how you’re approaching AI-native building in the wild! Or what you’d like us to cover. Hit reply.

That’s Shensi Ding, co-founder and CEO of Merge encapsulating their 2025.

Merge sits atop a thriving business — $75 million raised, industry leaders as customers, and a billion API requests flowing through their systems daily; easily the gold standard for API integrations. Couple this with a surge in demand for integrations amidst the AI search boom, the tailwinds favor them.

Their core product Merge Unified helps companies like Ramp, Brex, OpenAI, Mistral, JP Morgan, and Mastercard with ingestion, syncing data between systems, and powering deterministic workflows.

This solution also seemed to extend competently to AI search companies that needed a way to sync data from their customers’ core systems like Jira, Sharepoint, Google Drive, Salesforce, etc.

And then, in the midst of this boom, Shensi and team saw a brand new opportunity.

They sensed two trends: 1) it was a matter of time that the integration world would expand to AI agents, and 2) AI agents needed their own fundamentals for interoperability.

AI agents don’t just ingest and do deterministic actions, they act on-the-fly: create a task today, update a record tomorrow, send a message the next second.

This warranted a product of its own.

For this edition of PMF /evals, we sat down with Shensi to go behind the scenes of this change. We explored:

• The insights that led them to build another successful product (Merge Agent Handler),

• How they zeroed-in on a consumption-based pricing model without much data,

• Why they are investing in services and Forward Deployed Engineers, and more.

#1 ➤

Seeing the shift early

“Our Unified API is really great for ingestion and deterministic use cases like syncing invoices with your ERP system. But it wasn’t necessarily built for on-the-fly actions.” Shensi and team had been spotting signs.

On-the-fly actions. The kind of thing AI agents need to do.

Shensi had seen a lot of software businesses go multi-product. So instead of feeling threatened by this opening in the market, they started talking to customers and prospects, forming their opinion, akin to the process they had followed in 2020 — talking to 100 people over a period of 6 months — to validate what became Merge’s Unified API.

“We did a lot of research to find out where this new product could fit in. We talked to a lot of existing customers, new prospects, and we had a few different angles and thoughts about where the industry was going, but increasingly what became obvious was that there really needed to be high quality tools for agents. Full stop.”

They had their conviction in the merit of a tool for agents. But when Anthropic released MCP (Model Context Protocol), it became incredibly clear that this wasn’t going to be a feature bolted onto their existing product. Agent-to-tool connectivity deserved its own standard.

“We wanted to build a full end-to-end product that would be able to help companies connect their agents to third-party systems, manage credentials, observe their agents’ activities, and add customizable rules to make sure their agents don’t use PII or any other sensitive information inappropriately.”

This vision was only strengthened by the context and knowledge Merge’s team had from years of understanding API nuances, partner dynamics, and integration edge cases. “It became really clear that we were uniquely positioned to provide this core infrastructure for a lot of AI companies. We could apply a lot of lessons from our Unified API to this new product.” Shensi added.

#2 ➤

The return of custom deployment options

One thing Shensi didn’t expect to run into was the return of custom deployment options.

“Increasingly, because a lot of companies are worried about AI, some teams are almost moving away from the cloud. They prefer to make sure that their data is siloed and LLM providers aren’t training on their proprietary information.”

Managed VPCs, single tenants, BYOK, deployment options that seemed dead three years ago have become requirements for enterprise AI conversations because “you can’t be sure what buyers would want.”

There was another interesting emergence they ran into mid-implementation: some enterprises can’t even use AI because they’re stuck on legacy systems with no public APIs. Their data is trapped in systems that predate the cloud.

“For a lot of companies, they can’t even use AI because they’re currently using on-prem systems or legacy software that doesn’t have a public API and that a lot of vendors are not willing to connect to.”

Merge’s solution here has been to lean heavily into a role that is making waves: Forward-Deployed Engineers. This team builds custom connectors for on-prem and homegrown systems, making sure customers with system or security constraints are still able to deploy AI.

A recent a16z piece, “Trading Margin for Moat,” argues that during platform shifts, implementation-heavy companies can build durable advantages. You might have lower gross margins upfront, but that’s a trade-off for stickier customers and pricing power over time. Merge is betting on the same.

“When you have the best product, you have a lot of pricing power. Right now we want to make sure we have the level of service and flexibility that customers choose us and we become the best product over time.”

Shensi sees this as both a moat and a necessity — meeting customers where they are.

#3 ➤

Crafting a pricing model without enough data

Agent Handler launched with consumption-based pricing: pricing per tool call, with a free tier up to 1,000 calls.

Why consumption-based, we asked.

“We wanted to optimize for simplicity.” Shensi shared.

It was also heavily influenced by their target persona and their preferences. While they didn’t have a lot of data on pricing, they did have a sense of industry preferences for dev tools, preferred billing and experience, etc.

“A lot of developers prefer to purchase a product that they can just sign up for and start testing right away versus having to have a conversation first. We also noticed that devs were familiar with a unit of work. For simplicity we wanted to get something out there that made sense.”

Shensi shared that their finance team came up with the price point through math and a hunch, which is a likely state for anyone just launching an AI offering.

“We didn’t have a ton of data when it came to our pricing, but with the data that we did have, we were able to make an estimate of how much each tool call would cost.”

They weren’t trying to find perfection either. “The goal right now is to get customers that will be very happy with our product and are willing to give us a lot of feedback so that we can improve and be the best.”

Shensi shared that they do re-evaluate pricing every six months to make sure it still makes sense. “Usually it bubbles up from the sales team to our executive team, and we decide whether or not we have capacity and whether or not right now is the right time to evaluate.”

They only make a change when there’s new data, the industry has changed, the customer base has shifted, or something in the business is causing the old pricing to break. They are prepared to iterate. “We’ll continue to adjust our pricing as we get more information. We embrace this approach, especially for a new product, but in the absence of data you just have to do something.”

In addition to this, Merge also offers custom contracts for enterprises that need custom SLAs, more support, and custom support. It also helps them serve teams that are particular about a sense of predictability.

#4 ➤

Investing in professional services as a startup

Merge is also investing heavily in professional services, a move that might look wrong if you believe the standard playbook about software margins. But AI has broken that playbook, anyway.

“Services and partnerships will become increasingly more important,” says Shensi. Betting on this, she asked one of her team members to create a new function that combines services and partnerships.

This has also been informed by a significant change in their customer composition. “When we first started, we were mostly selling to startups, but over the past few years, we’ve mostly transitioned to supporting enterprise companies, and they need more support.”

When they pitched the services offering to existing customers, the response was overwhelming. “We ended up getting a lot more interest than we expected. And that gave us a lot more data for why we should start investing more in the team.”

It’s the same “margin for moat” logic as their FDE investment — meet customers where they are, earn pricing power later.

The team is growing rapidly. One person joined in December, another joined in January. Shensi thinks that the services aren’t a distraction from the product, but R&D that pays for itself and helps them offer best-in-class service.

“It is also helping us offer the last-mile customization that customers are asking for plus open up to customers that might have been hesitant before.”

#5 ➤

Looking ahead

Shensi thinks there’s a lot of opportunity for connectivity in the AI space that no one else is really targeting. “It’s becoming increasingly clear that context is king, especially when it comes to AI, and there’s a ceiling for what reasoning can do without data.”

Merge’s Unified API has helped SaaS companies offer integrations at scale successfully. With Merge Agent Handler, they are helping AI companies connect their agents to systems securely and without any headaches.

That’s them solving integrations, twice.

Shensi is thoroughly excited. “Every SaaS company will eventually become an AI company; so will a lot of non-SaaS companies. Increasingly a lot of buyers are expecting AI in products for personalization and better workflows.”

If she’s right, Merge is building the infrastructure layer that all of them will need.

“An AWS for AI.”

P.S. A big thanks to Shensi for taking the time to talk to us. And all the best to the Merge team for what’s ahead.

PMF /evals ◎ is just getting started. Tell us about how you’re approaching AI-native building in the wild! Or what you’d like us to cover. Hit reply.

We are taking a short break to regroup and will be back in Jan with our usual programming. For this edition, I wanted to leave you with a year-end read you can bookmark and make your way through, slowly, reflecting and visualizing as you plan for 2026 and close this year.

First is Ben Evans’ a16z interview on AI as the next platform shift. Ben is a revered tech analyst and a former a16z partner who does rigorous, data-backed analyses of macro and strategic trends in tech. Worth every minute of the watch.

And then you’ll find a roundup of the AI-native stories we’ve covered so far along with a mention of the themes they discuss, so you can skim easily and pick what hits home.

All yours now.

✨ First, Ben Evans on AI:

» Here is the presentation Ben refers to in this video.

🍿 Now, your roundup, ICYMI:

#1: CURATED PROFILE ON BOX’S AARON LEVIE

{cloud shift vs AI shift ; understanding business model threats; AI at Box;}

“It’s worth noting that for enterprise adoption of AI, it is still going to be a decade-long journey. We’re 20 years into cloud computing and the quarterly rate of cloud computing adoption is still growing at the same rates as ever before. It will be much faster with AI, but it will take time.”

#2: WISPR FLOW’S TANAY KOTHARI

{mindshare as a moat; consumer/prosumer AI-PMF; pricing for the world beyond the valley;}

“I don’t believe technology is a moat anymore. I think even data is not as strong of a moat as it used to be. The main moat that exists today in my opinion is mindshare and user loyalty.”

#3: DOCKET’S ARJUN PILLAI

{data infra advantages; three kinds of evals; doing pilots right;}

“We’ve chosen a consumption-based pricing model. We see consumption as a proxy for value. The more the customer uses the product, the more value they’re likely getting from it. That allows us to align what we charge with how our customers experience and measure success.”

#4: RUNLLM’S VIKRAM SREEKANTI

{selling to enterprises; incumbents’ least common denominator approach; moats in AI;}

“There’s stuff we need to do that’s not cool or sexy or fun, like building a deployment for Salesforce, Zendesk, Intercom and the 25 other support tools that people use because this is what it takes to build a good enterprise oriented product.”

#5: POCUS’ ALEXA GRABELL

{building enterprise AI; finding differentiation; pricing by personas;}

“We sell to CROs and a lot of them only think in seat-based models. So if you’re trying to sell usage-based or AI-based or work-based to a CRO, they’re not going to like it because it’s hard for them to predict.”

#6: SUPER’S CHRISTOPHE PASQUIER

{going multi-product; exploration-friendly pricing; “not a wrapper”;}

“When you try to package too many things in the buyer’s journey it becomes messy and counterproductive… we saw the incredible power of AI search on our knowledge base, knew it would 100x in value with external data, but we learnt from our mistake: it couldn’t be the same product.”

#7: PERCEPTIVEPANDA’S ANDRE VANIER

{pilot failure modes; understanding wedges; canonically great AI use case characteristics;}

“Frequently as founders, we tend to think of our wedge as the small version of the bigger thing that we will eventually do and there’s certainly a lot of truth to that, but it turns out that there is a very, very rich and nuanced history of how wedges have been successfully employed.”

#8: LIGHTSWITCH’S MONICA VILLAR

{consumer vs enterprise monetization; lessons from gaming apps; taxi-meter feeling;}

“Tactics like feature-based tiering, Van Westendorp to assess price sensitivity, and demand curve surveys typically fail when it comes to consumer businesses. The enterprise context is very different from the consumer context… enterprise has whale customers, teams that would pay 100 million, but “there is no $100M Netflix user”.

#9: SUDOWRITE’S AMIT GUPTA

{vertical-AI product thinking; designing margin-friendly free trials; 4 pricing models in action;}

“When you’re writing a novel, each inference call might use 10s of thousands or even 100s of thousands of tokens of input for context. These are very real numbers.”

We’re just getting started. See you in the new year with more stories and more analyses, rigorously reported to help you build in and win in AI. Stay tuned!

Also, do write in and say hi! if you have any thoughts/feedback to share. Cheers.

A 4th pricing attempt that is still not right isn’t a flaw in strategy or a confession but an honest take at the intricate craft of pricing anything AI: constantly evolving, and constantly required to evolve.

Amit is a serial founder, who had earlier built and sold Photojojo, after scaling it to $10M/year. He is also a Sci-Fi writer. Amit is now building Sudowrite with fellow entrepreneur and writer James Yu (YC S11 and founder of Parse).

Sudowrite was a very early entrant in AI, coinciding in birth with ChatGPT’s Model 3 back in 2020. Since then, Sudowrite has managed to garner wide-spread love from users, including prolific writers such as:

This edition of PMF /evals will feature notes from our chat with Amit where we peel back the layers on:

• 👀 the founding insight behind Sudowrite and why Amit chose fiction writing as a problem to take on,

• 🧱 the thoughtful tech decisions that enable their incredibly high quality of output,

• 🏷️ the nuanced evolution of their pricing model going from flat to credits spanning 4 different, in-time iterations,

• 🚴 how they designed an AI-enabled free trial that doesn’t burn a hole in their inference pocket, and more.

Read along and if someone forwarded you this newsletter, don’t forget to subscribe here.

#1 ➤

Building where foundation models fail to deliver

Amit recalls ChatGPT’s model 3 launch in 2020 and the frenzy it birthed, with the world flocking to build wrapper assistants for business use cases like marketing copy. Amit and team saw it but decided to take on a different, much taller challenge: building an AI assistant for fiction writing.

There were a couple of reasons behind this. One, Amit found the challenge of building a copy assistant entirely uninteresting. Two, Amit knew that the simplicity of the problem was going to inevitably attract a lot of competition.

Fiction was different.

It was both personally inspiring for Amit as well as complex enough to warrant building a thoughtful solution. “With fiction, there’s a large gradation of subjective quality. So we felt that there was more interesting work to be done here. You can’t ever get it right.” Amit shares.

But generative writing is a use case that foundation models seem to cover really well. So we had to ask: “Why do we need another tool? Why aren’t the foundation models sufficient?”

As a writer himself, Amit had spent a lot of time tinkering with the foundation models and had often walked away feeling both underwhelmed by its capabilities as well as frustratingly confined by the constraints.

First he ran into out-of-touch censoring. Fiction stories often involve mature themes that model makers consider taboo and censor.

“We quickly realized that this was going to be a huge problem for writers. They are used to writing in MS word, and MS never censors them.” Amit adds.

Then he ran into what could only be described as bland creativity.

Most of the foundation models are good at writing memos and emails, but they’re not good at writing dramatic narratives.”

Writers thrive off of reaching for a personal tempo, the best that foundation models could offer still felt way off compared to what writers wanted.

Sudowrite has a set of deep features designed to tackle this exact swath of problems: it allows writers to go as far as defining tone, setting up characters, choosing models of their choice for different steps, and feeding their own understanding of different writing elements into the tool. Foundation tools don’t offer any of these capabilities.

#2 ➤

A “non-generic” assistant: The tech decisions enabling Sudowrite’s high-quality output

Getting to this level of quality has been far from simple. Today, Sudowrite deploys a two-layered system, comprising a prompting layer and a model layer, with dedicated efforts in improving and innovating on both.

On the model layer, Sudowrite uses a few dozen models, all of which have been thoroughly evaluated for their specific edges/advantages and then brought in.

“Some models are better at extracting salient information from a user’s story bible, which is an area of Sudowrite that gathers the core elements of your story and acts as a source of truth as your work develops. Others are better at what we call the tip of the pen, which is actually writing the first draft prose as the final step. Each step along the pipeline, from a user input to a final version that the user likes, calls on multiple different, specialised models.”

This model-scape is how Sudowrite is able to deliver on quality.

In addition to these pre-selected models, they also allow users to choose models of their own preference for different steps of the process, something that Sudowrite’s users seem to have a strong interest in: “People who are just getting into writing with AI can’t be bothered to learn about different models, but those who’ve been working with this stuff for a year or longer often have favorite models.” Amit adds.

Amit and team are also doing a lot of the work in allowing authors to train models specific to their voice. They recently opened a beta program where authors could upload work to train a model just for their own use and writers have been lapping it up.

These user insights have reaffirmed Sudowrite’s original premise that fiction writers want non-generic assistants, going as far as wanting a model that can be trained to assist them for every step of their unique writing process.

#3 ➤

From flat fees to credits: Evolving through 4 pricing models

“This is the first time in SaaS that you also have real marginal cost to deliver the final product if you use AI. The era of zero marginal cost SaaS might be over.”

Amit opens his notes on pricing by acknowledging the elephant in the room, AI has changed the whole script of pricing software.

Today, Amit and team are on their fourth attempt at a pricing model, and he believes it’s still not right.

They started with a flat, $5/month — which was later changed to $10, then $20 — with no specific stated limit on usage. It soon started showing cracks, with users paying $20 a month while costing them $400 a month.

That’s when they decided to switch to AI words in an effort to find a way to correlate the value users were getting from the product with what they were paying for it. “It was crude because words aren’t really ‘value’, but it was as close as we could get in a meaningful way.” Amit shared.

Soon enough, as Sudowrite went multi-model, they had to switch again, this time from words to credits.

“With the level of flexibility that came with offering different language models, and the fact that we let users develop their own custom tools in Sudowrite, using whatever prompts and LLMs they want, it meant that someone could use the most expensive models at each step, inhaling multiple novels as input, and only output a single word, one word that cost us $30 to make. That was not sustainable. We couldn’t charge by words as an output metric.”

This usage-credits based model was their attempt to further align their costs with what they were charging customers. It also allowed them to offer greater flexibility and customization, as well as lower costs for the vast majority of their users. But it came with its own challenge: it was confusing.

The way Sudowrite is built, they charge different amounts of credits based on different models you use at different steps along the way, which meant that it was really hard to predict exactly how much pressing a button would cost you.

To help with the confusion, they decided to offer a one-week easy cancel policy. In addition, they also run coaching and education efforts to make sure that users are able to understand their credits and their implications as well as possible.

One useful parallel that Amit goes to is the apparel industry. Where, again, choosing the right size (plan) of t-shirt solely from photos is quite difficult.

“Ultimately, the solution that worked is that you can buy the t-shirt and you can exchange it if it’s not the right size. I think that’s the kind of solution where we’ve landed too. You can return or exchange any plan in a week if you buy it and it doesn’t fit.” Amit added.

Note: Check out this LinkedIn post by Chargebee’s CS leader Madhuri detailing the strategies that Fin (by Intercom) uses to ease difficulties around their outcome-based model priced on AI resolutions.

#4 ➤

Designing AI-enabled free trials that don’t burn a hole in your inference pocket

“All of our plans include free trials. We have had a similar journey of iterations with our free trials as we did with our overarching pricing models.”

Amit and team started with a flat 2-week trial, and then switched to a AI-words-first trial as they ran into the familiar set of challenges: of finding a balance between value for the customer and cost to the business.

Eventually they switched to a credits-based free trial where users get a limited number of credits to use in their trial.

Amit’s philosophy behind this was clear: They wanted to give people a taste of the product and time to see if it’s right for them but they also had to be mindful of the cost of servicing these trials. Amit shared that at one point a huge percentage of their overall inference costs was going towards servicing trials.

“When you’re writing a novel, each inference call might use 10s of thousands or even 100s of thousands of tokens of input for context. These are very real numbers.” Amit adds.

So they decided to cap the trials with a credits limit and make them shorter.

Interestingly enough, they saw that shorter free trials converted better. Amit’s hunch was that it created a sense of urgency and their ideal customer, professional authors, were serious about making a buying decision.

Amit and team have also made a deliberate effort to invest into education to ensure that users are enabled in their exploration and the business sees better returns from their free trial.

“We’ve taught classes for a few years now; we teach at least one class every day and we have different topics throughout the week. It’s a good way for people who are still learning the program to get up to speed without having a long trial potentially.”

They have also started experimenting with one-on-one onboarding where users can schedule a time with them and actually learn how to use the program with an expert over 20 or 30 minutes and then get more out of the trial.

Amit believes that creative tools can get complex, and educating people on how to use them and how to get the most out of them will become increasingly important.

#5 ➤

“Not for everyone”: The further-ordained position of pricing in finding PMF

Amit thinks of pricing as a fundamental lever of their product-market fit — as much about defining the ideal customer as it is about staying approachable.

Last year, they did a deep dive to understand who gets the most value from Sudowrite over time. They discovered that their core users were professional authors, people who’ve written more than one book and essentially earn their living through writing.

For these writers, Sudowrite was an easy purchase.

“It could probably be three times more expensive and still be valuable because it helps them write much faster and produce better books.” Amit found.

On the flip side, they saw that the cohort of aspiring writers and students considered $10-20 a month prohibitively expensive. This gap meant that they needed to pick a lane, there was no way to have a price point that could serve everyone.

It also gave them more clarity about where their PMF was going to come from, which then they doubled down on: “We’re focusing on being the highest quality solution in our segment, which means we can’t serve everyone.”

#6 ➤

Why “everyone will vibe-code their own software” might be an exaggeration

There has been much too much chatter about how vibe-coding could eat away SaaS or that people might start building their own software. If you’ve been on X at all, you know.

Amit has a more tempered take on this. He believes that as software becomes easier to build with AI, people will make more software for themselves. But the growing notion of personalized, bespoke software might not be as big as people think. He draws an analogy to help make sense of why:

“There was a period where personal publishing was really big and everyone published their own blog. Then it became easier to tweet and post on Facebook, and because it was easier, that’s what people did. Most people didn’t want to do the work of setting up a blog and trying to get readers even though it’s trivially easy.”

He thinks the same will be true of software for most people…

“If a solution exists and it’s good enough, they’ll prefer to use that than to create their own software and make the hundreds of decisions that go into making a piece of software.”

That said, Amit shares that there will be tremendous potential in allowing people to customize their software.

Someone could very well be like: “Oh, I hate how my video conferencing software works. I wish it had feature X. I’m not going to build a whole new Google Meet myself just to get X. But if I could prompt it and add that feature easily, I might customize it that way.” Amit adds.

P.S. A big thanks to Amit for taking the time to talk to us. And all the best to the Sudowrite team for what’s ahead.

PMF /evals ◎ is just getting started. Tell us about how you’re approaching AI-native building in the wild! Or what you’d like us to cover. Hit reply.

It is fairly recent that category-winners like Spotify, Snapchat, and Netflix have started experimenting with and championing another possibility: “stacking monetization models”, which is increasingly becoming a feature of winning companies according to Pete Flint, GP at NFX Ventures.

This innovation comes from a curse that has traditionally plagued consumer apps: poor revenue retention — where companies make significantly less from a cohort of users over time, unlike enterprise plays with a strong possibility of revenue expansion.

“The best subscription businesses retain 65%+ of their revenue after one year. Retention above 50% is considered in the top quartile and 42% as the median. Many companies hover between 40–45%.”, shared Parsa Saljoughian, VP of Strategic Finance, Whoop in this benchmark article.

AI is beginning to pierce through these constraints, and fast.

Olivia Moore, AI partner at a16z, shared some hard-to-ignore trends in a recent article pointing to this shake-up:

• AI has led to an explosion of the consumer appetite for software products, with consumer spending increasing 200% YoY to $1.1 billion,

• Companies are hitting 100s of millions in ARR in unprecedentedly short timeframes — outpacing even enterprise counterparts,

• And with new components like usage-based billing, revenue retention is evolving too, with some teams exceeding 100% (making more from a cohort of users compared to last year).

These are tectonic landscape changes. But a new challenge has emerged in tandem: even winning apps are hemorrhaging money. ChatGPT — the unquestionable leader in consumer AI apps — lost $5 billion last year while making $3.7 billion in revenue.

Even once the inference costs have gone down enough, monetization will remain the lever to pull: for retaining strong margins, continually grabbing market share, and surviving on good terms.

In this edition of PMF /evals, we will dissect consumer app monetization further with insights from Monica Villar of LightSwitch. Monica is an ardent student of economics, has been a part of subscription teams at Apple, American Express, and Nicklpass, and is now a founder helping folks in the thick of the consumer monetization problem.

✨ Brought to you by Chargebee. Chargebee helps AI-native, recurring revenue businesses scale with billing and monetization infrastructure built for speed, flexibility, and rapid iteration.

#1 ➤

Enterprise has intentionally tried to figure out value mapping — consumer hasn’t

In the context of consumer software, the flat access subscription model has one big flaw: it creates a massive value gap.

When you implement a flat access subscription, you end up with a massively skewed cost-to-value ratio for power users vs. everyone else.

Enterprise software is fundamentally different in this regard — when you sell to an SMB and a Microsoft, you can charge Microsoft a Microsoft amount of money.

This is exactly why, it’s the enterprise world, where there’s been a renaissance in pricing thinking/innovation over the last few decades:

#Act 1: Tiered, seat-based pricing ▶ Recognizing that a 10-person team and a 1,000-person Microsoft team derive vastly different value from the same product. Tiers were the first attempt to map value to different user segments.

#Act 2: Usage-based pricing ▶ Going beyond fixed tiers to charge based on actual consumption. This allowed even better flexibility in matching what you charge to the value delivered.

#Act 3: Outcome-based pricing ▶ Charging for results, not just access or usage. The AI era’s latest evolution.

#Act 4: Credits, agentic pricing, and hybrid models ▶ A flowering of new models, all with the common goal of deliberately mapping value to different kinds of users, at different levels.

“In the enterprise context, we have all collectively realized that flexible models — whether it’s usage based, outcome based, any variation — all comes back to the same premise that your customers are not one size fits all and that you need to use pricing as a mechanism to very efficiently map the right type of value to a customer.” — Monica captures the essence succinctly.

But in the consumer context, where there’s no 1000x variance, you need to think entirely differently. Unfortunately, that conversation is terribly underdeveloped right now:

“This thinking doesn’t remotely exist for this whole universe of software that only sells to individuals.

How do you monetize your product and grow a sustainable business when there’s no expansion revenue, no scaling up within organizations, and none of the traditional product-led growth tactics apply?

The internet needs new business models.”

It is beginning to happen in certain pockets, especially at scale.

Snapchat, the social platform, had monetized exclusively through ads for the longest time. They recently added a subscription offering called Snapchat+ which grew to 14 million subscribers in 2024, generating over $500 million in annualized revenue with 131% year-over-year growth.

Then there is Netflix which started at the other end, exclusively with subscriptions, and has very recently expanded to ads. The impact looks overwhelmingly positive: Netflix’s ad revenue is on track to increase from $1.4 billion in 2024 to $2.9 billion in 2025 and increase another 45% to $4.2 billion by 2026.

These numbers are a result of the deeper, intentional changes happening at Snap and Netflix, where they have started optimizing for ARPU (value from existing users) rather than just user growth.

Snap explicitly stated increasing ARPU as a goal in their earnings. Netflix has shifted from reporting “subscribers” to “Monthly Active Viewers” to capture the impact of their ad-monetization and engagement-expansion strategy.

Circling back to Pete Flint’s assessment, no matter where they started, consumer apps are seeing an important pricing evolution: stacking monetization models to capture more and more value.

#2 ➤

Whale Markets vs. Whale Customers: Monetizing in consumer vs enterprise

Monica says that even though enterprises pioneered monetization innovation, consumer apps can’t apply their tactics unthinkingly:

“Tactics like feature-based tiering, Van Westendorp to assess price sensitivity, and demand curve surveys typically fail when it comes to consumer businesses. The enterprise context is very different from the consumer context.”

Monica illustrated why: When an enterprise buyer lands on your pricing page, they come with a sense of the solution. When they see tier-based options, they can self-select and eventually expand organically.

For consumer businesses, Monica has seen that when products go from flat access to pricing in tiers, they lose money because it overcomplicates the buying decision for their audience. Fewer people convert when you overcomplicate the decision to pay.

Similarly, another popular enterprise exercise that doesn’t make sense for consumer products is figuring out the maximum willingness to pay across different segments.

To be clear, willingness-to-pay research itself isn’t bad, every consumer app should understand what their market will pay, just not the way enterprises do it. It doesn’t make practical sense to invest time into enterprise toolkits like Van Westendorp surveys, because the variance simply isn’t there.

It was transformational for enterprises because of the real variance in what you could charge customers. Enterprise has whale customers, teams that would pay 100 million, but “there is no $100M Netflix user”.

What Monica recommends instead is to focus on finding and capturing whale markets.

She believes that when selling to individuals, the alpha lies in getting more people to go from zero to something, not in figuring out how to get a paying customer to pay you more.

Take v0’s approach: they’re explicit about “building for the next 100 million builders”, a whale markets strategy. Even though more than 50% of their revenue comes from enterprise, their core goal and pricing strategy (freemium + $20 tier) is about reaching massive numbers of individual users.

The enterprise revenue is a natural consequence of individual adoption. Something that Olivia Moore has seen be true for a fair share of prosumer productivity startups.

Even Netflix’s metric evolution reflects this thinking. By moving from “subscribers” to “Monthly Active Viewers,” they signaled a focus on reach and engagement across their base, not just subscription expansion.

Interestingly enough, Eric Seufert of Mobile Dev Demo speculated and suggested that ChatGPT could actively consider monetizing their whale market with ads:

#3 ➤

Why subscriptions have stuck around, for better and for worse

Even the pioneer in AI, OpenAI, is trying to get their feet under them when it comes to monetization. Monica speculates that “they started with subscriptions because subscriptions are the least controversial, most legible to consumers, and well-understood.”

But the fact that Sam Altman is publicly pooling in ideas to build better subscription offerings is telling of the fact that subscriptions as they exist today are not going to work for these new businesses.

Going back to unthinkingly adopting enterprise tactics, even ChatGPT has been hesitating with rolling out usage-based for their consumers. Monica thinks there are a few lessons here to remember:

One, new unfamiliar models can introduce friction in converting non-paying users into paying customers. It can be daunting to put your card on something you don’t fully understand.

Two, usage-based or credit-based models can create a “taxi meter feeling” where consumers become overly conscious about “misspending” credits — curbing their exploration of the product. Something that Super’s Christophe Pasquier and Wispr Flow’s Tanay Kothari had also made a h/t to in their interviews.

“You get hooked to Netflix when you watch a bunch of things, you get hooked to Spotify when you listen to a bunch of songs, you get hooked to ChatGPT when you use it a bunch. By charging users for these actions, you are actually disincentivizing users from engaging with your product.” — Tanay Kothari

Monica shared something very similar, using Netflix as an example:

“When you subscribe to Netflix, you’re not paying to “watch a hundred movies” or any specific content. You’re paying for unlimited access to an entire library over a discrete period of time… The brilliance of this approach lies in the psychological effect it creates: after paying my monthly fee, it’s now on me to extract as much value as possible by consuming whatever content makes sense for me.”

This behavior is also what drives you to renew month after month.

#4 ➤

Beyond software: Lessons from gaming apps

A popular way a lot of consumer apps are thinking of monetizing cleverly is by adding enterprise features, as shared by Olivia Moore in her a16z article.

But there’s one important caveat: all of Olivia’s examples from the a16z article (and other consumer apps stories) are prosumer/productivity tools with enterprise revenue bridges, not pure consumer plays:

• Google AI, Krea, Grok → General AI assistants used by both consumers and businesses

• Gamma, Replit, Cursor, Lovable → Coding/productivity tools with clear enterprise applications

• ElevenLabs → Voice generation with HIPAA compliance for enterprise

These are tools that individuals use first, then bring to their workplaces, where 50%+ of revenue often comes from enterprise. This is fundamentally different from pure consumer apps like companion AI apps or entertainment products. For consumer app founders, Monica recommends thinking beyond software, specifically gaming apps.

She shared that contrary to what many assume, 3 of the 5 top-grossing consumer apps on the playstore are actually games. TikTok is at the top ($2.3B), then Google One ($1.4B), and then three games consecutively: Last War: Survival ($1B+), Royal Match ($915M), and Whiteout Survival ($941M) in 2024.

Monica goes on to share the crux of why games have been successful at maximizing ARPU:

“Mobile games maximize ARPU through a highly retentive free experience. This embodies the whale markets thesis: engaged users who interact deeply with the product get presented with low-commitment purchase opportunities. Multiple conversion chances exist as long as users stay engaged.”

Let’s take the game, Last War: Survival as an example. We will dig in to understand the sophistication of their monetization and why it managed to generate a whopping $1.5 billion in revenue (#3 in consumer apps) in 2024 despite only scaling in late 2023.

The game is completely free to play at its core, but layers in multiple monetization methods that feel natural to the experience. Here are the strategies it uses:

• Tiered conversion strategy: Low-cost entry offers ($1-$3) that provide significant value, making the first purchase easy.

• Dual acquisition approach: Combines gacha mechanics (miss-or-hit/slot machine rewards) with direct hero purchases, giving players choice.

• Premium passes: Battle pass-style systems that offer hero shards and unlock additional purchase options.

• Multiple monetization layers: Premium currency for progression, speed-ups, exclusive content unlocks.

• Event-driven monetization: Limited-time offers and bundles that create urgency without feeling pushy.

It’s a mechanism where spending ties seamlessly with enabling convenience and progression instead of feeling like a forced-upon extractive measure. Power users have clear, valuable reasons and ways to spend more.

“Compare this to current subscription app trends (long onboarding, hard paywalls) that make purchasing high-commitment and one-shot.” Monica calls out the limited view that subscription-only monetization can adopt.

Monica closes by stating a timeless reminder and an idea that will sail us into the future:

“The art of selling to people has existed for centuries, and there’s tremendous amount of wisdom to be drawn from industries like consumer package goods and telecom that excel at it.”

P.S. A big thanks to Monica for taking the time to talk to us and a h/t to all the sources and people we referenced/quoted to develop this story. Takes a village, indeed. :))

PMF /evals ◎ is just getting started. Tell us about how you’re approaching AI-native building in the wild! Or what you’d like us to cover. Hit reply.

Not a clever application or an incremental improvement, but a marvelously strong fit with a problem, constituting all the right elements to get to PMF.

If you look at the breakout successes in AI — GitHub Copilot, Intercom Fin, Fireflies, Midjourney — they share a few unmissable attributes:

1. Inverse Alignment: AI’s strengths map perfectly to the problem’s hairiest constraints.

2. Exponential Economics: The economics don’t just improve, the efficiency curve bends by orders of magnitude.

3. Erased Tradeoffs: A long-accepted pain or compromise suddenly disappears.

User research is a textbook example of this. “A canonically great AI use case” in the words of Jack Moshkovich, a partner at 8VC. Let’s see why:

“Traditionally, teams had to choose between high-fidelity but slow and expensive interviews or much lower signal options like basic surveys,” Jack shared in a post announcing their investment in the vertical.

AI erases this tradeoff.

Moreover, the problem was never a lack of intent or effort, but three unrelenting bottlenecks: real-time people-dependence, time commitment, and scheduling roulettes. These also happen to (perfectly) inversely align with AI’s most promising capabilities.

Here’s an illustration from Greylock condensing this:

These aren’t just investor notes, either.

Andre Vanier, co-founder and CEO of PerceptivePanda, and a product leader with two decades of user research experience, adds a first-hand observation specific to subscription businesses: “If you are running a subscription based business, it is really hard to understand why free trials don’t convert and why existing subscribers churn, especially if you have a scaled customer base.”

AI also inverses the economics here, allowing you to run these insight-yielding interviews, but at a fraction of the cost. Now they scale.

This fit is also what prompted Andre, a three-time founder and former product executive at Yahoo, to build in this space. At PerceptivePanda, he co-founded an AI-native user interview startup with pilot customers such as Koala, Outseta, and Joy.

We met Andre at Chargebee’s annual conference Beelieve, got talking, and followed up with a longer conversation to understand what he’s seeing in the field.

In this edition of PMF /evals, we borrow from Andre’s experience to understand how AI is transforming user interviews as a vertical.

Key insights we dig into:

• Why they use a deterministic layer of control on top of LLMs,

• Pilot design tactics to de-risk getting deprioritized, and the traps in the Valley’s wisdom of finding 2-3 customers who love you,

• Why the type of your wedge deserves as much attention as its size, and

• Giving away 500 AI interviews to encourage company-wide adoption.

#1 ➤

Why they use a deterministic layer of control on top of LLMs

“If you ask a generic AI to interview someone, the result wouldn’t be just kind of ok, it would be fundamentally discordant.”

Awkward at best, and maybe even off-putting. Andre has seen that most LLMs, in their raw state, would root on one particular topic and ask 10 to 20 questions in a go. “AI wouldn’t know when to go deep vs. when to circle back and go broad.”

This is because LLMs don’t understand the unspoken conventions of human speech, they aren’t sufficiently capable for this use case by themselves.

One would wonder why not just prompt a basic LLM to “act appropriately” and add some other context. Andre says that it doesn’t work because the LLMs don’t know what these nuances of interaction mean, or what “appropriate” really looks like.

So they approached the problem from a fundamentally different angle.

Andre’s co-founder Alex Djalali, who comes with deep NLP expertise, engineered a proprietary control layer to govern the AI as it conducts interviews. This control layer modelled the state of the dialogue while using APIs to orchestrate LLMs for follow-up questions.

This has helped them leverage the fluidity and dynamic nature of an LLM while combining it with the predictability and guardrailed nature of a deterministic system. This guardrailed system is also their edge.

#2 ➤

Pilot design tactics to de-risk getting deprioritized and protect against false positives, and biases

Much like what RunLLM’s Vikram Sreekanti had shared in his interview, Andre believes that the real investment in the pre-ROI stage of a pilot is time and trust.

So they wanted to pick customers that shared their conviction about where the world is going, the weight of the problem they are trying to solve, and PerceptivePanda’s ability to solve it.

There are a few things Andre wishes he had done differently during the pilot stage.

One, doing fewer pilots but networking more deeply with the leadership teams. Andre saw that B2B employment has a very high churn rate and unless you work directly with the founder or C-Suite, you run the risk of getting deprioritized or put on pause in case your champion leaves. To de-risk, you need more than one champion.

Second is a tactic that’s contrary to the popular advice of getting two or three customers to absolutely love you and then worrying about expanding. Andre saw that this can make you susceptible to biases that can prove fatal for expansion.

He realised that there are a lot of idiosyncrasies in every market and if you focus on just two to three companies, it’s possible that you end up with a solution/approach that does not generalize to other segments in the market.

He also warns against over-indexing on your network for similar reasons. In his experience, your network trusts you because they know you but the feedback you get from them might not be entirely generalizable.

What do you do in these cases?

Double or triple the number of pilots? Andre says that 3 vs. 10 is still statistically meaningless. He recommends staying aggressive with research work in parallel, instead.

What Andre did was actively explore market signals. They invested in listening to their current customers, potential customers, and even people who were unlikely to be customers, all to get a complete sense of the problem space.

These signals were: kinds of solutions these companies were evaluating, how laborious the existing processes were, and the ability of customers to quantify/articulate the benefit of moving to a better approach.

For this, he recommends supplementing your network by using userinterviews.com where you can access a large panel of folks based on industry title, functional expertise, etc. and interviewing them. This guards you against selection bias.

#3 ➤

Giving away 500 user interviews to encourage company-wide adoption

Andre and team have leaned strongly into the product-led growth ethos; they use price as a strategic lever to impact customer acquisition cost. This is based on a few convictions they have about the future.

One, they have a belief that the computational cost curve is going to continue to get forced down. While the inference costs on the deep thinking models has been rising, the LLM side of the equation has continued to get cheaper. And that’s what matters to them.

As a product that deals with real-time interactions where delays are not tolerated, the cost of deep thinking models is a non-concern for them.

On the pricing model, Andre and team have a usage-limited free plan model where the first 500 AI interviews are provided at no cost.

Their mindset here is similar to that shared by Super’s Christophe Pasquier and Wispr Flow’s Tanay Kothari — they want to prioritize habit formation and value creation over short-term cost optimization right now.

Andre is doing this by shifting their customer acquisition expenditures from ads, marketing, or other forms of distribution and putting that money back into the pocket of their customers. Not everyone can or should use this strategy, but it is a bet that they think is prudent.

Another reason for this is a focus on horizontal expansion. They aspire to ultimately enable customer 360 which is a holistic view of a customer that integrates data from a variety of touch points with a company, and not just one function.

When a team gets a generous 500 free interviews, more functions in a company can try their product and discover relevant use cases without feeling cost-constrained, this can improve pilot-to-paying conversion and retention.

Not just a hunch, they have already seen instances of horizontal expansion within orgs because of this approach.

#4 ➤

Why the type of your wedge matters as much as the size of your wedge

“Frequently as founders, we tend to think of our wedge as the small version of the bigger thing that we will eventually do and there’s certainly a lot of truth to that, but it turns out that there is a very, very rich and nuanced history of how wedges have been successfully employed.”

Every time there is a mention of a wedge, the narratives tend to heavily focus on its size. But one of the more important, often overlooked nuances is the type of the wedge.

Andre makes a h/t to the “12 Killer Wedges for Your Company’s Growth” article by NFX Ventures’s Pete Flint — a deep dive into each of the 12 basic wedge types that have been successfully used by different founders.

The wedge that Andre and team focused on is what Pete Flint calls the 10X ROI wedge. Here’s a description from the article:

“The “10X ROI wedge” works if you can instantly increase ROI or meaningfully cut costs. That’s going to give you instant traction. Many companies make this claim. Few are able to deliver. But those that do, grow fast.”

Andre and team focused on demonstrating a financially measurable benefit to customers so that they could easily stack-rank them against other priorities.

These ROI chats were based on tangible metrics such as conversion rate improvements and percentage gains in pipeline conversion.

This also points back to Andre’s focus on working with customers with conviction, as well as helping them build this conviction during the pilot phase.

P.S. A big thanks to Andre for taking the time to talk to us!

PMF /evals ◎ is just getting started. Tell us about how you’re approaching AI-native building in the wild! Or what you’d like us to cover. Hit reply.

Brought to you by Chargebee. Chargebee helps AI-native, recurring revenue businesses scale with billing and monetization infrastructure built for speed, flexibility, and rapid iteration.

“Notion launched competitors to Granola, Glean, and ChatGPT use cases

Figma launched competitors to Lovable, Framer, Canva, and Illustrator

Atlassian has launched competitors to Granola, Glean, Claude Integrations and more

Google just launched Prototyping and Codex, competitors

The list goes on…”

Brian Balfour, founder of Reforge, and a prolific writer on the business of tech recently made a stirring observation: we are officially in the era of everyone competing with everyone.

Point solutions, he believes, are under imminent threat. Bundling is happening all across, everywhere. And the PMF expansion (or second acts :)) blitz is accelerating as we speak.

Christophe Pasquier, co-founder of Slite (YC W16), in the meantime, has launched a new product that looks a lot like a point solution — BUT with a critical overlap with Brian’s description of winning startups: they too have bet on the AI flywheel of more data, albeit in an unconventional way.

Chris’ team turned what started as a feature into an independent product called Super, sold both with Slite and on its own. It does AI search for company data and extends on Slite’s mission to make work simpler and easier for teams.

We’ll dedicate this edition of PMF /evals to the key moments punctuating this decision, primarily:

1. the pricing challenge, market opportunity, and packaging confusion that became the bedrock of Super’s independence,

2. why they chose a seat-based + fair-usage pricing philosophy against a pure-play usage-based model, and more.

It all started when Chris saw ChatGPT and felt pulled to build it for private knowledge. This birthed the first AI feature within Slite called “Ask”.

Users loved it, but Chris had a hunch that they might need to expand and connect with more tools, including competitors like Notion and Confluence, but that was for later. They revisited the hunch once Ask had matured. “It was something that we wanted to pursue and put heavy investment into, if nothing, to just see if it can work or not.” Chris said.

Very wary of diluting the team’s focus, they created a commando team with Chris and 2 other people who enhanced Ask and built AskX (the new version). By December 2024, AskX was ready to be sold.

This is when they started running into a set of “pricing and real estate” challenges that convinced them that AskX warranted an independent existence — as Super. Chris goes deeper into these challenges below. 🖌️

#1 ➤

Three factors that shaped Super’s independence: Packaging, market opportunity, and pricing

“When you try to package too many things in the buyer’s journey it becomes messy and counterproductive.”

Even though both Slite and Super were technically solving the same problem, users approached them in markedly different ways.

Slite fell into the category of a knowledge base, but Super was a universal search or AI assistant. So as much as the products aligned, the buyer’s intent didn’t.

“We saw the incredible power of AI search on our knowledge base, knew it would 100x in value with external data, but we learnt from our mistake: it couldn’t be the same product.”

“We had already tried packaging features that were not really connected.”

Then came the too-big-to-ignore market opportunity.

If they supported non-Slite customers, the market would be massive. And they could: “We saw that we could deliver incredible value even if people didn’t use Slite”.

This unlocked a whole new user base for them. “Most of our customers now don’t have Slite.”

It wasn’t a part of their original plan, but it made sense. The reason is captured further in this post that Chris recently shared while launching support for Notion, a Slite competitor:

“You’d have told me we’d build a tight Notion integration 18 months ago, I’d have burst laughing… Slite is a direct Notion competitor…

but I respect Notion’s craft a lot, they built something special for a part of the market, and now a lot of knowledge is stored in their tools. Our entire purpose is to break team silos and solve access to team knowledge. With Super we could help Notion users, so letting them use Super felt like a no brainer.”

Now pricing…

Slite had always delivered an enviably strong bang-for-your-buck and Chris had a strong opinion that the search capability was table stakes to this promise. So, even with the upfront AI costs, they initially offered it at their regular price as a part of the deal.

Soon the costs started spiking and they knew they had to find a middle ground. Especially for the better models that costed them too much.

“We had a pricing model for the knowledge base and it was quite hard to include good LLMs in that price. It just cost too much.”

The quintessential AI pricing problem had hit them.

They decided on a win-win. All Slite customers get the base AI Search at no added expense. But for teams that want advanced features like external sources, customer assistance, automations, expensive models, they have the option to buy Super as an add-on.

Slite customers can get Super by paying $25/m/seat, which comes with a fair-usage limit on the number of queries they can run.

#2 ➤

Pricing mindfully for their stage: Going seat-based with fair pricing to encourage exploration

All AI tools are figuring this one out. Usage-based, outcome-based, seat-based… the frenzy is understandable, even underrated.

There is no “right” model, no playbook, but this choice affects your PMF journey more than it ever did — your margins, your adoption, your chances of standing out amidst the current rush.

Chris has been clear on this: it’s important for them to enable exploration. “We really want your team to not even have to think before using Super.” They want a pricing model that lets users explore freely (something that Wispr Flow’s Tanay Kothari makes a case for as well), find use cases, and form stickiness — all of this while ensuring the business is protected against bad margins.

So they are keeping it simple. Their current pricing model is seat-based with fair usage. Products like Fin by Intercom have popularized outcome-based pricing and some others like coding agents have leaned into usage-based models.

Chris thinks it is a good strategy for those orgs, but they wouldn’t use it at this stage. He believes usage-based, especially, can heavily limit exploration and create a barrier to usage and that’s the last thing they want.

He draws an internal use vs. external use comparison here. When external users interface with a product, they use it as they’d like, no holds barred. But internal users hesitate. When interacting with a usage-based product, they become extremely conscious of tokens/credits spent and start curbing their exploration.

This is bad for a product trying to build a habit.

“If ChatGPT cost $1 per request, we would not have discovered 90% of the use cases it could solve.” he adds. Interestingly enough, encouraging this kind of education/exploration of use cases was a big focus even for ChatGPT, as Krithika S, Ex-VP of Marketing, OpenAI had shared in a Lenny’s Podcast episode.

“When you think about all of the different stages of the funnel, awareness was clearly not the problem that ChatGPT or OpenAI had. Everyone knew of ChatGPT, but when you clicked one zoom level further, the thing that came up was, ‘I don’t know what to use it for. I don’t know what it replaces. Should I be using search for this? Should I be using ChatGPT for this? How can it even help me?’ And so the work of marketing ended up becoming, creating this sort of use case epiphany where people could say, ‘I had no idea ChatGPT could do that. And yeah, maybe I should be using it for X, Y, Z reason in my own life.’”

For Super too, this strategy did what was intended. Like in this one instance of unprompted exploration, where they recently had a customer use Super for the use case of performance reviews, which is not their primary use case at all but still has a massive market.

AI inference costs are real, though. Which is exactly what the fair usage element seeks to solve.

The “limit” component helps them with two things. One, it protects their margins. And two, it safeguards against platform-disrupting behavior by any one user. “We want to make sure indexing performance (how fast your sources are added) is never broken because 1 customer decides to index millions of docs every day.”

There is a fixed number against these limits under different plans. We asked Chris how they got to it, he shared:

“We compute the cost of storage, and the costs of our queries with best models out there (we assume for general, non-reasoning ones that the cost will remain similar while quality raises), and look at usage to make sure we have something non blocking for 99% of cases.”

That said, there might be a future where once Super has become sticky enough, along with a more exhaustive set of tools and use cases, that they may move to credits-based pricing. It stands to be seen.

#3 ➤

To close, here are some notes from Christophe’s LinkedIn expanding on the ideas we’ve discussed:

1️⃣ ☇ Dissecting the “it’s a wrapper” dismissal:

2️⃣ ☇ “You can’t do unlimited AI for $10/m” — and how a thorough pricing model can protect your margins from evaporating:

A big thanks to Chris and Ishaan for taking the time and collaborating on this story!

PMF /evals ◎ is just getting started. Tell us about how you’re approaching AI-native building in the wild! Or what you’d like us to cover. Hit reply.

Brought to you by Chargebee. Chargebee helps AI-native, recurring revenue businesses scale with billing and monetization infrastructure built for speed, flexibility, and rapid iteration.

When Alexa Grabell set out to build Pocus in 2021, she ended up building a whole new category. PLS, or product-led sales had been around but it was far from ubiquitous.

Alexa knew she was signing up for a tall order. So they decided to bet on a unique approach: building a community, a category, and a product at the same time.

“There would be no way we could build this robust of a product in this short of a timeframe if we didn’t do the category and community part.”

It worked and Pocus signed on customers like Canva, Linear, Loom, and Superhuman.

Now, Alexa and the team are setting out for another tall order. Building an evolved, AI-enabled Pocus and taking sales tools “from signal aggregation to intelligent guidance”.

Not a sci-fi shake-up of sales tooling or an all-sentient AI SDR, but a grounded sales solution that combines the best of Pocus 1.0 with the new-found super capabilities that AI can offer.

We sat down with Alexa (again) to borrow from her hard-won lessons from preparing for this evolution. Hopefully this will be useful for everyone navigating the pre-AI to with-AI journey. ↓ A TL;DR of some of the key insights that surfaced:

• Why 90% of AI tools are built wrong for enterprise,

• How Pocus is confidently differentiating from the incumbents,

• Why jumping to “using AI” is counterproductive as a builder, and

• Navigating AI pricing for a buyer persona that is too used to the predictability of the seat-based model.

🚫 Why 90% of AI tools are built wrong for enterprise

While a lot of AI startups are either focused on building for individuals or building point solutions for a single workflow, Pocus is taking a different approach. They are building for the entire sales teams at enterprises.

The reason lies subtly hidden in what made B2B explode in 2010, as Alexa shared in a LinkedIn post:

“There’s a reason B2B SaaS exploded in the 2010s. Companies realized that individual productivity tools (think: personal Excel sheets, local databases) were creating chaos at scale. The winning B2B products?

They built for teams first:

- Salesforce replaced individual contact lists

- Slack replaced person-to-person emails

- Figma replaced solo design files

- GitHub replaced individual code repositories

So why are AI companies ignoring this lesson?”

Pocus doesn’t want to make a champion out of one rep, they want to deliver success at scale. Alexa believes this org-wide consistency is a true unlock for any enterprise adoption of an AI tool.

When all reps are able to learn from all sources and all teammates, you get a true picture of what works, what doesn’t work, and iterate to collective success over time instead of getting siloed successes with outliers.

Another important reason behind this org-wide emphasis is Alexa’s understanding of AI’s fundamental tenet: it learns and gets smarter over time as you use it more.

“If you have just one rep using an AI point solution, it’s getting smarter only for that one person and their context. When the entire organization is able to use it and operate off of the same platform, it learns from all the reps and their context, and you get to realise compounded benefits.”

Building for enterprises is drastically different from building for individuals, though. Alexa shares two unique characteristics of selling AI to enterprises:

» First is the nature of evaluation. Enterprises want to know how they can use AI to roll out capabilities to their entire user base and hit revenue goals. You need to have a clear answer to how you enable this.

» Second is the role of the “unsexy features” like compliance — echoing RunLLM’s Vikram Sreekanti. To this effect, Pocus had to add a few tactical features such as:

• Retrieval intelligence on top of data (think CRM notes, call recordings, product data, and enablement materials),

• tools for admins to roll out features,

• reporting to help them monitor results, and

• dashboards for frontline managers.

🏰 How Pocus is confidently differentiating from the incumbents

Sales intelligence isn’t a space devoid of incumbents.

Alexa thinks they have 3.5 main differentiators against them.

The 0.5 first. Pocus is more intuitive and user-friendly. It isn’t enough to win, but it does make your case stronger when reps love you.

The other differentiators are more technical.

» First, Pocus has built a trademarked “Relevance Agent” that constantly learns from internal and external data to help reps understand their accounts deeply and suggest hyper-relevant next steps.

It sits on a sophisticated AI-oriented backend which if an incumbent tried to replicate would cost them a rebuild of their backend, Alexa shares.

» The second is the prescriptive nature of their approach.

Alexa, during her time in sales at a startup, saw how traditional sales leaders work. They would give junior sales reps a “phone book” and expect them to know what to do. It never worked.

Pocus is determined on being way more prescriptive, to the extent of surfacing signals like “here’s the account you need to focus on”, “these are the people you need to target within the account”, “here’s the messaging”, and so on. Incumbents don’t do this as well.

» The third is the way they tackle data overload.

On any day, reps using traditional systems have to go to Salesforce, Google Sheets, internal admin, Tableau, ZoomInfo, sales nav, company website, and a million other tabs to figure out who to go after and how.

Pocus consolidates this information and gives the “so what” of this data, including why it matters, and how someone can use it well.

“All of these become our right to differentiate and win.” says Alexa.

👀 Why jumping to “using AI” is counterproductive and how Pocus’ product development has embraced AI

“I think a lot of people today start building with a sentiment that they want to “use AI”. They start with the solution and then try to find a problem. We never try to do that.”

Alexa and team have not thrown their hat into the frenzy that is bolting-on AI for the sake of it. Instead AI has found its place as a potent, relevant consideration in their solution set.

They still start with the problem, and work towards finding the solution, the only difference is that now it’s second nature to see if the solution can be built better, faster, or cheaper with AI.

Alexa exemplified a recent build of their AI strategy feature to elaborate on this. This feature, a AI-heavy capability, does the job of creating an account plan for reps by digesting internal and external data.

Reflecting on how it came about, Alexa shared that it didn’t come from a place of “we need AI account plans”, instead it started with the problem like always.

Alexa and the team had seen reps spending too much time in bringing data together and doing the first level of research. Leaders, much alike, were also spending more time than necessary in trying to enable their reps with this step.

Their product development began with the problem, then mockups, followed by customer conversations, iteration, and then seeing where AI could make the solution faster, better, cheaper.

That said, there are some parts of their product development process where AI has changed things significantly and for the better.

Alexa shares that it has become much faster to prototype features and get to a working alpha or beta with AI. The wow moments, when customers see an AI feature, are also bigger. “When you take an otherwise manual workflow and automate it in seconds with AI, users find it fascinating.”

Alexa pauses to add one important caveat, as advice to other founders: AI has an effect.

“It is almost a given that you’ll have a wow moment the first time you demo an AI capability, but you also have to make sure that the wow moments happen on the third, fourth, fifth, sixth time as well, not just the first time.”

🏄‍♂️ Navigating AI pricing for a buyer persona that is too used to the predictability of the seat-based model

Historically, Pocus has been primarily seat-based. They would sell a plan based on the number of reps with a limit on first-party data access.

They are still using the same approach, with one change: usage credits. Instead of applying a limit on the CRM or warehouse data ingested, they have introduced credits for the additional data they are providing now akin to a hybrid model.

We asked Alexa if they considered any other pricing model such as usage-based or outcome-based. She brought in a personal anecdote about CROs:

“We sell to CROs and a lot of them only think in seat-based models. So if you’re trying to sell usage-based or AI-based or work-based to a CRO, they’re not going to like it because it’s hard for them to predict.”

What about the compute costs, one would wonder.

Alexa is aware that they are “threading a precarious needle” and that their pricing model will have to become usage-based in order to drive meaningful business while also covering costs.

For now, the way they are managing this is by selling credits in bundles. Bundles allow companies to buy credits in bulk instead of a pay-as-you-go model. This helps teams feel that costs are less uncertain and unpredictable.

Alexa can see a future moving away from seat-based but doesn’t think it will happen this year. CROs aren’t ready to buy in a new way and as a startup, Pocus doesn’t want friction. It would be interesting to see how this evolves.

In a different conversation, Sudowrite’s Amit Gupta had shared how difficult yet critical it is to find levers that correctly factor in your compute costs. They had a thorough evolution of their own, going from flat-fee to charging for outputs to model-specific usage-based credits in an effort to strike a balance between creating value for customers and covering their costs.

Pocus, with their usage-limited pricing have zeroed-in on four main levers here.

» First: first party data — how much data you are ingesting from your CRM/warehouse.

» Second: website visitors — the number of people who land on your website/month.

» Third: email lookup — used when you want to see contact information for a person.

» Fourth: Watched Accounts — number of prospects being scraped on using AI.

The rationale behind these was simple: it costs them money to perform these actions, it costs them more money to perform more of these actions, and it gives the end users more value the more they use them. It’s a win-win.

Currently, they offer a set limit for all four with the base plan and users have the option to buy more credits if they need more.

P.S. This interview was originally published on Relay here. It has been editorialized for clarity and flow. Thanks to Alexa and Sandy from Pocus for taking the time.

PMF /evals ◎ is just getting started. Tell us about how you’re approaching AI-native building in the wild! Or what you’d like us to cover. Hit reply.

Brought to you by Chargebee. Chargebee helps AI-native, recurring revenue businesses scale with billing and monetization infrastructure built for speed, flexibility, and rapid iteration.

Aaron Levie wasn’t supposed to end up in enterprise software. But in 2004, a research assignment, and a peculiar interest in storage, set him on a path few 22-year-olds dared to pitch.

While his classmates dissected footwear and apparel, Levie zeroed in on software, and then on 3 impossible-to-ignore shifts: mobile was breaking out (BlackBerry had arrived), the internet was going mainstream, and computing costs were crashing.

Aaron knew file-sharing was ready to be disrupted. “Everyone was charging too much money. Not enough storage, poor interfaces, no support for mobile.”

Atop these tailwinds, he set out to build Box with 3 friends, no business experience, and an investment of 300K USD from Marc Cuban.

Since then, Box has seen multiple strategic shifts under Aaron’s leadership, beginning with a pivot from consumer to enterprise prompted by monetization alarm bells.

Aaron has successfully navigated shifts before and he might do it again. In this edition of PMF /evals, we go into his thoughts on how the AI shift differs from the cloud, how they are fending off business model threats, what AI-native startups should build, how Box is deploying AI, and the advice he is giving founders today.

Quick note: If you enjoy reading this, make sure you share it with friends/colleagues. And if someone forwarded it to you, subscribe here to get the future editions.

Subscribe now

That’s from Aaron’s talk at Stanford in 2016, but it applies eerily well to the present moment. Including what he adds:

“If you are a media company, you are freaking out because of Netflix. If you are a life sciences company, you are freaking out because of 23andMe. If you are a financial services company, you are freaking out because of Stripe. If you are a hospitality company, you are freaking out because of Airbnb.”

🔍 How the AI shift is different from the cloud shift

The biggest difference, according to Aaron, comes down to one word: resistance. Or rather, the complete absence of it.

Cloud adoption was a decade-long slog of convincing skeptical enterprises. "You would go to banks and it was basically a non-starter. It was like, 'Oh this is nifty, when are you going to have the on-prem version?'"

Aaron's recent meetings, including with the same types of banks that once showed cloud vendors the door, have revealed that AI is the exact opposite.

The conversation isn't about whether AI matters, it's: "How many use cases can I apply this to? How do I lean in more? We've got 20 experiments running — we'd love to make some of them production-ready." Pushback, if ever, is purely technical.

Unlike cloud, which was an abstract concept for most business leaders, AI's capabilities are viscerally legible. They have seen what ChatGPT can do.

Aaron does add a caveat. He says that enterprise adoption won’t be a dreadful slog this time, but it will still take a while.

“It's worth noting that for enterprise adoption of AI, it is still going to be a decade-long journey. We're 20 years into cloud computing and the quarterly rate of cloud computing adoption is still growing at the same rates as ever before. It will be much faster with AI, but it will take time.”

More than one interviewer asked Aaron if AI agents are just hype, Aaron thinks the answer is vehemently present here, we just need to look at devs. "Developers had their ChatGPT moment with GitHub copilot at least 12 to 18 months before everybody had their ChatGPT moment on the consumer side."

The same thing is happening now with AI agents. "Cursor, Windsurf, etc are having their AI agent moments, and in a year and a half or so, the general population will have their AI agent moment… Agents are moving AI from a read experience to a read-write operation. It’s a matter of time.”

👀 Categories, verbs, and nouns: the breakfast menu for AI startups

AI has enabled an explosive expansion of categories. There are spaces that have no natural or meaningful tech incumbents like legal, healthcare, and consulting services. These are going to be generational opportunities.

For example, Aaron shares that for the first time ever, software creation has become a category.

He adds a crucial insight about AI economics: "AI expands the TAM of software from only being able to sell software to people. You were capped by the number of seats, now you have categories where the AI agent spend will be an order of magnitude more than the spend on software in a non-AI world."

If you need help in understanding where to look for ideas, Aaron shares an interesting framework in his YC interview: verbs and nouns.

Around 2022, he made a list of all the basic nouns and verbs of human life. We eat, sleep, travel, watch things, and get entertained.

“In 2008, everything was subpar: getting food, listening to music, travel, entertainment, hospitality. By 2015, most of them had been solved. This created a tough environment for startups because you could only build derivative things.”

The same thing happened in enterprise. Payroll, CRM, email, calendar — every problem had some incumbent or at-scale startup. “Today is the first time in a decade where there's a new set of nouns and verbs that startups are positioned to solve — because AI has changed the landscape enough to create opportunities.”

He believes these opportunities won't be obvious. It won't just be "CRM but with AI" and it won’t be a GPT wrapper because, “You don't want to do things that could be subsumed by a horizontal chat interface.”

The opportunity is: “What categories of professional services or work have no incumbent technology that AI agents can finally solve?”

His math is striking:

"Contract management software as a category is a couple billion dollars globally. Legal services is a couple hundred billion dollar category. Can software now take 10% of that spent through AI? Could that category become a $20 billion category instead of $2 billion?"

Aaron is also a big believer in verticalisation of AI: “We're in a five-year period right now where you're going to have to build agents for every vertical, every domain."

🌪️ “Wild times ahead:” Navigating business model threats from all angles

Aaron has been a student of software business models for decades. He accurately spotted the threat to their business before pivoting from consumer to enterprise:

"Companies like Google, Facebook, Apple, and Microsoft were eventually going to give consumers free storage… and it was going to be really hard to upsell people… on the other hand, large enterprises were spending millions of dollars, and we knew we could deliver it to them for maybe hundreds of thousands of dollars.”

Now he is at it again.

Aaron predicts that as long as there is a literal human element to the consumption of software, seats will stay. Instead pricing will likely evolve to a hybrid model similar to Windsurf and Cursor. In the case that seats do go away entirely, a more dangerous threat can loom.

They are already betting on hybrid. In addition to a seat-based model, Box is also using a credit systems. “So regardless of whether you ask a document question, you generate content in Box Notes or you go to a hub and ask a question, you get a fixed number of queries per month”, Aaron says in a TechCrunch article.

If they go over that, they can use from a shared pool of added credits that belong to the entire company. This allows power users to continue using their AI features as much as they’d like while ensuring the company’s costs are taken care of.

There is a different fear that Aaron has, though. "There is no AI budget right now. Everybody has some discretionary dollars for R&D/innovation, and that's probably the totality of the AI budget. We should, as a software industry, not expect that to be more than like 1 to 10% of an IT person's budget."

The customer fatigue problem compounds this: "We're overloading the customer with a thousand pitches that all are 'the AI of X' and the customer is exhausted." This is also why they decided to embed AI as core functionality rather than an expensive add-on.

📦 How Box does AI internally

Internally, Aaron wants ROI from AI adoption. Either in the form of productivity, output, efficiency gains, or better experience for users.

He thinks the most complicated task here is understanding “work” in a way that makes sense for your org. We have all seen the memos from orgs like Duolingo and Shopify, here is the mental model he has constructed to understand it:

"You could think about AI as going after a mass of work and use it to either reduce it/do it instantly. Or you can think about work as a timeline and try to get through each step and on to the next step faster.”

Box does the latter.

To figure out what to automate first, they created a framework:

X-axis: Frequency/Repeatability

Y-axis: Critical Thinking Required

In the business context, they have seen that the ROI sweet spot is: High critical thinking + reasonable frequency.

Of course, it wasn’t frictionless. Aaron's team has hit every wall:

• Search problems

• Data governance issues

• Agent chaining dead ends

Deploying an amazing agent to a 5,000-person company thinking all is good is naive. There is a possibility that employees could find corporate secrets they shouldn't access. So they had to be mindful.

To tackle the chaining issue, they stuck to using AI for tasks that are still "very task-centric" — doing a couple of steps, not 50. More autonomy often required more human oversight and became counterproductive.

🧱 An “effing no” to building their own models

On the other side of using AI, Box is also deploying AI capabilities to their customers. When asked if they considered building their own models, Aaron shared he spent exactly 10 mins before deciding, "No effing way are we going to get into this war."

“Why give yourself the brain aneurysm of competing for top researchers and spending $2 billion on compute for training runs?”

Instead, they make custom optimizations around current model limitations. In one example, they created "e-RAG" for better entity extraction, but they are not trying to build their own GPT.

🏰 Competitive advantages and possible creeps

On their ability to win the AI agent race for content management, he adds that their API-first approach will pay dividends in the agentic world where more agents interact with your systems than humans.

“There's a little bit less stress on the core innovator's dilemma piece because AI thrives on more data, more business workflows, more user interaction… SaaS platforms are in the most natural prime position to be deploying AI agents within their particular domain."

Aaron is also confident in their ability to evolve as an org. The pre-cloud to post-cloud change was a more rigorous technical change because orgs needed to rejig everything all the way to the infra layer. This time it is about adapting your consumption layer and how they operate. They have already started making these changes.

There are no obvious moats in AI today. Aaron has a heuristic though, one they have used to make Box sticky:

“Get data, do workflows, make sure you've got enough happening within your ecosystem that keeps customers to have some kind of compounding value."

More subtly: "As a single programmer, you could switch between Cursor, Windsurf, and Replit. But as a network of 500 engineers with embeddings on the whole codebase, change management becomes a lot harder."

There is another, scarier threat they face: their pie getting eaten into by foundational models. Anthropic recently released a new feature for directly uploading files to Claude, nothing as robust as Box, but a potential threat. Aaron isn’t too worried:

“When enterprises deploy AI at scale, they need security, permissions, and control. They need the user interface, they need powerful APIs, they want their choice of AI models, and they don’t want to be locked into one particular platform. What we’ve built is a system that lets you have all of those capabilities.”

✌️ Aaron’s advice to young founders:

Aaron’s Stanford talk had a few pointers. Here are three fresh ones with AI in mind:

First: Read The Innovator's Dilemma, Crossing the Chasm, and Blue Ocean Strategy. “I guarantee you'll be 10 times better off.”

Second: Have an incredible founding team. “Be in the grind with people because you're going to have more fun and you’ll be able to get through the difficult times.”

Third: Go after markets where AI fundamentally changes the very economics or actual process of that thing. “Always ride a tailwind.”

PMF /evals ◎ is just getting started. Tell us about how you're approaching AI-native building in the wild! Or what you’d like us to cover. Hit reply.

Brought to you by Chargebee. Chargebee helps AI-native, recurring revenue businesses scale with billing and monetization infrastructure built for speed, flexibility, and rapid iteration.

🔗 Links referenced:

a16z: Aaron Levie on AI's Enterprise Adoption

YC: Aaron Levie: Why Startups Win In The AI Era

The Logan Bartlett Show: Aaron Levie's AI Takes and Advice for First Time Founders

Box chat: AI-First Playbook: Inside Box CEO Aaron Levie’s strategy for turning AI hype into real ROI | Box

The MAD podcast: Box’s Big AI Leap: Aaron Levie on Agents & the Future of Work

Stanford talk: How to Build a Product II - Aaron Levie, co-founder of Box - Stanford CS183F: Startup School

SaaStr: Why Enterprise AI Adoption Is Moving 5-10X Faster Than Cloud with Box 's Aaron Levie and IBM's VP AI

SaaStr: What's New at Box with Aaron Levie, CEO and Co-Founder of Box and Jason Lemkin, Founder at SaaStr

The Logan Bartlett Show: Box CEO on Enterprise AI Trends No One is Talking About Yet

TechCrunch Article 1 and TechCrunch Article 2

Aaron’s LinkedIn and Aaron’s X/Twitter

Their Series A funding announcement noted that every single tier one venture fund in the valley uses Wispr Flow. Every so often there is an unprompted X post spelling out how much people love it.

They are also a part of ’s curated subscription bundle of some of the most innovative AI tools — right alongside Superhuman and Replit.

That’s hard-won traction — especially in a space where Apple, Google, and frontier model makers are all making bets to no real avail.

Curious how Wispr got here, we sat down with co-founder and CEO Tanay Kothari. 

What surfaced was the lesser-known backstory of their hardware beginnings, a view into their philosophy around pricing and retention, and Tanay’s obsession with building for everyone, not just the valley.

#1. How a contained pivot from hardware became Wispr Flow

When Wispr Flow started in 2021, it was with a bold bet to build a neural interface device that could translate speech into text even when you were surrounded by a crowd. Tanay got together a solid team with deep expertise in neuroscience, signal processing, and hardware.

Nothing short of an engineering feat, the device worked. But it didn’t land. 

"We were solving a problem that hadn’t formed yet" Tanay shared. People weren’t adopting voice as a medium the way they had anticipated.

It was becoming clear that getting people to move to voice AND to a hardware device was too tall an order to take up, there was too much friction, partially because other voice interfaces like Siri and Alexa had under-delivered.

“The key thing I learned about products and hitting PMF over the years was you can change habits in people one habit at a time, but you can't get them to do a number of different things all at once.”

The team stayed at it, tinkering, stumbling upon the idea to innovate on the software layer. 

It started as an internal project called “Project Flow” that could take your rambly thoughts and turn them into structured text. They opened up access for ten people and from there, it snowballed quickly.

“We initially had a group of 10 people using it, soon it was a group of 50 because they sent it to all of their friends and that just kept growing month over month.”

After 3 years of building on the problem, they started seeing market pull like never before, albeit for something they hadn’t planned on building. 

It was clear, people wanted Flow. Soon enough, Tanay made the tough call to pause the hardware business and redirected all focus to software.

#2. Going in-house with models to control accuracy, latency, and other core experiences

The core dictation models are all built in-house at Wispr. 

“GPT takes about three seconds to produce the first word, often referred to as time to first token, and then it takes a solid 20, 30 seconds to write the whole thing out.”

Tanay and the team had non-negotiables when it came to the quality of their responses. They knew quality was core to their user experience.

These non-negotiables were:

• Accuracy: As a keyboard, Flow couldn’t hallucinate. 

• Latency: Dictation needed to feel instant, not lagged.

• AI-speak: The output had to sound personal, not AI-churned. 

They assessed the foundational models, but all of them failed to deliver on the bar Tanay wanted the output to be delivered at.

That kind of latency, Tanay knew, could become the achilles heel for Flow. The three years of their hardware past had also given them a clear sense of the kind of rigor they would need to crack adoption.

“We realized that people want to see the entire output in under a second. There is no API to a foundation model in the world that guarantees that. And the only way we make that promise to our users is by having the inference, the model, and everything else hosted on our side.”

#3. Mindshare is the only moat, not tech, not data

“I don't believe technology is a moat anymore. I think even data is not as strong of a moat as it used to be. The main moat that exists today in my opinion is mindshare and user loyalty.”

Wispr’s team works with one mandate: build an incredible amount of user love. This dictates their product decisions, messaging, brand, everything. 

If you visit Wispr’s website, you’ll have a sense of what this means in action. The brand looks unique and recognizable, intentionally. Tanay believes in this so much that there is a dedicated awareness budget.

“We do performance marketing and we track the cost of acquisition and so on, but there's a budget we have allocated purely for awareness building and putting Flow in places where it is in front of millions of people.”

Another thing that Tanay is big on is “hitting the spot” with their marketing. Wispr’s marketing never uses jargon, talks like a friend, and focuses on driving an emotion home. 

Interestingly enough, Tanay derives a lot of this philosophy from commodity consumer brands like alcohol, food, and apparel. A glimpse below:

Sensing an urgency to crack messaging, brand, and distribution, we asked Tanay what makes him so marketing-first as a technical founder, what he shared is a worthy reminder for every founder especially when the world is in a dizzying flux:

“The one thing that can beat a startup with an excellent team, working at an insanely fast pace is someone with a distribution advantage.”

This also underscores Tanay’s take on the controversial AI debate of margins, where he defers focus to mindshare again. While there is a ton of chatter about AI margins behaving poorly in comparison to the SaaS-era margins of 80%, Tanay pays just enough heed to it.

“We’ve seen moments where teams spend a quarter optimizing infra, only to see model pricing drop right after. We’d rather keep moving.”

Because Wispr trains and serves their own models, their gross margins are close to 90%. Some other AI startups are operating with a margin of sub-20%. His mindset is the same in both cases: to not over-focus on margins.

Pricing changes often. Models get cheaper. What feels expensive today might stop being so expensive three months from now, or might not even be needed. 

#4. Determining staying power through a maniacal focus on retention

Early growth was never a reliable indicator for PMF, and with AI the adoption boom and bust has become even more common. Tanay was keenly aware of this from the beginning.

This is why Wispr’s team instead prioritised retention to determine their real staying power. “Were people coming back after the initial novelty wore off? Were they integrating the product into their routines?”

The team pulled in all retention tactics and notes, tracking their "smiley curve" for lapsed users, tracking their retention curve all the way up to week 26, and looking out for the health of their full funnel.  

“I care about weekly active users because if you're getting a hundred people to try your product and only five are using it a week after, that to me is a failure of either the product, the marketing, the audience that it was targeted to and beyond.”

As different schools of thought, with different products and categories, some teams accept low retention, and just keep filling the top of the funnel. Others go deep on habits and optimize for long-term usage. Tanay is in the second camp:

"As a prosumer product, we lean toward retention. That’s what we monetize. If you go the other route, that’s fine — but you need to know how much you’re spending to acquire users and how many of them are actually staying."

#5. Why the point of your pricing model is to help users extract the most value from your product 

“From a company perspective, I can see why usage-based makes sense. You’d think that your costs are usage based, so you should price it usage based. But if you make that decision without thinking deeply about it, it is lazy because you are putting the cognitive load on your user to figure out how to think about value from your product.”

Usage-based pricing is all the rage right now, but it wasn’t going to work for them for multiple reasons. The most obvious was general trust and familiarity. Tanay wanted to sell to “the average person”, not just the silicon valley.

"My dad is never going to understand token pricing. He shouldn’t have to."

Further he knew that paying monthly for a product is already a step away from the comfort zone for most people, and they didn't want to make it any more complicated than that. “I wanted them to think of this purchase the way they think of Netflix”.

Another challenge that Tanay discovered is the exploration barrier created by usage-based or consumption-based models, heavily underdiscussed right now.

He has a hunch, one shared by our co-founder Krish Subramanian and hinted at by Sam Altman, that usage-based pricing can make users (especially with consumer focused products) too conscious of every action and penalise exploration of the product — incredibly counterproductive when you are trying to build a new habit. This is what their pricing page looks like as of September ’25:

“You get hooked to Netflix when you watch a bunch of things, you get hooked to Spotify when you listen to a bunch of songs, you get hooked to ChatGPT when you use it a bunch. By charging users for these actions, you are actually disincentivizing users from engaging with your product”

The challenge of inference costs is real and Tanay gets it. The solution, he proposes, is to figure it out internally on the level of your models and other technical efforts instead of pushing it on to the user.

Tanay closed his thoughts on the pricing model by reflecting on the recent changes that Cursor and Windsurf made, switching to usage-based: 

“Cursor and Windsurf recently moved to usage-based and it makes sense because they tested flat and realised the changes that were needed. Plus they sell to developers. Developers get usage-based. Canva wouldn’t do that.”

Interestingly enough, even Cursor and Windsurf have had to face pushback from their customer base on this pricing change. Tanay’s point remains, pricing model decisions need to be made thoughtfully in tandem with the second-order effects he shared. 

#6. Monetization as the realest indicator of PMF

“Unless you have a very strong exception to make, monetization is and always will be the best way to measure product-market fit.”

Monetization has always been a behemoth to tackle, now with AI it is all the more hairy. When to switch it on? What to make of it? What does it say about PMF?

Tanay offers a useful frame to tie PMF and monetization:

“PMF at the end of the day is the simple question of: did you solve a problem that somebody cares about and you can repeatedly get to other people like them?...If that’s a yes, then it is rational that people should want to give you a share of that value by paying.”

Monetization potential is so crucial at Wispr that whenever they are thinking about a new feature or a new product offering, monetization is one of the main tests to pass. There are a few ways the team goes about this:

1. Checking current behavior: Are they already paying for a tool that does what your new feature would do? If someone's spending $20/month on an existing solution, that's stronger validation than any survey response.
   

2. Direct ask: Simply asking: "Would you pay $5/month for this feature?" Tanay shares that they use this simple test to “assess if users are at all willing to pay for something.”

Beyond serving as a cold test of your product’s real value, at times, it can also show you how much people love you. Tanay recollects when they had first turned on monetization with a big launch, about half of their existing user base started paying right there.

Moreover, their emails were flooded with messages saying it was the best purchase someone had made. That level of reaction was a solid affirmation of the PMF Wispr had attained.

H/t to Tanay for taking the time to chat with us and share his lessons so candidly. 🙌

PMF /evals ◎ is just getting started. Tell us about how you're approaching AI-native building in the wild! Or what you’d like us to cover. Hit reply.

Brought to you by Chargebee. Chargebee helps AI-native, recurring revenue businesses scale with billing and monetization infrastructure built for speed, flexibility, and rapid iteration.

Experimental budget or not, should you even care? 🤔

That’s a question Arjun Pillai found himself pondering over as the co-founder of Docket, an AI-native platform for revenue teams.

After all, conventional wisdom has directed founders to go seek “committed budget”. That is the budget worth fighting for, the sign that a customer is serious. Right?

Arjun recalls another AI-native challenge: when Docket kept surfacing the same answer over and over, not because the model was stuck, but because it was correct (it was pulling from the one source it was supposed to). The customer, expecting variation, felt it was broken. "Felt" because they had come to expect something new with every prompt.

In this edition, Arjun unpacks:

• ⚖️ Three layers of evals that Docket uses plus how they are tending to the gap between correct performance and an experience that “feels right”,

• 🐎 How you can build your data infrastructure as a Trojan horse with two strong reasons to help you justify the investment to yourself,

• 💰 Why a consumption-based model works for Docket (for now) and observations on the state of different pricing models — perception and challenges both,

• 🧪 Why experimental budgets are not bad at all along with some notes on the nuances of commitment hesitations from customers, and more.

In past lives, Arjun exited Profoundis to FullContact and Insent.ai to ZoomInfo.

Like journeys go, this experience laid a serendipitous foundation for founding Docket. For the stats-curious, Docket already has some remarkable customers like Pathify, Demandbase, and Whatfix, and they have raised 20.3M USD so far.

Arjun’s words, lightly edited for clarity and flow, appear below. ⤵

#1: Data infrastructure as a Trojan horse

I believe that data (and your data infra) is a Trojan horse.

My personal philosophy has been to build forward-looking data platforms. Systems that are capable of serving way more use cases than what customers see right now.

At Insent (acquired by ZoomInfo in ‘21), a startup I previously co-founded, we had a CDP underneath the core chat platform. It could identify people at the cookie level and connect that to accounts. 

And everything felt seamless because of that underlying CDP.

At Docket, customers see our AI agents, but the real magic is the Sales Knowledge Lake^TM underneath them. The Sales Knowledge Lake unifies data from various sources, cleanses it, learns from top sellers, and then applies reasoning models to generate pre-processed knowledge graphs.

Investing in this kind of depth is important for a couple of reasons. 

1. It addresses the fundamental problem that everybody's data is always broken. ZoomInfo had 35,000 customers with Salesforce. I've done integrations with hundreds of them and I never once saw clean data. And this is not changing anytime soon.

2. When you build deep data infrastructure, it gives you optionality. Your current product (or packaging) might fail, but if the underlying tech is solid, you'll still be in a good spot, through an exit or a pivot. 

My advice to founders is to put serious thought and effort into this layer.

#2: Three layers of evals

Building AI-native products is fundamentally different from building traditional SaaS. In SaaS, you define the logic: if a user clicks a button, X happens. You can build and test around predictable behavior.

AI doesn’t work that way. 

You can ask the same question ten times and get ten different answers. Even the underlying facts might shift based on specific phrasing, context, or how the model interprets available inputs.

That creates a very different product dynamic. It’s not just about producing a response, you also have to get the customer to agree that the response is good enough. That bar is subjective, and it moves.

So when you build AI products, you need to think in layers of evaluation. 

We use three:

1. Internal evals, using our own data and use cases

2. Public evals, to benchmark against what’s commonly available

3. Customer evals, which ultimately matter most — because if a customer doesn’t trust the output, the rest doesn’t count

When we built Docket, we didn’t say, “Here's the product, go use it.” We said, “Here’s what we think works. Can you tell us if it’s good, and where it breaks?” 

That feedback loop became core to our development.

There’s a good example from just last week. A customer was working on an RFP in Docket. They asked the system to refine its answer several times, but it kept returning the same result. They thought it was broken.

But when we looked closer, it wasn’t. There was only one valid source of truth for that particular question, a single fact that applied across the company. The system kept surfacing it because it was right. 

The consistency wasn’t a bug, it was a feature.

But it didn’t feel that way to the customer. And that gap between model performance and user perception is where a lot of the real work happens in AI. Especially when you're trying to build trust into the product.

#3: Why experimental budgets are not bad at all

I’ve never been concerned about whether a prospect’s budget is labeled “experimental” or “committed.” If a company is putting money toward a pilot with us, it means they believe something might work. 

That’s already a meaningful signal. It’s our job to validate that belief through outcomes.

At the same time, if someone puts in their experimental budget to run a pilot with us, I assume the deal isn’t closed. A prospect’s experimental budget isn’t a free pass, it's a budget you have to keep earning.

From our side, implementation happens during sales. Testing happens during sales. The deal’s considered closed only after all of that, when the buyer has confidence in pursuing a longer-term partnership with us. 

That’s why I don’t think it’s fair to insist on a fully committed budget upfront. 

Buyers are doing their own evaluation. Many are trying to understand what’s real and what’s hype. An experimental budget is a reasonable way for them to explore, and I think it will remain the norm for a while, especially in AI.

When we talk to companies about their GTM strategy for AI, we give two simple pieces of advice.

First: start with a pain point. Don’t go shopping for shiny features. Start with something real—something painful—and then go look for vendors solving that specific problem. Don’t begin with a long list of 150 AI tools.

Second: anchor on a KPI. If your goal is to improve win rates, or reduce the AE-to-quota ratio, use that as your filter. Then find out which vendors are most aligned with that outcome. 

That’s how one public company found us. In their research, Docket was one of the few solutions that made sense against their actual need.

#4: Pricing can be iterative and why a consumption-based model works for Docket (for now) 

Pricing has always been a moving target. And in the AI world, it’s evolving even faster.

We’ve seen software move through several pricing eras. 

Software licensing. SaaS. Then usage-based models with minimum commitments. More recently, outcome-based pricing has emerged, where companies like Intercom or Sierra charge based on ticket resolution: if we solve a problem, we charge a fee. 

Now there’s a new shift underway: agent-based pricing where if a human seller costs $140,000 a year, maybe an AI agent that performs similar work is worth $20,000 or $30,000. 

That framing is gaining traction, but like all pricing models, it depends on the context.

The truth is, there’s no universal “right” model. The kind of product you’re building, and the kind of company you’re selling to, both influence this decision.

Outcome-based pricing sounds ideal in theory, but in practice, it’s hard to model and harder to forecast. CFOs want predictability. 

If we say, “we’ll charge you based on outcomes,” the natural question is, “but how much should I budget for this year?” That tension matters.

With Docket’s newest launch, we’ve chosen a consumption-based pricing model. We see consumption as a proxy for value. The more the customer uses the product, the more value they’re likely getting from it. That allows us to align what we charge with how our customers experience and measure success.

This also means we’re not trying to maximize revenue on day one. I’m okay charging a little less initially if it helps build alignment and trust.

I’ve joked internally that the easiest way to try a new pricing model is to add the word “beta.” If we say, “AI seller — beta pricing,” we’ve created space to change it. 

Take off the beta when the model proves itself. 

Pricing can be iterative, as long as we’re transparent and fair.

We’re not Microsoft, we don’t need to lock in the model for the next decade.

#5: The M word: Margins

Early-stage startups don’t spend much time thinking about margins, and I don’t think they need to. The focus in those early days is on top-line growth, on proving that the market is real and that customers care.

This is especially true in AI. 

The cost of inference is dropping quickly. By some estimates, it’s improving by a factor of ten every 12 to 18 months. If I’m paying a dollar to OpenAI today, there’s a good chance I’ll be paying ten cents for the same capability next year.

Companies are built for the long term, not for the next quarter. If you’re thinking ten years out, short-term infrastructure costs don’t matter as much. There will always be time, and plenty of levers, to optimize when you need to.

That said, we’ve still made early decisions that helped. At Docket, a lot of our processes run through pre-processing and batch jobs instead of real-time inference. These are architectural choices. 

They let us scale usage without scaling cost at the same rate.

As a result, our margins are strong. We’re not paying a high marginal cost to serve each new seat, so we can take a longer view on pricing and expansion. Props to the engineering team for this.

#6: On Docket’s Series A: Raise before the wave hits

The windows of opportunity in AI are wild and unpredictable. 

We knew software programming would be one of the early ones, but did anyone expect Cursor to reach $500m ARR that quickly?

That’s the nature of these inflections. You don’t know which vertical will hit next, or how fast it’ll move. You just know that when it happens, it moves fast.

In that kind of environment, having a war chest matters. Especially if you want to build a significant company, not a lifestyle business. That’s what I’m trying to do with Docket.

You have to be ready to catch the wave when it comes. And that means having the capital, the hiring flexibility, the time advantage. Resources give you room to accelerate when others are still recalibrating.

Our Series A came together organically. Mayfield saw the opportunity and moved fast with us. It felt that was the right thing to do, not just because of where we were, but because of what could happen next.

The question I kept coming back to was: If this space catches fire, are we ready to move at that speed, or faster? Today, the answer is yes. Because we raised early enough to be ready.

PMF /evals ◎ is just getting started. Tell us about how you're approaching AI-native building in the wild! Or what you’d like us to cover. Hit reply.

Brought to you by Chargebee. Chargebee helps AI-native, recurring revenue businesses scale with billing and monetization infrastructure built for speed, flexibility, and rapid iteration.

“For me there was a visceral sense of your thing taking on a life of its own and lurching forward, like getting pressed into the back of your seat by a fast car or a plane taking off.” — Drew Houston, Dropbox

Product-market fit used to feel a certain way.

Forever ambiguous, but it had some familiar edges: usage spikes in unexpected pockets of users, ballooning waitlists, overzealous bug reports sweeping Slack. Some math. Some signals. Something. 

Not anymore, though. AI-native products are hitting record ARR with impossibly lean teams. Entire enterprise workflows have collapsed into simple prompts. 

Pricing models (not just price points) now shift mid-flight. Savvy, not-new-to-disruption incumbents are layering new S-curves on top of old moats. And the LLM makers you build on? They’re launching apps that look a lot like yours. 

It’s a fascinating flux state. ◎

So at Chargebee, we decided to do something about it. Because we’ve always loved telling builder stories, and you’ve been generous in loving them back (Thanks, as ever, for supporting Relay!).

Introducing PMF /evals ◎✨

a bi-weekly newsletter where we /evaluate AI-native PMF through first-hand stories of early-stage founders, operators scaling AI-native second acts, investors seeing the world unfold around this tech, and with some research/curation by us.

Here we will document what PMF really looks like today: the impossibly-hard-to-pin-down market pull, the enthralling new logic of monetization, the strange new shapes this thing is assuming overall. 

The aim? To surface truths that help the ecosystem make better bets; a second brain all over.

In this edition 1️⃣ we feature Vikram Sreekanti, the co-founder of RunLLM, an AI support startup trusted by teams at MotherDuck, Databricks, and LlamaIndex.

A software engineer and researcher with a PhD in Computer Science from UC Berkeley, he brings a builder’s eye — and a realist’s skepticism — to one of the most hyped categories in AI. Vikram also documents his observations in real time on his substack here.

Here is a TL;DR of what surfaced:

• why enterprises won’t put just any chatbot in front of a 7-figure deal,

• why you can’t bypass integrations-building if you are selling to enterprises,

• that time when someone said usage-based pricing has bad incentives,

• how incumbents are tripping over the “least common denominator”, and more.

Vikram’s words, lightly edited for clarity and flow, appear below. ⤵

#1:

AI for support is one of the hottest areas in enterprise. Look at the news, Decagon recently raised at a $1.5 billion valuation.

What people miss, though, is that most focus so far has been on high-volume, low-expertise support that would typically be outsourced to a high turnover and unreliable call center.

We aren’t doing that. We are doing B2B enterprise support. 

Even though the market calls it all “support”, enterprise support is peculiarly distinct in nature.

You can’t do support for a seven-figure customer buying a technical product the way you would do support for an e-commerce product.

If you look at some other, larger support products, I think they pretty explicitly tell customers that they're not oriented at doing the kind of advanced knowledge support that a more technical product is going to require.

That's a thing that we have heard from companies we talk to, all the way from series B to growth stage to public companies.

They have evaluated a bunch of products, but none of them really hit the mark in terms of being able to handle the complexity and the depth of the product that B2B enterprises are selling. This is what drives us.

#2:

All of us are technical people, so building the core product is a fun thing for us to do. But in building for enterprises, we’ve learnt that in addition to the core use case, we also need to build for scale, maturity, and customer workflows.

There are a lot of cool AI things to build like figuring out how to ask thoughtful questions to our end users, reorienting them if needed, helping them get unstuck, do all the things that a good support engineer would do — the magic bits of building with AI. 

But there's also stuff we need to do that's not cool or sexy or fun, like building a deployment for Salesforce, Zendesk, Intercom and the 25 other support tools that people use because this is what it takes to build a good enterprise oriented product.

We do focus a lot on answer quality in general because we know that our customers aren't going to put us out in front of their customers unless they really trust us with the actual answers. 

And then over time we've just been sort of slowly knocking down yet another integration, yet another connector, whatever it is.

#3:

We don't really view the Decagons and the Sierras of the world as competition. We're not coming across them in customer conversations.

Their support workflows are very action-oriented. They're trying to issue an order refund, send a password reset, rebook a flight, those types of actions. Our support is very knowledge-oriented.

In terms of competition, the closest is probably the incumbents like Zendesk AI, Intercom’s fin, Salesforce's Agentforce etc.

They certainly have the data advantage because they have decades of support tickets that they have answered. But from what we've heard and seen from customers, they don't have the expertise and incentives to go after the highly technical B2B support challenges.

I'll give you one funny anecdote. We were on a call with two Directors of Support at a Series C company that sells a database. They're Salesforce customers and they did a demo with the Agentforce team to see how they could use Agentforce for their support.

The Director of Support looks at us and asks, “Hey, by the way, what kind of products do you guys work with?” I thought he was asking about integrations so I started telling him about the different integrations, but he wanted to know the types of products our customers sell.

He then paused for a second and shared the full story. They had done a demo with the Agentforce team. The demo was about how they could do support automatically if, let’s say, a coffee machine stopped working.

What really got me was when this person said that “you could unplug a coffee machine if it doesn’t work, you can’t unplug a globally distributed database.”

The incentive for these larger companies is to go after the least common denominator for their customer base and Salesforce's customer base is everyone in the world, all the way from coffee machines to databases.

Whereas we focus very exclusively on those more advanced problems, which means that our product just feels right to this target audience.

The last part of the competitive landscape is of course startups.

Most of the startups we see are building generic RAG applications, which is a good starting point but largely ineffective.

Generic RAG doesn't understand the complexity of the product, so it's likely to provide vague/incorrect/generic answers to questions. And it doesn't enable you to support more advanced interactions like asking follow-up questions, analyzing logs, etc.

They also miss the bar in terms of integrating with every support tool and every documentation system and every wiki system.

#4:

I think monetization is an experimental process. 

You start with some reasonable assumption and then you do some back-of -the-envelope calculation.

Like “How long would it take a person? Okay,  well if it took a person X amount of time, what does that mean for how long it would take or how much it's going to save?” And then you take some percentage of that and you start there.

That's been a pretty good place for us to start.

I wouldn't say it's been met with universal acclaim, but it has actually been reasonably good in terms of a set of assumptions that we can make that help customers understand the right way to think about us.

To give you a counterpoint, we just closed a reasonably well-established Series D company. They did a fairly extensive set of testing, gave us a bunch of feedback, we iterated with all that feedback, they were pretty happy with it, and decided to commit.

They asked for pricing. We gave them our standard usage-based pricing on a model and their Head of Support reached out with a bunch of objections to it around bad incentives, feeling that we wanted them to use us more, et cetera, et cetera.

And so we got on a call with them. I asked about how much they thought was reasonable, what the expectation was, how they would make the case to the CFO, and so on. And we came to a number that we thought was mutually agreeable.

At the end of the call, I asked them to share a little bit about their expected usage. Then I took those numbers and I plugged them back into our model, and it was exactly the number that they would've paid under usage-based pricing, except we came up with a different model for them in this case. 

I think there's a little bit of a cultural shift that still needs to happen when it comes to usage-based pricing.

Many customers like it. But some have a negative reaction to it. We're still in a stage where getting customers on board is more of a premium than having principles about the right way to do pricing. But I do think that usage-based correlates everyone's incentives in a cleaner way. 

Even ticket resolution is a difficult thing to measure because oftentimes customers don't close tickets and oftentimes customers don't give any feedback, especially to an AI system.

So we’ve started with number of questions answered as our baseline.

Plus some of the other work that we do around updating documentation, finding issues, all of that has been fairly amenable to basically everyone that we work with except for a few specific cases.

#5:

I think with an AI company, there are three kinds of moats:

• data,

• UX, and

• integrations.

We answer, for our largest user bases, probably 3000 questions a week. We're gathering tons and tons of examples of areas where we're able to solve user problems, areas where we miss the mark, areas where we could be better. We're always looking to gather and categorize that data.

I would be lying if I told you I had a master plan for every data set that we're collecting. But we know this data will matter over time, whether it’s refining the service itself or helping customers improve metrics like resolution and deflection.

Enterprises are afraid of odd training on their data, and so there's some restrictions in terms of what we can do with that data, but in the grand scheme of things, there's still a lot of opportunities.

On the UX part, the main thing that we see as an opportunity is to continue to push the balance in terms of what we expose as capabilities to our users.

One criticism we received a lot even up to six months ago was that we present our product as “just another chatbot”.

A chat bot has a transactional nature. A question goes in and an answer comes out and we think there's a lot more that we can do beyond that. So we have deliberately focused on doing more, proactively.

Let’s say you ask a question, oftentimes there isn't just one answer to that question because it's a technical product, there are multiple ways to get to the same result.

So we don't just stop at one answer, we'll follow up and say, “Hey, this was the first answer, what we thought was the most relevant, but in case that doesn't work for you, by the time you come back, we're going to surface another way of doing things for you proactively”. 

And the other part of UX is acknowledging that this is end-user facing — our customers' customers see these things and so it is important to give our customers visibility and control into what's going on with our systems. That’s a part of UX we focus a lot on too.

The integrations bit is straightforward, when you talk to a new customer, being able to say "yes" to every integration question provides you an incredible amount of credibility and maturity.

It's no one's idea of fun to build yet-another Salesforce integration

but it makes the product usable for more and more customers, especially when selling to enterprises.

#6:

The most common objection that we hear is “It's great that you solve company X's problem”, even if company X is 10x as big as the prospect that we're talking to, “but our problems are different.”

What this really boils down to is a lack of trust. We all remember Facebook messenger's weird foray into chatbots directly in Messenger, that left a bad taste in people's mouths. Chatbots of the past were sort of goofy, robotic, and not really trustworthy. That mistrust still lingers.

So when people think, “Would I put a chatbot in front of my seven figure customer?” They're like, “Of course not. That sounds insane. Why would I do that?”

They are skeptical and you have to build trust with them. You have to sit down with them and say, “Look, I understand that this is the way you solve problems. Here's three examples of similar things we've seen with companies in your market or of your size or whatever it is.”

And ultimately what it will boil down to is them being able to see it work for them.

What we do is we show up to every customer call with a product demo pre-built on their documentation. We show up and say, “Hey, why don't you go to your Zendesk, pull in the last three support tickets, put them in the chat and try us out?”

90% of the time they'll say something like, “Oh, that's actually the answer I would've given.” And that's the aha moment that wins people. 

Another aspect of the trust is trust in the team. We know that every time we start talking to a new customer, we will need to build trust in the team because inevitably we're going to make mistakes every now and then.

We don't try to pretend otherwise. We accept when we make a mistake, we don’t set unrealistic expectations, and we learn from those mistakes both systematically or programmatically.

From a customer experience perspective as well, whenever a user gives us feedback, we take it very seriously and we proactively follow up with them. We tell them why something didn’t work, the improvements we can make, or if there are issues with their documentation, we surface them too.

It comes down to trust—in the product, and in the team behind it.

#7:

2025 is when we've really started to see a growth in inbound interest.

A common thing that we hear from customers we talk to is “I have tried a bunch of chat bots and they all suck, and yours was the first one I came across in company X's, product X's support system that I thought actually was giving high quality answers.” 

If you go to the MotherDuck site, there's an Ask AI chat widget there, and in the corner of that it says Powered by RunLLM, which is a great source of marketing for us.

SEO has also turned out to be a really productive area for us; there's very little in the way of people talking about support engineering right now. There's a lot of AI call center automation, very little AI support engineering. 

If you Google AI support engineering right now, you actually find that most of it is job postings. We're the only product that shows up on the first page of Google's results for AI support engineer. 

And then of course, there's always going to be outbound work. 

What is also working well with outbound is that this year people are starting to really look at AI and say, “It's cool that there's science project budget for AI, but this is the year that we actually have to move some KPIs, show some the board is going to kick our butts at the end of the year if we don't feel like we moved this metric with AI.

And support is still an obvious application area.

PMF /evals ◎ is just getting started. Tell us about how you're approaching AI-native building in the wild! Or what you’d like us to cover. Hit reply.

Brought to you by Chargebee. Chargebee helps AI-native, recurring revenue businesses scale with billing and monetization infrastructure built for speed, flexibility, and rapid iteration.

A fortnightly newsletter that brings you hand-curated pieces of advice drawn from the thoughtful founder-to-founder exchanges and interviews taking place on Relay (curated with 💛 at Chargebee for Startups) and the interwebz. So, stay tuned! 

In this edition, you’ll find the following instructive and inspiring pickings:

#1: Pendo’s co-founder and CEO, Todd Olson, on setting up a vision, a business case, and a bottom-up, team-wide resolve to deploy AI for the right (customer-first) reasons.

#2: tl;dv’s co-founder and CEO, Raphael Allstadt, describes the inward probing that led them towards an unlikely B2B growth medium/channel.

#3: Mailchimp’s co-founder, Ben Chestnut, recalls the unique deliberations behind why they never took the venture capital route, why they never hired a CS team, and why “the thing you’re good at, that’s probably going to be the thing that your company sucks at.”

#1: Driving an organic AI bet 🌱

(From: Pendo’s Todd Olson) (Source: The Product Experience Podcast)

Make sure that they tie back what they want to do with AI to real business outcomes, to real customer outcomes….They’re doing something instead of doing something else. It’s a prioritisation exercise. Everything is a prioritization exercise.

‘We thought this last thing was the most important thing and now you’re telling me you want to build this AI stuff. Is it because you think it’s cool? Is it really more important than that last thing? If you make a case that it’s really more important than that other thing, let’s rearrange…but there’s gotta be a strong why behind it.

One of the most interesting whys that even my board challenged me on when they thought we were going a little too slow and this is right around the time we did the hackathon.

Tech, in general, last year (maybe the last year and a half), the demand environment has gotten tighter. With the end of who zero interest rate period. In general, budgets are tightened. CIOs are saying they’re spending less. Growth is a little bit less across the board.

You know what’s not less? AI. AI is more.

One of the business outcomes for really investing in AI is that it’s one of the few areas where demand is at an all-time high.

If you want to meet customers where they are, you want to adjust to macro conditions, one of the things you can do is to see what’s hot and make sure you’re taking an advantage of it and giving it to customers.

If customers want AI? Build AI.

…

This is a creative conversation, customers are still not ready for it. But the thing that’s interesting is they want to be partnering with companies that are going to be the leaders and not the laggards. Last thing you want to do is decide, ‘well, next year is really our year for AI.’

Because guess what? Whether it’s adding a lot of value to customers or not, that’s kind of irrelevant. We’re learning! And I’m not saying that the AI that Pendo is shipping today is AI that’s perfect.

Or that’s ultimately what our customers want. But boy we’re getting smarter and smarter. And we’re getting it into people’s hands. They’re touching it.

That’s how you build great products. You iterate. And if you’re not shipping you’re not iterating. Us getting something out there fast and knowing that we’re constantly iterating, constantly improving as the technology is shifting, we’re just setting ourselves to be a leader and not a laggard in this technology by not waiting.

Not waiting until it’s a mandate or requirement that people have to have it.

…

It’s a journey.

…

One of the things I realized was that if we don’t sort of organise it and create a strategy and a vision around how AI can help improve our product experience, we’re going to get technology for technology’s sake. Just like lots of other trends we’ve seen in the past.

AI is no different from any amazing technology transformation that we had. Remember the early days of mobile? The number of things that did not belong as a mobile app that we had to all download and then a bunch of them went away…

To this day, Pendo doesn’t have a mobile app. To be honest, most people are using us on a computer. We had one for a very short period of time because it was cool, we could build it.

AI is really no different. I think step one was getting our leadership team. We did an offsite for four and half hours, where we were like, ‘what are all the possibilities, and what do our customers really want from us?’ From Pendo. You got to think that other vendors will do other things and we can let them do those things….

We kind of went back to some of the basics.

Focused on things like jobs to be done. What job are we actually automating? If the whole concept of generative AI is automation, what are we trying to automate? What are the really painful things that we think that this technology can automate? How painful are those things? Is it valuable to the user?

That’s kind of where we started. We started with that vision.

The second major thing that we did was…’now that we had a vision, we wanted to unlock creativity for our engineering team.

The reality is if it’s really me, our head of product, and three other executives come up with an AI strategy…it’s just not going to be as good if you don’t enlist all the support of all the really smart people here at Pendo.

So we did a hackathon. Hackathons are primarily product and engineering events at Pendo. But it does often bleed into other organisations. We let anyone who wants to join it, join it….We decided that we didn’t want to force people to do AI projects. That didn’t align with our culture and what the hackathon was about.

However, we do give out cash prizes for the winners. And we said ‘we’re going to give out cash prizes only for AI projects.’ I will say that someone did a non-AI project which was amazingly cool, we still gave out cash prize for that…

After that we had 38 examples of how our products can use AI… Maybe not all aligned with our vision. But a bunch [of them] were. We then started to find the ones that aligned really well and we started productizing them.

Share

#2: B2B TikTok 🎭

(From: tl;dv’s Raphael Allstadt) (Source: Relay)

When we started out, we had this intuition to not just look at the same growth channels that our competitors were focusing on. We wanted to look inward first and understand what we were drawn to, what we were good at, and what we would enjoy.

That early probing made us realize that we were definitely excited about the idea of humor in communication. We also found that this was something clearly missing in the software industry. More, importantly, we felt we were good at it.

Not necessarily being funny, but at least knowing what’s funny.

So we partnered with a cool creator. And gathered great confidence from our Product Hunt launch video, which won the demo of the year award. We saw that that attempt really resonated with people and asked ourselves, ‘what’s the channel for videos?’

Those were two. Instagram and TikTok. So picking them was a sort of natural evolution for us. We hit it off quite fast, especially with Instagram. We were essentially doing work humor that was funny for everyone and no one.

Our videos started amassing a lot of engagement and reach. Ex: we were featured on 9GAG etc. But we also learned that nobody tied those viral videos with what we did. They didn’t connect us with anything beyond being funny.

So another round of introspection followed. Who did we really want to target? The product org. Not just product/engineering but sales and customer success as well, given that they all sit at the core collaboration that informs products.

We started making jokes about these folks.

With that we were able to align our impressions, our views, our follower growth, and our engagement with those who’d actually care about tl;dv. That’s when we started to see impact. Essentially, a correlation between a clip going viral and the product getting attention from a specific persona.

That being said, one important thing to mention for any social channel is that you always begin with top-of-the-funnel metrics. You just need to have the faith that if it’s the right audience, in the end, they’ll have enough brand love to check out the product.

So, it’s definitely a long-term practice, but it works!

How did we act on the learning that we had to be more focused? We started throwing in one or two product jokes in 10 generic ones.

Whenever you’re shifting positioning on a social channel, you must do it slowly and gradually. You need to be patient around such changes because your audience has a certain set of established expectations.

Within four months or so (yes, it took a while) of making this transition, we saw that we were going even more viral because we were niche. A niche, we discovered, is always bigger than one thinks.

We went deep enough that some people in the audience asked us, ‘are you secretly in our meetings?’ or “how did you get so good at understanding my day-to-day struggles?”

Of course, the best and the worst of companies, we all struggle with the same core issues; there’s just a difference of magnitudes. Say: a clash between growth incentives and inner alignment. Or a lack of empathy for a different team.

Plus, the viewers told us things such as, ‘this is happening to me, but with a variation.’ We listened and addressed that with another skit. Getting to a real symbiosis with our followers.

Share

#3: Notes from Mailchimp’s 20-year journey 📖

(From: Mailchimp’s Ben Chestnut) (Source: MicroConf)

When we started, it was during the dot-com burst. Nobody was getting money. All the VCs had gone bust. So many years passed before the economy came back again and VCs started knocking. But, by then, it was too late.

We were already making 10s of millions of dollars. We were on the verge of making 100s of millions. And I could see over the horizon, hitting a billion.

When they [VCs] came knocking, it was nice. It was good for my ego. But I couldn’t figure out what I would spend the money on. We were already making enough of it. That was really the most matter-of-fact reason.

The other one was that what I noticed every time they came knocking, they would really just give me someone else’s playbook. They would say ‘take our money and you can do what Constant Contact did.’

They really wanted me to get the next best IPO. And it’s understandable, that’s their job, to get exits for the LPs….

As entrepreneur, I’m sure, you all can relate I wanted to build my own thing my way. This is my baby, I was going to do it, Bruce Lee style. This is the way. My way. So that was one of the reasons. I couldn’t articulate it back then.

It was sort of the lack of respect that they had for the thing that I was creating. I didn’t want to be a copy cat. And I didn’t want to take orders from some nerdy MBA VC to become a copy cat. That’s the worst.

…

What we learned the hard way was we were serving small businesses…I used to tell all new employees, I used to draw this diagram, it was two mountains. And I was like, ‘in our business, you have [the] small business mountain or [the] enterprise mountain, you pick one of those two.’

And in the middle I picked death valley.

Mid-market is Death Valley.

You have clients who have the ambitions of a large business and the budget of a small business. And a lot of those managers who’re running marketing in those mid-sized companies, they’re not exactly your cream of the crop with a lot of experience in marketing.

They don’t know how important marketing is as an investment. They’re very picky, choosy customers. They’re high maintenance. So we focused really on the bare-bones, just the smallest of the smallest customers.

You just lose your shirt if you have customer success managers trying to help them give them any kind of service. It was all kind of self-serve for us…These are people who would sign up for an account, run it for 10 months, shut their business down, go get a job. Two years later figure out, ‘oh that’s how you run a business.’

Start up a new business. Revitalize that old Mailchimp account and then pay us money. That’s like a three/four year sales pipeline, you don’t make money with customer success from that kind of clientele.

We served enterprise for sure. But we chose to serve them through the API. If you’re an enterprise, if you have an engineer, read our API docs, build it yourself, and we’ll give you volume discount. That was the way we chose to do it.

We experimented with customer success. It’s also a cultural thing. We could not quite get that culture into our business. If you look at HubSpot, they’ve got that nailed, but it’s a very different business vs ours.

…

What I learned later on is that the thing you’re good at, that’s probably going to be the thing that your company sucks at. You can’t scale it because you’ve still got your hands into it.

I was really good at design, creativity, and branding. And that team, I noticed after 10 years, wasn’t scaling like the rest. Like support, customer service, operations, I sucked at them so I always hired the best people I could find for that.

That side of my business scaled and was very, very efficient. I think because I didn’t have experience, it kept my mind open to always hire operational people. I always called them that. ‘Those *operational* people over there.’

My passion was not just creativity but fostering a creative culture. That’s the thing that I focused on the most. So, we got to 12,00 people, In think because I wasn’t the one managing them all. I always had great COOs.

Share

🤝 Founder social:

Thanks for reading! 🌻

Team Relay (Chargebee for Startups)

A fortnightly newsletter that brings you hand-curated pieces of advice drawn from the thoughtful founder-to-founder exchanges and interviews taking place on Relay (curated with 💛 at Chargebee for Startups) and the interwebz. So, stay tuned! 

In this edition, you’ll find the following instructive and inspiring pickings:

#1: Pilot’s co-founder and CEO, Waseem Daher, makes a sharp-edged case against building a particular type of business — “lots of smart people have started…with this exact thesis…Almost all of them have failed” — that they themselves have been crafting successfully for years. 

#2: Panobi’s co-founder (previously, Slack’s Head of Growth), Merci Victoria Grace, lifts the potentially dizzying fog of shutting down her startup with an immediate, exacting critique of their original assumptions and concludes what likely went wrong. 

#3: Canopy’s (previously, Know Your Team) CEO, Claire Lew, writes evocatively about the practical (and the philosophical) realizations she had after accomplishing an all-encompassing, time-boxed rebrand as a 3-person team.  

#1: “Don’t start a tech-enabled service” 👩‍💻

(From: Pilot’s Waseem Daher) (Source: Startup Real Talk)

Here’s what makes it tough:

You’re going to fool yourself with traction

Congratulations: your {law firm, accounting firm, recruiting firm, travel agent, executive assistant, SDR, website design} service just launched. 

You work your network to validate the idea, and you’re able to get your first few customers. You delight your customers by unscalably providing the service yourself, by hand, as you write the first few lines of code.

You continue to do this, it goes well, and you grow the team to keep up with demand—and one day you wake up and you’ve hit the holy grail $1M ARR mark.

You’re crushing it, right?

Well, yes and no. Congratulations: you have built yourself a nice accounting practice. The problem is that you’ve validated that you could start a small accounting firm. Unfortunately, it’s well-known that this is possible: there are literally thousands of small accounting firms in the United States.

The hard part isn’t the first $1m. It’s the first $10m, where you really start to feel the pain of scaling beyond where a normal services firm goes. Your initial signals of success actually lead you astray. You think you’re winning, but your early progress has nothing to do with whether or not you can achieve scale.

And the reason that achieving scale is hard is because…

You’re not going to achieve 100% automation

As you start to tackle the problem, you’ll learn that software can do a lot of the heavy lifting. You’ll also learn that it can’t do 100% of it.

That’s a problem, because your customer will not tolerate the problem being only 95% solved. Their alternative is “work with a legacy accountant,” which can, by definition, provide 100% of the service to them.

(Separately, my view on Pilot is: even if the computer could do 100% of the task, you’d still want a talented accountant in the loop, to provide a compelling customer experience for your client.)

So you are necessarily going to have to staff up a fairly large team of accountants.

Which means…

It’s the worst of both worlds

Imagine all the hard parts of running a B2B SaaS software company. Now add in all of the hard parts of running an accounting firm. Congratulations, this is your new life.

…

You’re going to have bad margins

Your margins want to be bad. By default, you’re a services firm, and services firms have bad margins. (They certainly do not have SaaS margins.) Those margins tend to get worse, not better, with scale—because scale requires overhead.

Unless you’re paying very close attention, you will be unhappy with the margin profile of the business. More importantly, your investors will be unhappy with the margin profile of the business.

The single most common reason that investors passed on the Series A, Series B, and Series C was “I just don’t believe that the margin profile will actually be good.”

They happen to be wrong in the case of Pilot specifically, but this is absolutely one of the main causes of death of this shape of business. You’re unprofitable and your margins are bad, but you continue scaling. One day you wake up without access to further capital and unsustainable burn, and you have to shutter the business.

Good luck

I’m very bullish on this sort of company. (If I weren’t, I wouldn’t be running one.) I think a number of iconic, enduring companies are going to be formed using this model—and it’s going to be amazing for customers.

But it’s definitely going to be a slog.

To the founders of the new crop of “Pilot for X”—best of luck!

Share

Related Relay reading:

Basis’ co-founder, Bebe Kim, on failing to build a tech-enabled law firm and how that influenced her next venture (“The experience of co-founding and shutting down my next business, Atrium, came with a lot of lessons, and reframed a lot of the perceptions that usually dictate Silicon Valley bets.”)

#2: (Controversial) takes from not finding PMF ♨️

(From: Panobi’s Merci Victoria Grace) (Source: LinkedIn)

1️⃣ I’ll never do my own customer research again. People are too nice to your face. In the future I’ll only have proxies do my interviews for me and remain just doing founder sales myself.

2️⃣ I’ll also reach out to folks at companies that are struggling to see what problems are painful for them. I didn’t grok until too late that my challenges running growth at Slack were champagne problems… and so were the problems of the successful companies that I spent most of my customer development time with like Notion, Figma, and Postman.

3️⃣ Most individual contributors like their silos, they feel safe in them. If your product requires individual PMs to opt into a new system – of which the primary benefits accrue to the larger org and not the person themselves – I think it won’t work.

4️⃣ On one hand, a silo is a prison where people feel trapped and frustrated. This is probably a high performer mindset. But most people are – statistically speaking – average performers.

5️⃣ On the other hand, a silo is more like a private office compared to an open floor plan where your monitor is always in view of your colleagues and your boss. People with this mindset who are working outside of their silo feel exposed and unable to control how they and their work is perceived.

6️⃣ Only executives and high performers care about disrupting silos because they either have or feel they have the agency to create meaning out of a more comprehensive data set.

7️⃣ Most people value the small tuning work that they do to present a story around data. They don’t actually want the data to be shared with anyone at the company without their ability to interpret it for them.

8️⃣ Most people don’t want to be more efficient. They feel like the time they spend moving data around, updating docs, and holding update-based meetings is actually what they get paid to do instead of busywork that steals time from driving outcomes.

9️⃣ It’s also possible that people who become PMs simply... enjoy creating and updating docs?

🤔 🔥

Share

Related Relay reading:

Kairn’s co-founder, Patricia Bernasconi, on how building too close to user love became a problem (“What we mistakenly assumed from the problem validation research was that just because teams (yes, not just people) already pay for project management apps, they’d be willing to pay for a new one.”)

#3: Lessons from a 3-week rebrand 🌳

(From: Canopy’s Claire Lew) (Source: Canopy Newsletter)

As CEO, here are my 4 personal biggest takeaways from doing a 3-week rebrand…

(1) Be willing to walk away.

With each passing day that we worked on the rebrand, I would pause, and ask our team:  “Are we happy with where this is at? Does it make sense to continue? We can always turn back. We can always walk away.”

If we didn’t find a new name we liked in 1 week, we wouldn’t continue.

If we didn’t come up with a logo we liked by end of the 2nd week, we’d stop. 

We would do the rebrand if and only if, each step of the way, we felt that it was 100% what was best for the business and for our customers. We always gave ourselves the option to walk away.

Sunk cost fallacy is seductive swamp – and avoiding it means being explicit about the other roads through the woods we could take.

After all, we didn’t have to change our company name. Our previous company name (“Know Your Team”) was just fine. It’d worked well enough for the past 10 years. It probably could work pretty well for the next ten. 

We didn’t have to change anything. Keeping that perspective, and giving ourselves permission to change our mind, helped stay calm throughout the process. 

…

(2) Accept the existential.

…

A rebrand is a much more emotional process than we credit it as.  It’s not like choosing a new marketing channel, or deciding you’re going all-in on a different product strategy.  Rebranding is an inherently personal, subjective decision.  There really isn’t anyone who can tell you what the “right” name is or “right” logo is, no matter how much research you do, how many books you read, or how many podcasts you listen to. 

So accept the subjectivity. Know that for the most part, you’re leaning on personal taste. Of course, do root it in what you can find to be objective, as it makes sense (for example, how does this actually support your vision and business goals). But don’t be surprised by the emotional nature of the process.

Acknowledge and accept the existential identity crisis it will undoubtedly trigger.

(3) Keep the decision committee small.

When you rebrand, it’s tempting to crowdsource opinions. However, often the more voices in the room, the harder it is to hear any voice.

As a result, we kept the decision committee exceedingly small. Just 3 people: Myself (CEO), co-founder and CTO Daniel, and our designer James. Then, I would have final say on the decision.

We’re a small company to begin with, so we don't have precedents for 20-person committees, which is helpful. But this intentional move to make sure only 3 people were the ones contributing names, giving feedback on logo kept us moving fast.

…

(4) Prioritize only what people are going to care about.

When you start creating the list of to-dos for a rebrand, you start to realize very quickly that there’s a million things you’d like to change.

Logo appears in X thousand many places.  Name is fossilized in X hundreds of podcasts, videos, other assets.

The list starts to become overwhelming long.

So we decided to be ruthless about was only changing what was absolutely necessary – and that’s it.

No one’s really going to care if the logo isn’t updated on an obscure marketing PDF right away.

As long as we covered the basics, we were good.

…

Full caveat, it’s still a lot of work.  Whether it’s 3 weeks, 3 months, or a year – whatever time you set aside for your rebrand (or any big project) will feel exhausting. 

I needed real time to rest, after we launched the rebrand. And I think I still need more recovery time, more than I’m willing to admit.

The lessons I shared here are by no means a “shortcut” or hack. No, no, no. All meaningful, high-quality projects require effort and time.  

Rather, I hope they give insight into how a big meaty project can be done on a shorter timescale with a smaller team, than we originally think. 

Share

Related Relay reading:

Claire’s Relay AMA on unconventional pricing models, avoiding “judo solutions,” writing everything herself, and more!

🤝 Founder social:

Thanks for reading! 🌻

Team Relay (Chargebee for Startups)

A fortnightly newsletter that brings you hand-curated pieces of advice drawn from the thoughtful founder-to-founder exchanges and interviews taking place on Relay (curated with 💛 at Chargebee for Startups) and the interwebz. So, stay tuned! 

In this edition, you’ll find the following instructive and inspiring pickings:

#1: HashiCorp’s co-founder, Mitchell Hashimoto, relates how his role evolved from a founder-does-everything start to a spirited reclaiming of what he had always wanted to do. 

#2: Amplitude’s co-founder and CEO, Spenser Skates, explains how new innovations and mature product processes often pull in opposite directions; and how that realization led to their second largest success. 

#3: tl;dv’s co-founder and CEO, Raphael Allstadt, goes beyond the surface sheen of their bold, original foray into comic B2B skits and talks about the thinking behind running with this unconventional channel. 

🗞 Recently on Relay:

➡️ Heuristics and Hunches (April 19th) — “I Don’t Think You Can Have a Great Product Unless It’s Freemium” and Other Notes on Growing in an Established SaaS Category with Loops’ Co-Founder, Chris Frantz

— Not just another email tool
— Selling into an established category
— Moving towards a self-thinking app
— Aiming for an unclaimed vertical
— Introducing freemium confidently

#1: CEO -> CTO -> IC 🪃

(From: HashiCorp’s Mitchell Hashimoto) (Source: The Stack Overflow Podcast)

I was naive, I guess, or so young, mentally. As I had never done a startup before…I didn’t really know what to expect. My original thought was that I was so passionate about building these tools, I would start this company and I would be able to spend all my time building these tools.

That couldn’t have been farther from the truth and reality.

One thing I quickly realized was (and I tell new founders all the time) when you start a company around your passion, it’s not to enable you to work on that passion, it’s enabling you to enable others to work on that passion and just realize that dream or that vision.

So you don’t get to really be hands-on. I would say, I was hands-on. As much as I could…for the first year to three years, depending on how activity ebbed and flowed. But I was splitting my time so much with the managerial duties and executive duties.

As much as you could call someone an executive at that stage, but it was like financial management, fundraising plans, ensuring growth, thinking big picture, how are we going to commercialise, thinking about the next X hires that we have to hire. Things like that.

It settled in really quickly. I think within the first eighteen months I was programming as much as I could but probably working very long hours to do other stuff too.

…

One thing I don’t want people to think is like I hated not programming… I didn’t dislike those [other] jobs. I found being a manager really rewarding in a lot of ways. I loved helping people grow and being a good manager (I tried to be, most times hopefully).

But what I realized was that I drive a lot of my passion really was programming. The thing that drove me was really programming. So while I found the management stuff interesting and fulfilling in certain ways, it was taking away from this thing that gave me the most happiness.

I wasn’t worried about losing skills or anything. Because one thing I always think about is that a good way to identify your passion or something you really care about is that I always made time for programming, even when it probably wasn’t healthiest to do so.

I could have worked a 10-hr day without programming once, then I would program for two hours. I didn’t have to. I probably shouldn’t have. But I wanted to. I think seeing that over a long time, just made me realize, that was the direction I wanted to head back towards, eventually.

….

The CTO thing happened, three-four years into the company…That was really a realization of me actually, actively disliking the CEO role. The executive team building, the financial planning. Fundraising. All that sort of stuff.

…

My point of view is 1) I didn’t like that, 2) I had an investor, someone on our bored who told me something, they never pressured us…But one thing he told us was start-ups have a 90% failure rate and so for an investor to invest in a startup, it’s because they think they’ve identified some sort of trait that can help beat those odds.

For us, that trait was always our engineering vision and our engineering ability. So, me spending time being a CEO was putting us into the odds of the 90%. It wasn’t giving us the advantage we needed to beat that 90%….

He just made me think if this isn’t something that you want to be passionate about, do what makes you happy and do what would give company the best chance for success.. And me and Armon both thought that was being engineering leaders. Not necessarily having a keyboard. But being engineering leaders rather than CEO….

Another way to look at it is that we’re capable of being a CEO but we would make mistakes. We wanted to hire someone who had already made the mistakes and knew what to do. While we wouldn’t make mistakes and knew what to do on the engineering side.

…

As the company kept growing, the job description changes a bit. As we started to hire more VPs, introduced some senior directors, my depth away from actual engineering was growing. It became much more of a headcount planning/financial planning, long-term visioning type of role and I wanted to get closer back to solving a specific problem.

At the sane time, we had grown to so many more people and so many more products. So the stuff I was doing as a CTO had to be done but it would take all my time. I couldn’t program anymore, I had to get rid of that part.

…

So I had this fork on the read: Do I keep doing this thing, which I don’t dislike? Or do I get back to programming, is there a way to get back to programming that I know I absolutely love?

Share

#2: New products as pre-PMF startups 💥

(From: Amplitude’s Spenser Skates) (Source: SaaStr)

It was hard enough to catch lightning in a bottle and find product-market fit for your first product. How the hell are you going to find it for your second product? I want to share a few tips that we learned after a bunch of failed attempts.

A lot of companies will create a new product innovation team, but the problem is they have that innovation team sit right next to the central product team…The problem is the things that work for a mature product at scale are completely the opposite of what you want to do when driving innovation for the first time.

As an example, your product team may set quarterly goals and have quarterly plans. It turns out while you’re trying to innovate, quarterly goals don’t really help you that much.

They may set long-term product roadmaps to be able to communicate to customers. Again, when you are innovating on a new product you don’t have a long term product roadmap because what you may do may change next week or even tomorrow.

Another thing that’s done is coordinating with other teams whether it’s go-to-market or compliance teams like security. Or teams like support teams…makes a ton of sense for a mature product with tons of users and is exactly the opposite for how you want to run a new innovative experiment.

All those things that help a central product team be successful are huge distractions for your early team. So, as a leader, your job is actually to quester them off from the rest of the organization and protect the team from those distractions so they’re able to innovate and drive really, really quickly.

…

The very first product we did successfully was experimentation. Outside of analytics. First time we tried to do It, we had this as part of the quarterly planning process. We ended up failing and having to redo it.

What we did the subsequent year is we sequester them up, we didn’t include them as part of quarterly planning. The only meeting we had was a meeting every two weeks, where I would ask the same question: how many paying customers do you guys have this week?

And get them to maniacally focus on that.

And not worry about everything else.

Don’t worry about how it integrates with our existing product, don’t worry about what the long term roadmap is, don’t worry about what metric what metric we’re trying to optimise yet, don’t worry about compliance…just get something out there that has paying customers. Do that.

We ended up doing this to great success. Experiment is now our second-largest product after the core analytics.

Share

#3: The best marketing is unattributable 👟

(From: tl;dv’s Raphael Allstadt) (Source: Relay)

I wish I knew how best to measure this work.

When you scroll LinkedIn these days, you can easily come across posts that are fixated on either killing/championing marketing channels. “Don’t do SEO,” says one expert, “do it like this,” says another. “Do paid.” “Don’t do paid.”

This particular moment is quite interesting as well, where a lot of go-to-market work stands to get (and is already) getting disrupted because of AI. We’re yet to see how that plays out.

There’s this “hard to swallow pills” meme which essentially says that the best marketing is unattributable. I think that’s completely true. It just makes logical sense.

If you reach people that you know can buy your product and you make them laugh/learn often, if your product is actually valuable, everything comes together.

Because humans aren’t dumb.

When we do customer interviews, almost 50% say they discovered tl;dv through socials.

As we scale, we’d definitely: track/benchmark more, diversify our funnel better, and run a lot more experiments. But, right now (a sort of 0-1 stage), I feel our decisions can be very qualitative and gut-driven.

This certainly varies from channel to channel and how attributable it is. With paid social, you can definitely put a number to every ad unit. In terms of what it takes to acquire a customer and such.

The funny influencer campaigns we’ve been running are much harder to attribute. So our goal is to grow profitably across our marketing mix. With a healthy knowledge of what we can/cannot measure.

Put another way, I don’t lose sleep at night because we don’t have a number attached to all our marketing channels. Primarily because attribution, to me, is in general a false flag. As it’s the final click.

People who discover us on TikTok, google us because that’s more convenient for them. We’d count that as a brand search on Google with no mention of TikTok.

Even self-reported attribution is broken. As everyone forgets things. That’s why we don’t obsess too much over attribution. We constantly review the overall marketing mix and see what makes sense for our goals.

Then, the last thing here is that if it’s obvious, it’s obvious. We ran an influencer spot and there was an immediate, obvious spike right after. So we made sure to optimize our spend there and got the most out of it.

I know this is not super scientific. But that’s it. That’s how we’ve been making it work.

Share

🤝 Founder social:

Thanks for reading! 🌻

Team Relay (Chargebee for Startups)