Conor founded Glofox in Dublin in 2012, we (Notion Capital) invested in 2019 and he went on to build it into a platform serving studios and gyms in sixty countries. He relocated with his family to the US (first LA and then Miami), sold to ABC Fitness in 2021, and then spent the years since as CRO inside the American acquirer.

At the time of the acquisition the US was 40 to 50 per cent of revenue and 60 to 70 per cent of new business. It is around 70 per cent of the company today.

Somewhere in the middle of the conversation he described a decision that made much of that possible, and handed my own words back to me. They had chosen, he said, to be uncomfortably narrow about who they serve.

Narrowness is measured in refusals

Most stated ICPs are descriptive rather than narrow. They name a segment, a size band, a job title and a use case, and they exclude almost nothing the company was ever realistically going to win. A definition of your ideal customer that is not uncomfortable, that costs you no revenue is worthless.

Uncomfortable is the operative word here.

If the definition does not hurt, if it does not produce a steady stream of deals you turn away that you could plausibly have closed, then nothing has changed.

The discomfort is the evidence that a choice has been made. It is felt most acutely by the people carrying the number, which is why narrowness is so hard to maintain.

Glofox narrowed from gyms to studios. A modality, in other words, rather than a size band or a geography.

Once they were precise about who they served (and who they didn’t), two things became obvious almost immediately.

There were far fewer studios in Europe than the ICP implied there should be, and the American market was running two to three years ahead of Europe for exactly the segment they had chosen.

So they put a few feelers out. What came back surprised them. Conor discovered sales cycles in the US were faster than in Europe.

American studio owners were quicker to buy from an unknown Irish brand than Irish studio owners were, because they already knew what problem they were trying to solve and cared far more about that than about who was solving it.

Then came the harder decision. They could have built where Mindbody (a key competitor) was not looking, taking the quieter ground around the edges. Instead they went at the incumbent on home soil, because, as Conor put it, you cannot become a category leader by assembling a patchwork of small European markets with different buying cycles, different compliance regimes and different localisation requirements.

Higher risk for sure, but considerably higher return. The ICP work and the internationalisation strategy turned out to be the same piece of work.

The narrower you are about who you serve, the more dispersed those customers turn out to be. Precision about “who” is also a decision about “where”, whether or not you notice you are making it. And that can cause significant challenges.

Mike Coscetta, who runs the commercial side of Mews, puts the negative case first:

Treat the US as a single market, he says, and you will go bankrupt inside six months simply trying to build awareness across it.

The cost of breadth can kill you, and no amount of conviction makes that cost affordable. His remedy is what he calls hyper-segmentation, by which he means working out the smallest niche you can go after and winning one genuinely referenceable customer inside it.

He does not think founders take referenceability seriously enough; he says he would pay a customer for it if that were what it took. As a demonstration of the method, Mews spent something in the order of eight thousand dollars on expert network calls to understand an association of Asian-American hotel owners that operates very differently from the customers Mews already knew. That research reshaped the pitch and the pricing, and it produced forty wins in thirty days.

Rich Mason, who runs revenue for Paddle, takes the argument down a level. When he turned his attention to Paddle’s US market, one of several problems he inherited was an ICP problem, and his diagnosis is precise.

The customer profile that works in a fragmented Europe does not port to America, because US companies can scale to enormous size while thinking only about their home market, which changes the job to be done and therefore the whole value proposition. The ICP has to be derived again for the market, rather than translated into it.

What Rich did next is important: he clustered people where the concentration of their ICP actually sits.

Conor did the same thing, replacing a first-come-first-served coverage model with genuine territory mapping once he understood how studios clustered by city and by state. The country, the region and the city are all answers to the same question; the question starts with who and where follows.

The cost of being narrow has fallen

There has always been a penalty attached to narrowness.

Narrow used to mean “expensive to reach”, because the customers were scattered and in Glofox’s instance each one was small, and the cost of finding them ate into whatever margin the focus had earned.

That penalty is declining. Inbound is getting more expensive and less reliable as language models chew through search. Outbound, prohibitive for years at an SMB price point, has become cost-comparable.

His teams now run it with Clay and Nooks, having moved most of an inbound SEO function across into outbound prospecting, and the outbound funnel produces a higher average sale price and better unit economics than the inbound one it replaced. That reverses roughly fifteen years of received wisdom about how to sell to small businesses.

The interesting part is not the tooling. It is that geographic dispersion is now far less of a tax on precision. You can now be narrow and everywhere at once, which even five years ago was considerably harder.

Narrow twice

One more thing from that conversation. Inside the studio segment, Glofox did not only hold an ICP. They held what Conor called an aspiring ICP.

Three years ago boxing was the modality with momentum. They formed a view that Reformer Pilates was about to go vertical, and made investment decisions on the strength of that view before they had product-market fit in it. Where the growth would appear by region, which product gaps to close, what it would cost to acquire those customers early enough to grow alongside them.

I have argued here that your ICP is a posterior, not a prior, derived from the customers who actually expanded rather than asserted at the offsite. I stand by that, and the aspiring ICP is unmistakably a prior. The two sit together more comfortably than they look, though, and it is worth saying why.

The posterior tells you who is expanding, and it remains the only honest basis for how you sell today. The aspiring ICP is a bet placed ahead of the evidence, about who will be expanding next.

What makes it a legitimate bet is that it is narrow. A narrow prior is testable, cheap to be wrong about, and produces a clear signal within a couple of quarters. The broad prior that most companies hold does none of those things, and mostly it delays the moment anyone has to be accountable for a view.

The failure mode is real

When I run premortems, two of the answers that come back most often are that the ICP was too broad, and that the company went too narrow too soon. Both happen.

A founder said to me a while ago that you have to go wide to be able to go narrow, and that is true of the discovery period, when the point of the exercise is to find out what you do not know. Jean-Baptiste Poljak at Upciti spent his first six months in New York doing precisely that, meeting hundreds of prospective customers to work out which of three use cases resonated. Parking won, and the shape of the US business followed from that answer, including who he then hired.

But the discovery period must end.

The test I would offer is simple enough. Name the customers you turned away last quarter that you could have won. If the list is empty, you do not have an ICP. And if you cannot say who you are deliberately building for eighteen months from now, what you have is a description of the present rather than a position in the market.

I put that question to Dan Steinman, who spent two decades defining the Customer Success discipline, first at Marketo and then as Chief Customer Officer at Gainsight. Nobody I know has thought harder about what it means to help a customer realise value.

We’ve been working together for nearly ten years. We didn’t always agree on this one.

Customer number thirty-one

I started by asking Dan when founders know it’s time to hand customer relationships to someone else.

His answer had nothing to do with ARR, headcount or the number of customers. Instead, he described two meetings. In the first, an investor asks, “What’s happening with customer number three?” Everyone in the room knows. In the second, the same person asks, “What’s happening with customer number thirty-one?” The room falls silent. “That’s when you need Customer Success,” Dan said.

The point isn’t that founders stop speaking to customers. Quite the opposite. In the early days, founders are the Customer Success Function. Every customer conversation teaches them something about their product, their market and their business. The problem is that this shared understanding gradually disappears as the company grows.

Dan sees Customer Success as the function that preserves it. Not simply by managing customer relationships, but by ensuring the organisation continues to understand where customers are succeeding, where they’re struggling, where value is being created and how the product must improve.

That, he argued, is why Customer Success belongs at the executive table.

If leadership teams review sales every week but rarely discuss customers, they’re ignoring one of the richest sources of strategic learning available to them.

The doctor and the diagnosis

Once we’d established that Customer Success exists to improve the product and organisational learning, the obvious next question was whether AI changes that role.

My instinct is that it must. As software becomes increasingly autonomous, products are taking on work that was once performed by people. Surely that changes what Customer Success looks like?

Dan wasn’t convinced. He suggested asking the same question about sales: “Does the need to sell an AI-native product fundamentally change what makes someone a great salesperson?”

Probably not. While the tools evolve, the work doesn’t. The technology has changed every decade. The objective hasn’t: build stuff, sell stuff, make sure the customers receive the value they expect.

Customer Success, he argued, is no different. Its purpose remains understanding customers, helping them achieve the outcomes they bought the product to deliver, and ensuring the company learns and improves based on those interactions.

He compared it to medicine. Doctors have always adopted new technologies. Diagnostics have become more sophisticated. AI will undoubtedly improve the speed and accuracy of many decisions. None of that changes the doctor’s fundamental responsibility: understanding a patient’s health and helping improve it. Customer Success follows the same pattern.

I’m not convinced.

AI doesn’t simply give products better tools. Increasingly, the product itself is becoming capable of observing its own performance, learning from that performance and adapting accordingly. The question, I thought, wasn’t whether Customer Success changes, but which responsibilities migrate into the product, and which remain irreducibly human.

I agreed with the principle, but I wondered whether we were describing two different responsibilities.

One is outward looking: helping customers realise the value they expected when they invested in your product.

The other is inward looking: ensuring everything the company learns from those customers feeds back into building a better product.

Both felt equally important to me. Dan didn’t disagree. In fact, he argued that they’d always been inseparable. Customer health has always sat at the centre of Customer Success. The role has never been simply to maintain relationships. It has been to understand whether customers are genuinely becoming more successful, why they are, or aren’t, and what the company should do about it.

That raised another question: If Customer Success is ultimately responsible for helping customers realise value, how should companies think about measuring value when the customer may no longer be the person doing the work?

Observing value

Historically, software has relied on proxies. Logins, clicks, feature adoption; all of them stand-ins for value because value itself was never visible. The assumption was that activity correlated with worth.

But what happens when there are no users? If AI agents are performing the work, activity doesn’t disappear, it moves beyond human observation. A product may soon know far more about the work it performs than any human ever could. A warehouse robot knows every movement it makes. An AI agent knows every decision it takes. They don’t need to estimate activity through logins or clicks; they observe the work directly.

Which sounds like the end of the problem, but it isn’t.

Why? Because observing work is not the same as observing success.

A product may know it has completed a task faster, more accurately and at lower cost. It cannot know whether the organisation has used that improvement well.

I asked Dan how Customer Success should think about value in that world. He argued it had always been difficult.

“We’ve never actually measured value. We’ve measured things that we hope correlate with value.”

ROI models may help during a sales process, but they rarely survive first contact with reality. Every customer defines success slightly differently.

His approach was refreshingly simple, ask the customer this one question: “How will you know this investment has been worthwhile?”

Once you have that answer, identify the behaviours and capabilities most likely to predict it. At Gainsight that became the basis of customer health: a small number of capabilities that consistently appeared in successful customers, used as leading indicators rather than waiting for renewals to reveal whether things had gone well.

But Dan was equally clear about the limits. A customer may be using every feature you expect and still not be successful. They may find the product frustrating, their organisation may have changed, the executive sponsor may have left. None of that shows up in the metrics.

“The signals tell you where to look. The customer tells you what they mean.”

Which took us back to where we started. Perhaps the real purpose of Customer Success isn’t measuring value at all. It’s about preserving the organisation’s ability to see it.

Success is a shared responsibility

Thinking about value inevitably led us to another question. Who is actually responsible for creating it?

One of the recurring themes I see across our portfolio is that customers must be active participants in their own success. They don’t simply buy the solution and start using it.

They connect data, approve agents, redesign workflows and, in the end, change how they operate. The supplier provides the technology. The customer has to change the company around it.

Value is created where supplier capability and customer commitment meet.

I asked Dan how Customer Success should respond. His answer came via a conversation he’d recently had with a founder who described a person he felt he was missing: someone who wakes up every day thinking about whether customers are genuinely getting value. Someone who makes sure the right people are involved, the necessary decisions are being made and both sides are doing what they had committed to do.

Dan’s response was immediate: “You’ve just described a great Customer Success Manager.”

Then he made a point I’d never heard expressed quite so clearly: “One of the most valuable things a Customer Success Manager can do is tell a customer they are unlikely to achieve the outcome they want.”

Not because the product has failed, but because they have. IT hasn’t provided access to the right data. The executive sponsor has disengaged. The implementation team hasn’t prioritised it.

Those are uncomfortable conversations. But avoiding them doesn’t improve customer success, it merely delays failure.

This reminded me of a company I’d been speaking to recently. They only had the capacity to onboard three new customers during the coming year. Three! Which changes the qualification question entirely, because the scarce thing is no longer the prospect’s budget but the supplier’s ability to deliver, and every slot they fill is a slot they cannot offer to someone better suited. So rather than working out how to persuade a prospect to buy, we talked about a different question: why should we choose you?

It sounds backwards until you follow it through.

If, as stated above, value is indeed created where supplier capability and customer commitment meet, then the customer’s willingness to do their share is not something you want to discover during onboarding, it is a qualification criterion. A prospect with budget and no executive sponsor is worse than no prospect at all, because they consume a year of delivery capacity and produce a reference you cannot use. Choosing the wrong customer is every bit as risky as choosing the wrong supplier.

Dan agreed. Great sales organisations don’t simply qualify whether a prospect can buy; they qualify whether they’re likely to succeed.

Does Customer Success become part of the product?

If products are increasingly responsible for creating value, shouldn’t they also become responsible for observing it?

I’ve been exploring this idea for some time under the label native observability: the idea that when a platform does the work, the outcome of that work becomes visible in the platform’s own data, without anyone having to be asked. Mews can see what a hotel earns per available room because every booking, rate change and check-in passes through it. The evidence doesn’t depend on the relationship.

I think products will become the primary locus of operational observation. They will recognise patterns, detect failure, surface opportunities and adapt without waiting for a human to intervene. In most domains they’ll understand far more about their own performance than any operator could.

What I’m less certain about is whether that extends to organisational success.

Dan’s instinct was that he’d seen this movie before. This was the problem Gainsight originally set out to solve: not simply collecting product data, but combining every signal that mattered (usage, support, commercial history, marketing engagement, renewals) into a complete picture of customer health. Could more of that capability move into the product itself? Absolutely.

But he was clear about the edge. Some of the signals that decide whether a customer succeeds have never been in the software at all. Organisational change. executive sponsorship, internal politics, competing priorities. A product can see behaviour. It cannot see the organisation around it.

This is a distinction I’ve been circling. Native observability gets you everything inside the system’s boundary, far more than we’ve ever had. In Mews’ case, their platform can now show that the hotel’s revenue per room went up. What it can’t tell is whether revenue per room was what the group was trying to move this year. It can’t tell if the executive who signed the contract still works there.

Product performance and organisational success are not synonymous.

AI-native products will become extraordinarily good at understanding performance.

Success remains a human judgement.

Organisations aren’t collections of workflows. Leadership changes, strategies evolve, teams restructure, budgets move. Companies sometimes become more efficient without becoming more successful. Those aren’t missing data points, they’re a different kind of information. A product can increasingly understand the system it operates within. It still can’t understand the organisation that created the system in the first place.

That doesn’t eliminate Customer Success, but I suspect it changes where the leverage sits.

A conversation with the founder of a robotics company came back to me. He described the progression from one operator controlling one robot, to ten, then a hundred, and eventually perhaps a thousand. The operator never disappears. Technology extends the reach of their judgement. The robots understand how they’re performing; the operator understands whether the operation is working.

Dan thought Customer Success follows the same path. When Gainsight was being built, he said, the question was never whether technology would replace Customer Success Managers. It was whether it could let each one carry twice as many customers, with better information and better judgement. AI simply extends that trajectory.

Which is roughly where we ended up. We agreed on the destination, but not quite on the mechanism. Dan sees AI increasing human leverage. I’m drawn to the idea that products will take over understanding and improving the work itself, leaving people to judge whether the work is making the organisation more successful. Both may turn out to be true, but either way the centre of gravity is moving.

Where to look, how to think

Products will get better at understanding and improving the work they perform. Neither of us doubts that. Whether they will ever understand if that improvement is translating into success for the customer is a different kind of intelligence: judgement about strategy, priorities, leadership and change, and about the organisation around the product rather than the product itself.

Which brings me back to something Dan said earlier, because I think it was the truest thing either of us managed. The signals are about to get much richer and cheaper. That doesn’t change what they’re for.

“The signals tell you where to look. The customer tells you what they mean.”

what problem do you solve?

who do you solve it for?

what value do you create?

The best founders don’t answer those questions once. They return to them relentlessly, allowing every experiment, customer conversation and product release to refine their understanding of what matters most.

They sound almost too simple to matter. Most founders have an answer to each of them before we’ve even finished our first meeting.

The interesting part isn’t the first answer. It’s what follows.

I’ve come to believe that almost every strategic decision a company makes is downstream of these three questions. Product strategy, go-to-market, pricing, hiring, partnerships, fundraising and even organisational design all depend, ultimately, on how well you answer these three fundamental questions.

If those answers are weak or based on outdated assumptions, everything built on top of them begins to drift.

Which is why we don’t move on to a different set of questions.

We come back to the same ones:

• Is this THE most important problem you could solve?

• Are these THE customers who value that solution the most?

• Is this THE greatest value you could create for them?

Those aren’t refinements. They change the nature of the conversation.

The objective is no longer to define your business, it is to improve your understanding of it.

• Every customer conversation becomes evidence.

• Every product release becomes evidence.

• Every commercial success or failure becomes evidence.

Each iteration should sharpen your understanding of the problem, the customer and the value you create.

Over time, those three questions become less like a checklist and more like a learning loop:

• What problem creates the most value?

• Who values that problem the most?

• How do we maximise the value we create?

The framework scales because it is recursive.

Sometimes you refine the answers. Sometimes you discover there are multiple answers. In either case, every coherent business inside the company must answer the same three questions: what problem creates the most value, who values it most, and how do we maximise that value?

That is exactly what you should expect.

The best founders don’t cling to their first answers. They improve them.

AI doesn’t change this discipline. It accelerates it.

Founders can now build prototypes in hours rather than weeks, test ideas more quickly, speak to more customers, analyse more feedback and iterate at a pace that would have seemed impossible only a few years ago. The learning loop becomes dramatically faster.

That changes the economics of company building.

The cost of building the wrong thing falls because building itself becomes cheaper. The opportunity cost of pursuing the wrong problem, serving the wrong customers or creating too little value rises because you can now move so much faster in the wrong direction.

Sometimes those questions converge on a single, sharper answer. Sometimes they branch into entirely new opportunities, revealing that what appeared to be one business is, in fact, several value-creation systems waiting to emerge.

This is why I increasingly find myself returning to a different question.

What matters?

Not as an abstract philosophical question, but as a practical one.

• What problem matters most?

• Which customers matter most?

• What creates the most value for them?

Everything else follows from there.

Your product roadmap is simply an expression of the problems you have decided matter. Your go-to-market strategy is an expression of the customers you believe matter most. Your pricing reflects the value you think you create. Your hiring plan reflects the capabilities you believe matter most. Your partnerships, your investments and your organisational priorities all emerge from the same underlying understanding.

If that understanding improves, every one of those decisions improves with it.

Then someone I work with said: “Explain the problem to me.”

In trying to explain it, I finally understood what the work actually was.

The task required reconciling almost ten years of management accounts, contracts, correspondence and email into a single, internally consistent picture. My limitation wasn’t effort. It was my ability to hold ten years of unstructured information in my head at once and reason across it. And anyone who knows me will know I never had a chance.

It took me almost two hours to discover what actually mattered and define the problem.

Once I had, the Claude completed the work in minutes.

That experience captures something fundamental about the AI era.

For most of the software era, execution was scarce.

Building software was expensive. Research was expensive. Writing code, creating campaigns, qualifying prospects, analysing markets and coordinating teams all required significant human effort. Companies differentiated by executing better than their competitors because execution itself was difficult.

AI is changing that balance.

Execution is becoming dramatically cheaper. Building, analysing, researching, writing and coordinating are no longer the constraints they once were. As these activities become increasingly abundant, they cease to be the primary source of competitive advantage.

The obvious question is: what replaces them?

The answer, I think, is remarkably simple.

Understanding what matters.

That sounds obvious, but it changes everything.

In product, the question is no longer whether you can build something. The more important question is whether you should and whether building it creates meaningful and compounding value for customers.

The same is true in go-to-market.

AI can help identify prospects, personalise outreach, generate campaigns and analyse performance. But none of those capabilities tell you which customers matters most, which problems they are truly trying to solve or which markets are truly worth pursuing. Those remain questions of judgement.

The pattern repeats inside organisations themselves. AI can summarise meetings, generate plans and coordinate work, but it should not decide what deserves the team’s attention. It shouldn’t be determining which trade-offs matter, which priorities should dominate or which opportunities justify saying no to everything else.

In every case, while execution becomes easier deciding what matters becomes ever more important.

Leaders have establishing a vision and organising execution. Increasingly, they create value by focusing attention.

Their job is to help the organisation distinguish signal from noise, identify what creates the greatest value for customers and shareholders, and ensure that extraordinary execution is applied to the few things that genuinely deserve it, the few things that genuinely matter.

I’ve written before that startups are “machines for discovering truth”. Perhaps another way of saying the same thing is that they are “machines for discovering what matters”.

AI dramatically expands what a small team can build, but it also dramatically expands the number of things that could be built. Every reduction in the cost of execution increases the importance of choosing wisely. The easier it becomes to do almost everything, the more valuable it becomes to understand what is actually worth doing.

AI will continue to make execution faster, cheaper and more accessible.

It will not answer the most important question.

What matters?

Stephen Millard is Partner and Chief Platform Officer at Notion Capital, Europe’s leading early-stage technology investor.

When we think about everything we are taught about building companies, the frameworks, the benchmarks, the motions, are nearly all an attempt to capture what tends to work across enough companies to be named best practice. Which is a problem, because the thing that works across enough companies to be named a best practice is, by definition, an average, a path to the middle of the pack. And in the world of venture capital middle of the pack is failure.

A best practice is, by definition, average practice, and in a world of power laws the average is a mediocrity you should not accept.

This is why I keep coming back to principles rather than playbooks, and I have made the case against the playbook more than one way before: That a playbook is someone else’s answer, derived against a market that was theirs and not yours; that a playbook encodes the limits of a world that has since moved on.

Both are true and both matter. But the power law gives the harder objection underneath both, which is that even a playbook that is yours may still only carry you to the average, and the average is not the game you are in.

Take go-to-market, where the playbooks are most lovingly maintained.

I recently wrote up what is actually working with AI in go-to-market, drawn from a Notion Capital workshop led by Harrison Rose, together with Al Simpson and Even Walser, and the example that has stayed with me is the most ordinary one imaginable: targeting customers by their job title. We have done it for decades, not because the title ever told us that someone would buy, but because it was something we could hold on to. Asking “who is actually responsible for the pain, who has the need and budget?” was not something a team could do at scale, so the title became the stand-in.

And here is the part worth considering. The job title was never best practice. Nobody ever proved it was the right way to target a market; it was simply inherited, passed from one company to the next until it felt like the natural order of things. That makes it received wisdom, the more dangerous relation of best practice. Because best practice at least had to work at least once to earn the name, whereas received wisdom only had to go unquestioned long enough that we forgot it had ever been a choice.

Best practice carries you to the average. Received wisdom cannot even manage that.

What was fascinating about the discussion is that the constraint which forced the workaround in the first place is dissolving.

Even Walser described the shift as moving from a system of record, to a system of context, to a system of action, and making the business readable in that way is the precondition for all of it. I have written about why most companies don’t yet have that readability. But readability used just to run the old motion faster is not the prize. Ask rather who is responsible for the outcome you seek and the proxy disappears, because you can now read responsibility directly, from reporting lines and tenure and what someone actually owns. The annual process of building target personas, the quarterly forecast, the experiment gated behind weeks of coordination: each was received wisdom built for a world of scarce information and even scarcer ability to execute.

In short, received wisdom is a simplistic answer to yesterday’s question.

What replaces it is the discipline of deriving your own buying personas against your own market and your own customers, and then proving the result honestly, by seeing whether customers are genuinely and measurably more successful because of you in their own data. That is harder and slower and far less comfortable than adopting the thing everyone agrees on, and it is the only work with any chance of making you the exception to the rule.

None of which means everything inherited is worthless.

Some of it encodes hard truth, paid for in other people’s failures, and discarding it on principle is its own kind of foolishness. Knowing which to keep and which to throw out is judgement, and judgement is the part that remains stubbornly human. But the default posture has to flip. Not “this is best practice, so adopt it,” but “this is best practice, which means it is average, so what would beat it?” And when the thing turns out to be received wisdom rather than best practice, the question is sharper still, because you are no longer weighing the average against the exception, you are asking whether it was ever true at all.

Be wary, above all, of anyone sure they already know.

The most useful sentence I have is still “I don’t know.” It is the honest starting point for the only question that matters, which is not how to reproduce what worked for the average company, but what the exception would do here, now, in a game whose rules are still being written.

Adopt the best practice, absorb the received wisdom, and you will, reliably, arrive at the average.

Stephen Millard is Partner and Chief Platform Officer at Notion Capital, Europe’s leading early-stage investor in B2B software and AI-native companies.

But I have always thought one line is back to front:

I need you more than want you, and I want you for all time.

Surely everything we know about ourselves and people in general is that desire comes first and need follows, so in my head I have always sung it backwards:

I want you more than need you, and I need you for all time.

The more time I spend with AI-native companies, the more my backwards version describes the ones growing fastest right now.

The want comes first.

The need follows.

And then, if you do the work, the need fuels more desire.

Adoption begins with want, not need

Look at how the fastest-growing AI products are actually being adopted. Someone tries one, falls for it, and is using it daily before anyone has written a business case.

Lovable is the clearest example I can point to: a product people fall for long before procurement knew it existed, spreading from one person to the next on something much closer to wonder and delight.

For thirty years B2B software was built to stop exactly this. The buying process existed to filter out desire and focus only on business needs. BANT, SPICED, MEDDIC all follow this similar path and we were all trained to ask whether the organisation genuinely requires the thing and whether it was ready to buy.

That world is changing.

The user has displaced the buyer and want is displacing need as the leading edge of B2B tech adoption.

The reflex is to call the products winning this way sub-standard, and by the old definition of business-ready maybe they are, lacking the single sign-on, the audit trails, the reassuringly slow implementation.

I could make those challenges of Claude.

As an enterprise-ready product it lacks much I expect. But I want it and so too do many of my colleagues. We started with one user, me, and now have an enterprise licence. It still falls down in many ways, but we now all want it and, of course, need it.

But, want alone won’t keep a customer

Want is a wonderful way in, but on its own it will never hold a customer in place. After all, desire is fragile. The tool everyone loves today is the one nobody mentions tomorrow, often with nothing in the customer’s actual situation having changed to explain their dampened ardour.

A company that reads the wanting as a settled relationship is building on fickle affection and so nobody asks the only question that matters over time:

When the wanting fades, what holds the customer in place?

The “aha” turns want into need

I was working through this with Dave Peterson, former CMO of Celonis and Author of Play Bigger and he put it differently:

Companies sell on needs, but the products that endure are the ones that create wants from those needs, because the want is what motivates long-term consumption.

That sounds backwards too, until you see what want he wants :-)

The want at the start is a pull the customer feels before any proof. The want Dave is describing is the attachment that forms afterwards, once the need has been fulfilled and the value shown.

This circles back to my recent article on native observability.

Desire may create adoption, but observable transformation, the evidence of real need, is what turns adoption into enduring power.

As Dave describes:

The thing that carries you from one customer to the next is the “aha moment”. At Celonis it was the moment a customer saw, for the first time, the inefficiency they had been living with and were blind to: the leakage, the broken process, the money walking quietly out of the door. Once a customer sees that, they cannot unsee it.

The need is now obvious, proven and theirs.

Proven need hardens into deeper want

The proven need, delivered again and again, hardens into a new want.

It sounds like this:

I can’t imagine working without them.

I have written before that this sentence is what you are really building towards, and that when a customer says it, means it, and can put hard numbers behind it, you have power.

Now the want is no longer the unearned pull from the front of the relationship.

It is a want underpinned by proven need.

The danger is mistaking the first want for the last

Which brings the founder to a question worth being ruthless about, in the spirit of the first principle: you must not fool yourself, and you are the easiest person to fool.

It is easy, when growth is fast and the desire is loud, to believe you have reached nirvana, while in reality you have just started dating.

Lovable and Anthropic are the sharpest test of this in the market today, precisely because they have manufactured want at an unheard-of scale.

The question is not whether people want it; plainly they do.

It is whether that want has yet passed through an aha moment and come back as need. Whether the customers reaching for it could, six months on, point to something measurably changed in their world that they could not imagine giving up.

You tell the difference if you can actually see the value you claim showing up in the customer’s own world and their own data.

The companies that win today are creating extraordinary levels of want. The companies that endure will be those that can make the resulting transformation natively observable.

That is what separates a customer who wants you from one who wants you and needs you for all time.

Stephen Millard is Partner and Chief Platform Officer at Notion Capital, Europe’s leading early-stage investor in B2B software and AI-native companies.

NRR is misunderstood. It isn’t a growth metric, it’s a power metric and a lag indicator.

The leading indicator is customer transformation: what measurably changed in the customer’s business because of you. Most vendor ROI claims are performative.

What separates evidence from eloquence is native observability. AI moves work inside the platform; the outcome must move with it.

Expansion isn’t transformation

In October last year I wrote The Measures of Success in the AI Age. The argument still stands. Speed and expansion matter. Defensibility and power matter. NRR is the honest metric. Free cash flow is the endgame.

But, there is a layer above it that I failed to consider.

NRR tells you customers stayed and bought more. It doesn’t tell you whether you actually changed anything. That distinction, between customers who stayed because the alternative was unthinkable, and customers who stayed because the renewal was deemed painful, is what separates companies with real power from companies that just look like they have it.

NRR has been mis-filed

The metric literature is well-developed in two layers and absent in a third.

Growth efficiency is mature. Bessemer’s Efficiency Score, the Burn Multiple, CAC payback and Magic Number. These are all credible sources, all measuring the same underlying question: how much capital are you burning to add a dollar of new revenue. It’s an acquisition layer.

Operational efficiency is also mature. ARR per head, gross margin, opex ratios, free cash flow margin, the Rule of 40. How efficient is the engine at any given growth rate. An operating layer.

The third layer - the measure of power - has no formalised metrics. Hamilton Helmer’s 7 Powers is the closest theoretical anchor. Helmer’s Fundamental Equation of Strategy gives us a way to think about market scale and differential margin. But no operator-facing metric pack measures power directly. ICONIQ, the most credible source in this space, has started to acknowledge the gap, naming “product engagement and business outcomes” as the emerging signals of value in the AI age.

NRR: a power metric hiding in plain sight

NRR has lived in the middle of every metric pack for a decade, but noone really knows how to describe it.

It isn’t a growth efficiency metric; the growth efficiency cluster measures acquisition.

It isn’t an operational efficiency metric; operational efficiency measures productivity.

It’s a power metric: NRR above 100% is the commercial fingerprint of customers who expand because they can’t or won’t leave.

This reframes the architecture, into three layers, not two:

• The growth layer: how much capital you burn per dollar of new revenue.

• The operational layer: how efficiently the engine runs at scale. ARR per head, gross margin, the cost structure of the business.

• The power layer: how durable the revenue is, and how persistent the value and the market cap it produces, with NRR as the lag indicator.

The leading indicator is transformation

If NRR is the lag, what’s the lead? Customer transformation, that’s what.

In short, what has measurably changed in the customer’s business because of you. Not what they used or how often they logged in, but what changed. Pricing realisation, gross margin durability, win rate against incumbents all sit here too as lag indicators of power.

This is the layer most companies capture anecdotally at best.

Real transformation or anecodotal?

The risk in writing this piece is that “Transformation,” “ROI” and “Impact” are exactly the words software vendors have been spraying across customer success decks for ever.

As a former CMO, guilty as charged. We built ROI calculators to flatter ourselves, with case studies cherry-picked to impress our investors.

Let’s be honest, most of it is performative.

The discipline that distinguishes real impact from anecdotal is a single question: would a sceptical third party, the CFO who didn’t sign the contractor or the new CEO who inherits the vendor list, credit you for the change you claim?

Customer transformation must become “natively observable”

If impact data relies on the customer’s own attribution being generous, it is vanity. If it survives someone whose only fealty is to observable data, it is real.

That test is harder than it sounds. What makes it pass is whether the transformation is natively observable, i.e. visible in the system’s own data, in real time, because the system is doing the work that produces the outcome.

I’m borrowing the word observability from software engineering where it has a precise meaning; the ability to understand what’s happening inside a system from the signals it produces. The idea transfers neatly I think . A platform with native observability of its customer’s outcome is doing something similar: reading the impact on the customer’s business from the results the platform itself is producing.

Native observability is what separates evidence from eloquence. Where it exists, the sceptical third party can run the test for themselves.

Examples that pass muster

Mews is the cleanest example in our portfolio. Mews is the operating system for hotels; every booking, every payment, every rate change, every check-in, every housekeeping task flows through the platform. Which means the customer’s outcome is natively observable inside Mews itself. The platform sees how much revenue each available room earns, how much profit survives the cost of running it, how much of the business comes direct rather than through expensive third-party booking sites, how well staffing matches demand; all in real time, in the system’s own data, because the system is doing the work that produces it. A new CFO inheriting the contract doesn’t need to take anyone’s word for whether Mews is delivering results. The performance is right there, in the platform, and can be set against industry benchmarks for context. The attribution does not depend on the relationship. The transformation either shows up in the data the platform is producing or it does not. Notice what Mews is selling: not capability, i.e. “you will be able to do X,” but operational fitness for a competitive market. The customer is buying the ability to run a better business. That is the natural language of a power proposition applied at the customer level.

Palantir illustrates a different facet; institutional transformation rather than transactional throughput. Foundry and AIP (Palantir’s data platform and its AI layer) don’t sit in the customer’s stack; they restructure how the customer operates. Decisions get made through the platform, not next to it. The transformation shows up in decision making, cycle times, coordination, all observable to a sceptical CIO without anyone needing to flatter the contract. The market doubted Palantir for years because it was reading SaaS signals while institutional dependency was forming underneath. The re-rating, when it came, wasn’t the market predicting the future. It was the market finally recognising operational leverage that already existed. NRR was the lag indicator of a power that had already formed.

Adyen illustrates a different form of SaaS expansion economics: measurable operational improvement tied directly to customer revenue. Their proposition is that tighter integration across acquiring, routing, fraud, and orchestration improves payment performance, particularly authorisation rates. For merchants operating at scale, even modest improvements can translate into millions in incremental revenue. Crucially, these outcomes are measurable. Merchants monitor authorisation performance obsessively because it sits directly on the revenue path. Over time, Adyen’s expansion dynamics suggest that customers continue to see enough economic value to deepen the relationship across markets, products, and payment flows. That is materially different from expansion driven primarily by seat growth or contractual mechanics. The vendor grows because the customer’s underlying business performs better. In fintech infrastructure, that is one of the strongest forms of retention there is.

What links the three? In each case the platform is the path to the outcome, not adjacent to it. The transformation is something the system produces and the customer can see.

This also manifests in market-mispricing

If power is what produces durable revenue, i.e. the ability to defend prices and margins, to expand existing accounts, to win the best customers, then revenue from a power-rich company is worth materially more than revenue from a power-poor one. The first will persist and compound. The second will erode under competitive pressure.

Markets are not recognising this distinction, rather they are pricing growth, revenue and operational efficiency. They are not pricing power, because the metrics that would let them price power are not yet formalised. So they average, meaning companies with strong customer transformation, durable pricing and a base that compounds get valued on the same revenue multiples as companies that look identical on the surface but whose revenue base will fragment the moment a sharper competitor arrives. The first group is underpriced. The second is overpriced. The mean is roughly right and tells you nothing about either.

The 2022 SaaS reset showed this in slow motion. Companies whose growth had outrun their retention were marked down severely once NRR became visible. Companies whose growth was anchored in NRR kept their multiples.

In the AI age, the gap between power and its lag indicators is widening, not closing. Revenue is being added faster than ever. The data that would tell you whether any of it will persist is being generated faster than ever too, but only by the companies that have built the discipline to capture it. The market will eventually re-rate around this, as it always does.

Why this piece and why now?

There is a reason this is now a piece worth writing.

Native observability is no longer optional. A small number of platforms, the systems of record at the centre of their customer’s operations, could see whether the outcome was happening from inside their own data: Mews and Adyen do. Palantir, once it is deeply deployed, can.

That privilege is no longer optional because AI is changing where work happens.

When the system does the work (an AI SDR running the outbound, an AI agent drafting the contract, an AI analyst building a model) the work moves inside the platform, and the outcome must become directly observable therein.

The implication for measuring customer transformation is clear. You can read usage continuously, synthesise CS call transcripts against outcomes, compare expanders against non-expanders inside the same segment and surface the differences in days rather than quarters and you can do so for every customer - large or small. And the cost and time of doing so have collapsed.

The companies that exploit this will pull away from the ones that don’t. Not because their customers transform more, in fact their customers might transform exactly the same amount, but because they will know it and will be able to feed the evidence back into the system that generates the next sale.

What it changes downstream

If transformation is upstream of NRR, the implications for how the company reads its own data are significant.

The ICP-as-posterior argument I made perviously was that you derive your ideal customer from the customers who actually expanded. That holds, but the loop should be reading transformation data as well as commercial data. Otherwise you are deriving from a posterior that overweights customers who renewed for reasons unrelated to the value you delivered and underweights customers who genuinely restructured their business around you but happened to expand on a slower cycle.

The same is true of the derivation loop more broadly. Hypothesise, instrument, evidence, update. If the evidence layer is only commercial, the GTM system you derive will be a system that optimises for renewal and expansion probability, not for transformation probability. Those look identical for a while, until one day they don’t.

Where this leaves us

NRR remains the lag metric of commercial compounding.

What sits above it and tells you commercial compounding will continue, is the lead metrics of customer transformation, measured by observers with no incentive for flattery.

Native observability makes that measurement possible without the vendor telling the story.

The companies that understand this will know things their competitors don’t.

1. They will see drift months before the renewal data shows it;

2. They will price their own future revenue more accurately than the market does and have a clear narrative to assert it; and

3. They will derive a sharper ICPs, sharper growth motions and a sharper product roadmap.

Ultimately, the future belongs to the companies whose power is evidenced directly from the transformation they deliver.

Stephen Millard is Partner and Chief Platform Officer at Notion Capital, Europe’s leading early-stage investor in B2B software and AI-native companies.

I’d had echoes of it earlier at Oracle, bigger company, more layers, more people, but at MessageLabs it landed harder, because of the speed we were moving at.

This is the thing the productivity narrative misses. AI doesn’t relieve cognitive load, it concentrates it.

When a hundred people did the work a handful of people now do, the load was distributed. Bad days for one person didn’t drag the system. The friction of coordination was real, but so was the redundancy. Now that ten people carry the same load the friction is gone but so too is the redundancy. Every person is closer to the edge every day.

This is a human performance story

We talk about this as if it’s a productivity story. It isn’t, it’s a human performance story, and it has specific components.

Decision volume goes up per person, not down. AI handles more of the execution. The decisions that remain - what to build, who to hire, what to stop, what to bet on next - are the ones AI cannot make for you. Each individual on a lean team makes more of these decisions, in less time, with more consequences.

Context-switching cost rises. When ten people own what used to be ten teams, each person carries multiple domains at once. They are constantly re-entering work they last touched yesterday or last week, paying the cognitive tax of remembering where they were each time.

The rate of unlearning accelerates. I have written about reinvention as the operational reality of AI-native companies. Ship of Theseus from the outside: the company looks unchanged. Ship of Theseus from the inside: the humans on board are constantly being asked to unlearn what they were doing well three months ago and learn something new.

The celebration of reinvention is honest. The cost of reinvention falls on the people inside, and we rarely name it.

The emotional weight. When ten people carry what a hundred used to, every outcome is closer to every person. There is nowhere to hide. The CSM who would have been one of fifteen is now one of three. The PM who would have been one of twenty is on their own. Accountability is massively concentrated.

This is what makes the cognitive load extraordinary. It isn’t that there is more work. It is that the work, the decisions and the responsibility all sit with fewer people.

In lean teams, clarity, cohesion and alignment matter more, not less.

The cost of misalignment is no longer one of a hundred people pulling slightly awry. It is one of ten. This is one of the reasons we continue to invest in working with The Table Group, continually revisiting their insights for creating clarity and building high performance teams.

All this means the work of leadership in a lean AI-native company is not what most founders trained for. The instincts from the SaaS era to hire ahead of the curve, layer in management, distribute the load do not work when the company is lean by design. You cannot hire your way out of the cognitive load problem, you can only work hard to mitigate it.

What that looks like in practice is unglamorous:

• It looks like saying out loud what matters this quarter and what does not, and saying it often enough that it begins to feel unnecessary.

• It looks like protecting time for reflection and deep thinking.

• It looks like making fewer, better decisions.

• It looks like being explicit about who does what, because in a ten-person team there is no room for the implicit assumption that someone else has got it.

• It also looks like a different kind of care for the people on the team; designing the work so the load is bearable, while watching for the signals that someone is past the edge.

We are non-linear

It is vital that we remember that humans do not scale linearly, and that the ones who appear to are possibly the ones closest to breaking.

The companies that win this era won’t just be those that operate lean. They will be those that understand the cost of operating lean and design for it by being clear about what the lean team is being asked to carry, and equally clear about what it is not.

I have written elsewhere about velocity outrunning judgement at the founder level. This is the team-level version of the same problem. The founder’s judgement degrades under load. So does the team’s. And once the team’s judgement degrades, the company drifts in ways that take longer to detect and longer to fix than a single founder making a single bad call.

The binding constraint in the AI-native company is no longer how fast the machine can adapt.

The binding constraint is how fast the humans inside it can adapt.

Stephen Millard is Partner and Chief Platform Officer at Notion Capital, Europe’s leading early-stage investor in B2B software and AI-native companies.

In the SaaS era we got away with this fiction. the world followed a relatively sequential model - MVP, PMF, GTMF. But in this era telling yourself you have reached product-market fit is akin to a form of self-deception.

Marc Andreessen coined the phrase product-market fit in 2007, defining it as “being in a good market with a product that can satisfy that market.” But implicit in that definition was the idea that both the product and the market were stable enough to fit together. In reality they never fully were. In the AI era they categorically are not. Products evolve continuously, markets adapt in real time, and customer expectations shift faster than organisations can institutionalise what they have learned. Product-market fit is no longer a destination, rather it is a moving feast.

A market is not a fixed thing you can fit a product to, rather it is a moving set of buyers, with shifting needs and emotions in a changing context, who are themselves being reshaped by the technology, products and hype surrounding us all.

Similarly, a product is no longer a fixed thing you can hold up against the market. A product is a moving set of capabilities, being reinvented as the technology underneath it changes. I’ve written previously about products having no permanence. The idea that you can hold one up against the other and declare “fit” assumes a stillness on both sides that does not exist.

In the SaaS era the approximation worked. Products and markets moved slowly. Founders could hold both still in their minds long enough to believe they had fitted them together. The fiction was useful, but it was still a fiction.

In the AI era the approximation has stopped working. Both sides are moving so fast that the fiction is actively misleading. You cannot fit a product that is being rebuilt quarterly to a market that is emerging or evolving as you sell into it. There is no still moment in which the fitting happens. The founder who says “we have reached PMF” is describing something they cannot be seen and doesn’t exist.

Product-market fit was always a fiction. In the SaaS era it was a useful one. In this era it is a dangerous one.

Richard Feynman wrote that the first principle is that you must not fool yourself, and you are the easiest person to fool. PMF in the AI era is the easiest fooling there is. It gives you a clean narrative. It rewards you with relief. It tells your board what they want to hear. It lets your VCs put a milestone in their LP update. It is also not true.

What founders are actually doing when they declare PMF is taking a slice of evidence (some buyers signed, some logos landed, some early success looks good) and treating it as if it confirms a state of the company. It doesn’t. It confirms that one cohort of buyer, in one configuration of the product, in one state of the market, in one month, behaved in a particular way. That reading will be different in a few months, weeks or even days.

The damage of treating the reading as a state is the same damage I have written about elsewhere. Founders who think they have reached PMF stop running discovery loops. They stop interviewing customers. They scale at velocity while the fit underneath them quietly degrades. Their VCs encourage this, because the language is the language. Pre-PMF is risk. Post-PMF is execution. The company is now in execution mode. Head down. Scale. By the time reality hits, six months have gone by and the company is rebuilding from a worse position than it started.

What is actually happening is something more honest and more useful:

1. Some segments are working. Some are not.

2. Some configurations of the product are landing. Some are not.

3. Time to value is compressing for one cohort and lengthening for another.

The work is to read those signals and to do so continuously by segment, and then to update the company’s posture on the evidence.

That is not PMF. That is the job. The thing the PMF language has been hiding from founders for fifteen years.

I argued in “Your ICP Is a Posterior, Not a Prior” that your ICP is derived after the fact, modelled on the customers who actually succeeded and expanded. The same is true here. There is no PMF to declare. There is only the continuous, by-segment reading of what is actually happening. The founders who do that work are the ones still expanding two years on. The founders who declared PMF and went heads-down into execution are the ones explaining to their boards why the numbers are softening.

I am going to stop using the phrase. The industry will not stop overnight, it is far too entrenched, but I refuse to use it as if it describes a real thing. The next time a founder tells me they have reached PMF, my response will be the same…

The cost of language is not abstract. Every month a company runs on the PMF-posture is a month it is not reading the evidence honestly. By the time reality hits, the fit has already shifted. The phrase has cost more founders more time, money and lost opportunity than any single hiring mistake or product decision I have seen.

Product-market fit was always a fiction. In the SaaS era it was a useful one. In this era it is a dangerous one. The day you tell yourself you have reached it is the day you have started to fool yourself.

Stephen Millard is Partner and Chief Platform Officer at Notion Capital, Europe's leading early-stage investor in B2B software and AI-native companies.

Take Aikido, one of Europe’s hottest unicorns. Enterprise cybersecurity had a settled playbook which Aikido ignored, because it didn’t match their read of where the market and buyers were going.

They made a series of “counter-intuitive bets:” PLG in a category built on top-down CISO selling; Public pricing in a category built on bespoke quotes; Helping customers buy on their agenda and timeline in a category trained to push on its own.

Underneath, a deeper shift: AI was accelerating software production faster than the fragmented security stack could absorb, and Aikido collapsed the complexity before the market knew it needed it.

That is the difference between a derived GTM and a contrived one. A contrived one assumes a certain way of doing things. A derived GTM reads the buyer, spots the shift before incumbents, and builds the motion the market is pulling toward based first on intuition and then backed by evidence.

The most dangerous GTM playbook is the one that sounds most credible. It might come from your last company, a trusted advisor, an investor director, or the standard operator canon.

Credibility is exactly what makes it dangerous and it stops you interrogating the assumptions underneath. Different product. Different market. Different customer maturity. Different time. The playbook that won “then and there” is worth studying, but not copying.

The GTM system is a posterior too

In “Your ICP is a Posterior, Not a Prior,” my argument was that your ideal customer profile isn’t the hypothesis you write at the offsite. That’s where you start, but your true ICP is something you discover after the fact, modelled on the customers who actually expanded. A posterior, not a prior.

The same logic runs one level up: the go-to-market system itself is a posterior.

GTM systems aren’t templates or models you choose. They’re posteriors you derive. The best systems for your company emerge from running tight loops - intuition, hypothesis, evidence, update - against your actual market, your actual customers, your actual product.

Import one wholesale - from your last company, from a competitor, from an investor’s pattern-match - and two things happen. You substitute someone else’s posterior for your prior. And you skip the work that would have built the muscle to run the system once it’s yours. That muscle is what compounds. The system is just it’s manifestation.

That’s the difference between contrived and derived. A contrived system is inherited. A derived system is earned.

The derivation loop

Deriving a system isn’t vague, and it isn’t necessarily slow. It’s a loop with four moves.

Hypothesise. Name what you currently believe: how you sell / who buys, why, how, when, at what price, through what motion. Make the beliefs specific enough to be wrong.

Instrument. Decide in advance what would tell you they’re wrong. Build the measurement before you build the system.

Evidence. Run with your eyes open for what the market is actually saying. Disconfirming evidence is worth more than confirming evidence.

Update. Change the hypothesis with evidence.

Run 10 / 20 / 30 discovery meetings and you have human pattern recognition. ‘This part of the market is not that warm. Let’s move on.’ Five out of 10 customers are unhappy. ‘Let’s dig deeper here.’ Run a short feedback loop with product and engineering: ‘Why are they unhappy? It’s because of this feature lacking. Can we build that?’”

Hypothesise, instrument, evidence, update.

You loop this discovery process on a given part of the system until it stops surprising you and then you develop a durable infrastructure around the part that works, and keep the discovery loop running on the parts that haven’t.

The templates stopped standing still

This has always been true.

ICP and GTM are unstable by design, in the words of Daniela Amodei, “things will never be chill again.”

When Kelly Loftus joined Anthropic to build go-to-market, she didn’t install a quota system, because adoption couldn’t be forecast. She built the team around feedback instead, “shadow targets” rather than quotas, designed to scale from ten people to hundreds without the structure a SaaS playbook would have called for. Anthropic’s go-to-market team has since gone from fewer than ten people to more than 150 - source.

Anthropic services demand it can’t forecast. Aikido fulfils demand on the customers own timeline. The same instinct, help the customer buy, derived from opposite market structures, neither one copied. Neither of these companies is running a playbook; but each is running loops based on bets. That is the trap in benchmarking against the best: in an AI-native market they have no settled motion to hand you. But they have methods you can copy.

Contriving looks like wisdom

Contriving rarely looks like the laziness it actually is. Paradoxically it looks like wisdom. Five patterns to watch for in the people you hire, and in yourself:

The previous-company playbook. “We did this at my last company and it worked.” Maybe. Re-derive to be sure the conditions still hold.

The investor recommendation: this is working really well at one of our other companies , you should try it. Treat it as a hypothesis to consider, not a directive to follow.

The VP importing a methodology from their last three companies is more dangerous than the junior still learning it.

The framework as authority. MEDDIC is a fantastic tool for diagnosis, not a playbook to follow.

Activity mistaken for evidence. Meetings run, calls made, demos given. The metrics that matters are what you learned, not what you did.

The discipline runs the other way too.

Deriving can mean choosing not to build the thing the orthodoxy says you should. Should an agent handle customer reviews? Perhaps. But most founders will decline, preferring the human touch.

What transfers?

So if the system doesn’t transfer, what does?

The willingness to sit in uncertainty long enough for evidence to arrive. The patience to hold conviction lightly until it’s earned. The discipline to write hypotheses down so you can tell when they were wrong.

It is also the part AI can’t do for you. AI shortens the loop - evidence is cheaper to gather and faster to read. AI is really good at improving existing things. But when you’re trying to develop a GTM System, an AI is not going to tell you, ‘No, you should try this business model.’

That is the trap with a contrived system. You can be very efficient at doing the wrong thing. You can be very efficient driving your sales organisation in the wrong direction.

The previous journey taught you a lot. But treat it as a prior to test, not a conclusion to execute. The system you ended up with was not the system to adopt again. But the muscle that derived it is.

That’s a prior. And it’s almost always wrong. And that’s OK.

The ideal customer profile that matters is the one you discover after the fact — modelled on the customers who actually expanded. Not the ones you hoped would. Not the ones who looked like they should. The ones who did.

Your ICP is a posterior, not a prior.

The bowtie was always a loop

Jacco van der Kooij reframed SaaS GTM with the bowtie: the funnel was a story about acquisition, the bowtie a story about success. Left side wins customers. Right side delivers the value. We’ve repeated that language with founders for the better part of a decade, and it’s still right.

But the line of his that has aged best is quieter: an ideal customer is one that expands.

Not one you can win. Not one with the right job title or the right industry code or the right initial use case. One that expands.

That single shift - from “who can we close” to “who will compound” - is what turns the bowtie from a static diagram into a loop. Because once expansion is the criterion, every cycle of the GTM machine produces evidence about what your ICP actually is, not what you thought it was.

From hypothesis to evidence

When you start out, you have to guess. You write down who you think will expand: the segment, the size, the trigger, the use case. That’s a prior, a necessary fiction. After all you have to point the company somewhere.

But the company that beats you doesn’t run on its prior for very long. It runs on its posterior.

After the first ten, fifty, two hundred customers, the data is there to be read:

• Who actually expanded?

• What did they have in common - not on paper, but in their behaviour, their onboarding pattern, their time to value?

• Which of them looked identical to customers who didn’t expand?

That last question is where most teams underread their own data.

The posterior is built from the contrast between expanders and look-alike non-expanders, not from the expanders alone. Two customers in the same segment, same size, same use case; one tripled, one churned. That’s not noise. That’s exactly where the ICP lives.

The posterior is what you target with, based on what you have learned, not what you hoped for at the offsite.

You don’t have one posterior, you have many

Here is what most GTM models miss. A real company isn’t running one loop on one posterior. It’s running multiple loops on multiple posteriors, and each one is at a different level of maturity.

At any given time you’ll have:

• Segments in discovery. Early signal, qualitative posterior, a handful of expanders and a thesis. The question you’re asking is is there something here? The work is interviews, pattern recognition, and the discipline of writing the posterior down even when it’s based on five customers.

• Maturing segments under development. The pattern is forming. The question shifts to what does this segment need to expand reliably? The work is tightening how we sell and customers buy, sharpening onboarding and messaging, refining pricing, watching time to value compress. The posterior is hardening but still being calibrated.

• Proven segments in distribution mode. The posterior is sharp. The question becomes how fast can we find many more of these? The work is acquisition at scale - outbound, partnerships, content, channel, pointed at the highest-resolution version of the ICP you have.

These three loops run concurrently. Not sequentially across the company, concurrently across segments.

The mistake most teams make is to treat the company as if it has one maturity level. Half the time they’re behind on their strongest segment and ahead of themselves on their weakest - pushing a distribution motion at a posterior that’s still in discovery, or running another founder-led discovery cycle on a segment that’s already been figured out.

• A young company has one posterior, in discovery.

• A growing company adds a second under development.

• A scaling company is running all three loops at once: discovering, developing and distributing.

The transition from one to the next isn’t a stage you graduate through, it’s a portfolio you accumulate.

Where AI actually changes the loop

The reason this is becoming a competitive differentiator now, rather than a permanent feature of how good companies have always operated, is that the cost of running this portfolio of loops has collapsed.

Three concrete shifts:

Signal extraction. Synthesising expansion patterns from CS calls, support tickets, and product usage used to be a quarterly project for a small team. It’s now something you can run in real-time, in software, against every account in every segment. The pattern that took a year to surface in 2006 surfaces in a fortnight in 2026.

Concurrent loops at full intensity. Running “discover, develop and distribute” motions at the same time used to mean spreading a finite team thinly across all three. AI lifts the constraint that made attention scarce. You can hold a discovery thesis for a new segment, a development sprint for a maturing one, and a distribution motion for a proven one - all at full intensity, all in parallel, many times over.

Closing the loop into the org. What you learn about each posterior has to change targeting, onboarding, messaging and pricing for that segment specifically. AI is making the connective tissue between “we learned this about segment B” and “the system now behaves differently” thinner. Briefs, pitch updates, onboarding paths and lead scoring can move per-segment, at the pace of the learning, rather than at the pace of the next company-wide planning cycle.

The principle - your ICP is a posterior - is old. I did this work at Oracle in the 90s and MessageLabs in the noughties. What’s new is the speed at which posteriors can update, and the fact that you can now run a portfolio of them concurrently without the org seizing up. That’s the unlock and it’s genuinely incredible.

The hard part is organisational

Most teams that fail at this don’t fail at the principle. They fail at the wiring.

The posteriors have to live somewhere. Someone has to own them. CS has to feed them. Marketing has to read them. Product has to weight roadmap decisions against them. RevOps has to score against them. Without that wiring, ICP is a slide deck, not a system, and the loops never compound.

The most underrated GTM hire in a Series B company isn’t another AE. It’s the person who owns the posteriors; the synthesis between what CS sees, what product measures, and what marketing targets, across every active segment. Call them what you like: Chief Posterior Officer, if you’ll allow it.

The job of this CPO isn’t to maintain one document. It’s to manage a portfolio: which posteriors are still in discovery, which are developing, which are ready for prime-time distribution, and which are starved of attention as they don’t stack up.

And then, eventually, you commit

There’s a moment in every segment’s life where its posterior becomes conviction. The data is sharp enough, the pattern is clear enough, and the right move is to stop running discover on that segment and start betting hard.

But that moment arrives per-posterior, not company-wide. You commit to distributing segment A while you keep discovering segment B and developing segment C. The discipline is knowing which is which.

That’s a different problem - the commitment problem - and its probably worth its own post. But it only arrives if you’ve been reading the posteriors all along.

The principle, distilled

Most GTM systems are still designed to win customers. The best are designed to learn from the ones who expand, find more like them, and run that play across multiple segments at once.

Your ICP is not a hypothesis you start with. It’s a portfolio of answers you arrive at, again and again, by reading what the company has actually produced. Done well, the loops compound: sharper targeting, faster onboarding, higher expansion probability, and a clearer view of which segment to bet on next.

The principle is old. The velocity is new. The concurrency is what makes it feel different in 2026. The discipline - to keep updating the posteriors, and to keep rewiring the org around them - is the work.

Read this next…

Stephen Millard is Partner and Chief Platform Officer at Notion Capital, Europe’s leading early-stage investor in B2B software and AI-native companies.

At it’s heart is the simple tenet: Explore / Exploit.

Marc Roberge neatly calls the sequence PMF, GTMF, Growth and Moat. I had my own: Start, Build, Scale.

What we now realise is that for AI-native companies, that sequence is irrelevant.

What replaces it is three streams, running concurrently, though at different degrees of intensity, from day one.

1. Discover — the work a company does on itself to find what’s real: the problem we solve, the customer we choose to serve, the value we intend to deliver. Are these worth betting on?

2. Develop — turning that discovered truth into repeatable motion that survives without the founder in the room.

3. Distribute & Compound - maximising distribution to reach every buyer while compounding value through data flywheels, reference gravity, capability stacking.

These are not three phases you graduate through, rather they are streams that intensify across a company’s life.

Discover is near full volume pre-PMF. Distribution is maxed out post GTMF. But none of the motions are ever fully abandoned. Nothing graduates. Nothing gets abandoned. The streams rebalance.

Here’s why the sequence breaks.

In prior software waves, you could hold one thing still while you worked on another.

If the problem was understood, you could focus on product.

If the product was stable, you could focus on establishing a repeatable GTM.

If the category was mature, the buyer knew what they were buying, so you could focus on maximising scale.

Sequencing was possible because the world was standing still.

AI-native companies get none of that:

• The category is forming while you sell into it.

• The product surface is shifting under you as model capabilities improve month by month.

• The buyer’s understanding of what they need is changing faster than your pitch deck.

Multiple moving reference frames, all at once.

In that environment, the explore/exploit tension stops being a philosophical question about balance and becomes the operating condition of the company.

You cannot exploit what you haven’t explored; the ground moves too fast. There’s no point exploring if you never exploit; the window closes too fast.

Sequential GTM assumes a stable world to sequence against which founders no longer have.

So the work has to be re-described.

Not as phases you move through, but as streams that run concurrently.

If you enjoyed this article, go here next:

Stephen Millard is Partner and Chief Platform Officer at Notion Capital, Europe’s leading early-stage investor in B2B software and AI-native companies.

AI makes information capture and coordination cheap and has the capacity to enable a shift from hierarchy to intelligence. But most startups are not constrained by coordination. What they lack is knowledge and know-how. Most are focused on working out what matters, what signals to trust, and how to make the right calls.

From Hierarchy to Intelligence is a convincing argument for removing hierarchies. They are not sacred, just until now the best available tool.

A system that routes information through layers of humans introduces exactly what you would expect: latency, distortion, politics. The context that leaves the engineer’s desk is not the same context that reaches the CEO.

That degradation is not a management failure. It is human and structural.

AI changes this.

A system that maintains a continuously updated model of the business, what is being built, what is blocked, where resources are allocated, what is working, can carry the information hierarchy was designed to translate. Without the lag and without the loss of context.

Most companies are still using AI as a copilot to an unchanged structure, rather than questioning the structure itself.

So the premise should be taken seriously.

The real shift

Information capture and resulting coordination is becoming cheap.

So judgment is is what matters and is expensive - in every sense of the word.

That is the shift.

AI handles routing, alignment, and increasing parts of execution. But it does not own consequences.

It can assist judgment, but cannot be accountable for it.

So advantage moves to the companies that:

• build the best context

• learn the fastest

• and apply judgment at the right time, place and speed

The hidden assumptions

But, the model rests on two conditions most founders do not have…

1. Machine-readable work

Everything creates artefacts. Decisions, discussions, code, plans, progress. This becomes the raw material for a company-level model.

That is not a given. It is a design choice.

Most early-stage companies are still operating in messy reality:

• context in the founders head

• partial documentation

• unstructured conversations

• decisions made, but not captured as data

You cannot build an intelligence layer on top of context that does not exist.

2. Signals you can trust

Most founders do not have clear, immediate, measurable signals.

They are still asking:

• Is this the right problem?

• Are these the right customers?

• Are we delivering sufficient value for the customer not to cancel on us?

This is not just an information problem. It is a discovery problem.

And while AI is improving how discovery happens — generating tests, synthesising feedback, accelerating iteration — it does not remove the need to decide what is true.

So while it dramatically changes the speed of discovery, it does not remove the importance of being right.

Where the model breaks down

1. Discovery is not solved, it is accelerated

The line between information and discovery is blurring.

AI can run more tests, generate more insights, surface patterns faster.

But this creates a new problem.

More signal. Less clarity.

The constraint is no longer access to information.

It is deciding what matters.

2. Companies fail on judgment, not coordination

Startups rarely fail because information moved too slowly.

They fail because they:

• solved the wrong problem

• misread the customer

• scaled something that did not work

AI does not fix this.

It makes it happen faster.

3. Hierarchies bundle more than coordination

Hierarchy did more than move information.

At its best, it bundled:

• context

• decisions and accountability

• coaching

AI replaces the first.

It does not yet replace the others.

Removing structure without replacing these functions does not create intelligence.

It creates inconsistency, at speed.

4. This is domain-dependent

The model works where:

• signal is clear

• feedback is fast

• truth is measurable

For example: payments, marketplaces and high-frequency systems.

It breaks down where:

• Problems are ambiguous

• ICP is unclear

• Value is subjective

• The product is still being discovered

Which, lets be honest, describes most early-stage companies.

The irony is that as signal improves, the model becomes more viable.

But by then, the judgment infrastructure is already built.

What founders should actually do

Not remove structure, or simply add AI.

But focus on how the company learns, makes decisions, places bets, takes intuitive leaps.

1. Use AI to compress time, not replace thinking

Speed up:

• analysis

• iteration

• execution

But be deliberate about:

• direction

• prioritisation

• which signals to trust

AI accelerates your judgment.

Your challenge is to make sure it is worth accelerating.

2. Build learning loops, not just systems

The goal is not more output.

It is better decisions over time.

Ask:

• What did we learn from this cycle?

• Where was our model of the customer wrong?

• Where did speed help, and where did it amplify a mistake?

The companies that win will not just move faster.

They will learn faster.

3. Design explicitly for judgment

Before restructuring, be clear:

• Who makes the call when the model is wrong?

• What context do they have that the system does not?

• How do we detect drift early?

Judgment does not emerge from architecture.

It is built through people and their experience, good and bad.

4. Stay close to the edge

AI will optimise for known problems and existing intent.

It will not discover what customers cannot yet articulate.

That remains a human task.

The founders who stay closest to that edge will have better inputs than those who run the company through dashboards.

A more useful frame

The question is not whether hierarchy disappears.

It is whether you can replace what it actually did.

A more useful way to think about the AI-native company:

• Signal: what do we trust as truth?

• Learning loops: how we update our model of the world?

• Judgment: who makes decisions under uncertainty?

• Execution: where should AI and humans operate?

Most companies will rebuild around these layers, not simply flatten.

The bottom line

AI removes the cost of coordination.

It does not remove the cost of being wrong.

The companies that win will not be the ones that flatten fastest.

They will be the ones that use infrastructure to: learn faster, build better context and apply judgment

Descarte’s Error

We talk a lot about the science of company building. Scaling. Sales. Metrics. Process.

That all matters.

But it misses something fundamental. A company is not a machine. It is a creative act. A series of beliefs. A process of learning and a system of decisions. And the quality of those decisions is set by the founder’s ability to think clearly under pressure. That’s the job.

Antonio Damasio, studied patients who had lost access to the emotional centres of the brain, and wrote about this extensively in a fascinating book “Descarte’s Error.” These patients could still reason. Analyse. Compare options. What they couldn’t do was decide. They had no instinct. No judgment. No ability to choose between two perfectly logical paths. Damasio’s conclusion was simple. Reason and emotion are not separate systems. They are one.

Since René Descartes, Western thinking has tended to treat reason as rational, logical, and objective, and emotion as irrational, disruptive, and something to control. Antonio Damasio argues this is the mistake. That is the error.

So when founders talk about intuition, about gut feel, about the creative leap that no framework could have predicted, that isn’t soft. That is their unique and personal system working.

And when that system degrades, decision quality goes with it.

The AI era has compressed the timeline of company building.

What used to take years now happens in months.

Idea to product in days. Product to revenue in weeks. Revenue and the associated pressure arriving almost immediately. VCs breathing down your neck for faster growth.

It feels like progress and may be progress. But it creates a new constraint.

Cognitive load.

Not just hours worked. Not just effort. The sheer volume of decisions. The speed. The ambiguity. The constant context switching. The sense that everything might matter.

At some point, the system overloads.

And when it does, so too does our ability to make good decisions. Some just spin their wheels. That’s my modus operandi. Many just keep going.

This is where most founders get it wrong.

They think failure looks like collapse. It doesn’t. Products keep shipping. The team stays busy. The metrics keep moving. Everything looks fine. While underneath, something has changed.

Decisions get revisited. Priorities drift. Teams interpret things differently. Conviction drops, even as activity rises.

The danger isn’t paralysis. It’s the illusion of progress. Velocity without judgment.

That’s where companies drift.

There are clear moments where this risk is at its highest…

When we transition from one phase to the next.

Phase transitions play out in many ways, but let’s keep it simple - one of the earliest and the hardest - is the transition from chaos to order. From the founder doing everything themselves to building a system that works without them constantly in the loop. From instinct-led execution to something that needs to repeat, repeat, repeat.

This has always been hard.

What has changed is when it arrives.

You now hit this moment earlier than you expect. Before your instincts are fully formed. Before you have made enough mistakes to trust your own judgment. Before the company is ready for the systems it suddenly needs.

At exactly that moment, everything accelerates. More decisions. More people. More complexity. Less time to think.

This is the cusp. And it demands your best thinking at the exact moment your capacity to think is most compromised.

If Damasio is right, this isn’t just operational, it is neurological.

The system you rely on to make decisions degrades under sustained cognitive load. Not dramatically. Quietly. You don’t notice it happening. You just start making slightly worse decisions. Then slightly worse again. And because execution is faster than ever, those decisions compound before you have time to correct them.

The cusp doesn’t announce itself

It looks like momentum. More customers. More hires. More opportunities.

But underneath, the signals are there.

You are involved in too many decisions. The same questions keep coming back. Alignment takes more effort than it should. You feel busy, but less clear.

Not failure.

Drift.

The founders who get through this know it will come and they change how they operate as it hits.

The first shift is to reduce the surface area of decisions.

You cannot scale decision volume linearly with company growth. Most companies try anyway. They add more products, more markets, more initiatives. All running in parallel. That is where complexity takes over.

Frank Slootman describes this shift as moving from parallel priorities to sequential execution.

Not everything at once. One thing at a time. Done properly. Sequential execution of what matters most creates real momentum.

Frank Slootman explains this challenge he put to his leaders: “If you only achieved one thing this quarter or this year, what would it be?

But this only works if you separate three things that are often blurred together:

1. The vision must be stable. It rarely changes. This is true of the very best companies, even decades later.

2. The strategy can move. Year to year, sometimes faster.

3. What matters right now can change very fast: quarter to quarter, perhaps even month to month.

This framing shows up consistently in the work of The Table Group. The confusion isn’t usually about any one of these. It’s that they get blurred together.

The Table Group Clarity Framework is deceptively simple, yet powerful. Two questions are critical:

• What is most important right now?

• And, who does what?

When everything starts to feel important. New priorities get added without old ones being removed. The system fills up.

At that point, it doesn’t feel like a loss of control. It feels like progress. More happening. More motion. More output.

But underneath, decision quality is degrading.

So your job becomes very simple. And very hard.

Be absolutely clear on what is most important right now and who does what.

Be equally clear on what does not.

And be willing to stop things that were previously important.

That is where most companies fail. Not in starting things, but in stopping them.

Scale does not come from doing more things.

It comes from doing fewer things, in sequence, at a higher standard, with absolute clarity on what matters now.

The second shift is recognising that your job is no longer to make more decisions.

It is to make fewer, better ones.

That requires space. Time to think. Fewer inputs, not more.

If your calendar is full, your judgment will be compromised. There is no way around that.

The third shift is alignment.

Misalignment compounds faster than poor execution.

If people are pulling in slightly different directions, the system generates noise and noise kills judgment.

You have to be explicit. What matters now. What does not.

And then repeat it until it feels unnecessary.

Which of course it isn’t.

If you are on the cusp, or approaching it, the signals are usually there.

Decisions get revisited. Different parts of the company describe priorities differently. Output is high but conviction is low. You feel increasingly reactive.

That is not scale.

That is drift starting.

I have seen this often enough to know it is coming. I have not found a way to stop it.

The only consistent advantage I have seen is this: the founders who make it through are the ones who knew this moment existed. They did not know when it would arrive. But they recognised that it would.

Whether that is enough, I genuinely do not know.

What to do now

If this resonates, don’t overthink it. Start here.

1. Reduce what’s active.

Look at everything currently in motion across the business. Products, markets, initiatives, hires. Then ask one question: What actually matters right now?

Not this quarter in theory. Not strategically. Right now. Pick one or two things. Stop the rest.

2. Create space to think.

If your calendar is full, your judgment is compromised. Block time. Protect it. No inputs. No meetings. Just thinking. This is not a luxury. It is the job.

3. Make decisions once.

If the same decisions keep coming back, you don’t have alignment. Be explicit. What matters and what doesn’t. Then repeat it until it feels unnecessary.

Why because what breaks companies is a loss of judgement.

Lastly, watch for drift. Not failure. Drift.

Decisions being revisited. Priorities interpreted differently. High activity, low conviction. If you see it, don’t push harder.

Slow down. Reduce. Create clarity.

Scale and speed won’t break your company.

But a loss of judgment will.

Stephen Millard is an Operating Partner at Notion Capital, Europe’s leading early stage investor.

In a recurring revenue model, growth is not driven by acquisition alone but by the system that follows: how quickly customers activate, how reliably they reach outcomes, and how consistently that success turns into expansion and advocacy. This is not a function. It is an operating system. And it determines whether growth compounds or unravels.

I like winning, but I hate losing

When I was growing up, I was obsessed with playing rugby. It was the one thing that really defined me. I wasn’t particularly academic, artistic or creative. I had no idea what I wanted to do with my life. But I was good at this. I’d be playing competitive rugby once, twice, sometimes three times a week - school, club, county…

I really liked it when we won. But far more than anything, I hated it when we lost.

I remember my mother buying me a mug. A Snoopy mug, with a forlorn Charlie Brown on with his baseball bat slung over his shoulder. And it said:

“Winning may not be everything. But losing isn’t anything.”

I loved that. I still do. Because it captures something profoundly important for any recurring revenue business - SaaS, AI-native, whatever you’re building.

Winning customers matters enormously. Of course it does. Without new customers there is no future. But losing them? Just like me at school, and Charlie Brown, that’s just the worst. And yet, as an industry, we remain obsessed with acquisition and almost entirely blind to what comes after

An Obsession That Leads Us Astray

There is an understandable addiction to customer acquisition. To the win. To the new logo. It shows up in the metrics we celebrate, the board updates we write, the headlines we chase. New bookings. New customers. New markets.

It is understandable. It is rational. And it is only about 5% of the job.

My good friend Dan Steinman (who built the customer success category at Gainsight) said it better than I ever could: winning customers is 5% of the job. Making them incredibly successful, turning them into advocates, ensuring they see demonstrable value from the problem you’re solving, that is the other 95%.

The balance is out of whack.

We’re not giving anywhere near enough thought to how we are going to deliver the repeatable experience that ensures customers get the value they were promised. This commitment to value needs to run through the whole selling / buying process, through onboarding, into the daily rhythm of use.

An Old Truth That Hasn’t Changed

I used to work at Oracle. Larry Ellison was famous for saying: “If you don’t build the product or sell the product, tell me very carefully what you do.” (And as I worked in marketing, I’d be found hiding in the toilets. It was a sales culture through and through!).

It was a worldview built for a different era. Build it. Sell it. That was the game.

The game has changed. The new operating principle, the one that has to define every company in this era, is this:

Build to solve the problem. Sell to communicate the value. Deliver to compound value.

That is not a slogan, rather it is a structural commitment. It is woven into every stage, into how you build the product, into the promises made during the sales process, into how you onboard, activate, and expand. Every part of the machine has to be oriented around it.

Businesses exist to create value, shareholder, employee and most critical customer value.

Why This Matters More Now, Not Less

AI-native businesses move incredibly fast. They create applications and products that were previously impossible and do so incredibly fast. They solve highly complex problems with lean teams and extraordinary leverage.

But here is the risk: speed of customer acquisition without depth of value delivery is a trap. It feels like momentum. And then the renewals, or lack thereof, come around.

The three most common failure modes I see across portfolio companies are unchanged by AI: ability to sell, ability to onboard, and ability to prove value. Two of those three happen after the deal is signed. The technology changes. The failure modes do not.

AI makes it easier than ever to generate curiosity. It makes it easier to fake value. A customer who has been wowed by a demo but cannot point to a tangible outcome six months later is not a customer, they are a churn risk with a contract attached.

The Philosophical Shift

What is required here is not a process change. It is a belief change.

It starts with how you think about the company itself.

The best founders I know are not obsessed with their product, they are obsessed with their problem. Products are finite. Problems are infinite.

Build your identity around the problem and you can reinvent the product as many times as necessary. Never losing sight of the problem you solve, the customers you choose to serve and the value you exist to deliver.

The same logic applies to customers. Your relationship with a customer does not end at signature. In a recurring revenue business, signature is the beginning. The entire commercial relationship that follows - expansion, renewal, advocacy - depends on what happens next. And what happens next depends entirely on whether you have designed a machine that systematically, reliably delivers the value you promised.

NRR (net revenue retention) is the metric that cannot be faked. Above 100% means your existing customers are growing their spend. It means they are getting value. It means the 95% is working. Below 100% means no amount of new logo growth will save you. You are filling a leaking bucket.

So What Do You Actually Do?

Philosophy shifts the priority. But infrastructure is what acts on it.

The question is: do you have a systematic way of capturing the signals that tell you whether your customers are actually getting value? Not whether they are using the product. Whether the problem is being solved.

That requires four things:

1. Define the outcomes, not the activities. Before a customer goes live, be explicit - with them and with yourself - about what success looks like. Not logins. Not seats used. Actual outcomes. What will be measurably different in their world if your product is working?

2. Build milestone markers into the journey. Map the customer journey from signature to first value moment. Time to first meaningful use. Time to first demonstrable outcome. Time to expansion conversation. These are leading indicators that tell you early whether a customer is on track or at risk.

3. Capture and synthesise the data. This needs to be systematic, not heroic. Customer success cannot run on the intuition of a good CSM. You need the signals (product usage, engagement patterns, outcome data) synthesised in a way that allows you to act proactively rather than reactively.

4. Create the repeatable experience. The best customer outcomes are not accidents. They are the result of a deliberate, repeatable playbook executed consistently across every customer, every segment, every market. Activation playbooks. Onboarding frameworks. Health scoring. Escalation protocols. None of this is glamorous. All of it is essential.

   ---

The Compounding Effect

Here is what makes this all worth it.

Customers who achieve value stay. Customers who stay expand. Customers who do both become advocates.

Advocates - customers and users - become your best salespeople: referencing you, referring you, reducing your cost of acquisition and increasing your close rate on net new business.

This is the flywheel that enduring companies are built on. Not the new logo count. The compounding value created for customers who love you.

Power — real competitive power — does not come from having the best product in a market where every product can be replicated in weeks. It comes from having customers who would not leave even if they could. That is what we should be building toward.

So yes — win. Win hard and fast. The obsession with customer acquisition is not wrong. It is just incomplete.

Losing isn’t anything. But the 95% between winning and them never wanting to leave — that is everything.

—

Stephen Millard is Chief Platform Officer and Partner at Notion Capital.

First principles thinking is the same discipline applied deliberately. Strip away inherited logic and reduce a problem to its components. Rebuild from what is true, not what is familiar or oft-repeated.

In Why Information Grows, César A. Hidalgo argues that information is physical order created in a universe that naturally drifts toward disorder. He describes this as entropy. Economies grow not because they produce more stuff, but because they accumulate, organise and embed knowledge into increasingly complex forms that function.

Order does not emerge automatically. It requires structure, constraint and coordination. Entropy has its own momentum. Order does not, so we have to work hard, constantly, to create it.

Startups operate under the same conditions.

Without discipline, ideas multiply, priorities blur, and everything begins to feel equally urgent.

First principles thinking is the mechanism founders use to impose order on what I like to call “constructive chaos.”

What needs to be true?

When I start working with the founders of a new business, post our first investment, my first question (after diving into what the business does, the problem it solves, the vision, the traction etc) is:

“What needs to be true to achieve the stated goals?”

Then:

“What do you need to learn?”

Then:

“What do you need to prove?”

This is an exercise in focusing attention on the topics that truly matter. Too much of what is discussed in venture capital are lag indicators: growth, customers, revenue, expansion. These are outcomes. What really matters are the lessons that need to be learned, the experiments that need to run, and the hypotheses that need to be proved (or disproved).

Early stage companies are not scaling machines. They are learning systems. Their progress depends on discovering what is true about the problem, the customer, and the value they can create.

To push this further I often conduct a premortem: Imagine it’s 24 months from now and your assumptions were all wrong. You are way off track. Why? What went wrong? I’ve written about this many times before.

Let’s take this back to first principles…

Start with (and stay with) the problem, not the product

My good friend Chris Tottman is fond of saying: “Good founders obsess with the product. Great founders obsess with the problem.”

Who cares about this problem and why?

Why has it not yet been solved?

What has changed that now makes it possible to solve this previously unsolved and incredibly valuable problem?

There is a useful loop I return to that helps isolate the essence of any business:

1. Is this the biggest, baddest problem you can solve?

2. Are these the best customers to solve it for?

3. Is this the most value you can create?

Round and round we go. In my opinion, the early stages of a business are all about this.

Ignore the noise

Startup ecosystems generate enormous noise.

Raise more.
Hire more.
Go faster.
Relocate to the US.
Go product-led.

Sometimes these things are correct. Often they are not.

First principles thinking doesn’t mean rejecting all advice.

It means understanding why something works or doesn’t.

The more possibilities that exist, the more important judgment becomes. This is the true value of a human being in an AI-native business.

Show me the money

Businesses exist to create value. Value for customers, shareholders and founders. This has to start with the customer:

• If a customer buys, what makes them stay?

• Why would they spend more over time?

• How do they receive more value the longer they work with us?

If customers don’t stay and don’t grow, your business won’t compound.

Be honest about what’s really happening

This is all about intellectual honesty. Remember the maxim: the first principle is to not fool yourself, and you’re the easiest person to fool.

• Why are people buying from you?

• How are they using your solution?

• What observable behaviour tells you a customer has truly adopted your product rather than just experimenting with it?

Be clear about how you will create compounding value (for your customers and you).

Ask instead, what do we need to be world-class at?

What gives us leverage?

AI increases leverage, but it also increases entropy.

AI dramatically increases what small teams can build. But it also increases the number of things that could be built.

The easier it becomes to create products, features and experiments, the harder it becomes to decide which ones actually matter.

In other words, AI increases leverage, but it also increases entropy.

When constraints disappear, discipline matters more. Founders must decide which problems are worth solving, which signals truly matter and therefore what to build.

I was listening to a podcast with Ben Peters, CEO of Cogna, and what he said struck me: “I have a compulsion to not just build a startup, but to build this startup, to solve this problem. And to endlessly experiment to prove what works.”

Being a first principles thinker as a venture-backed founder means build your company based on what is actually true, not on what other startups look like. You build it by understanding what you need to learn, by what you need to prove and by constantly experimenting, while obsessing with the problem you solve, the customers you serve and the value you deliver.

In the end, a startup is simply a machine for discovering truth.

----------------------------

Read: Why information grows, by Cesar Hidalgo

Listen: AI is rebuilding the economy with Harriet Green and Ben Peters.

It is way too easy to be caught up in our own hype and statistically unlikely outcomes. As VCs and Founders we believe in the possibility of highly improbable outcomes, of course we do. We need that confidence and conviction to take the leap into the unknown. But without disciplined reflection, that same conviction can easily become self deception. And that is where avoidable mistakes begin.

I talk about principles not playbooks and and often return to the wisdom of physicist Richard Feynman, quoted in the title, that hits so hard, “the first principle in life is that you must not fool yourself - and you are the easiest person to fool.” I think we can all see the wisdom in that. What he is describing is intellectual honesty.

Its too easy to believe our own hypotheses, too easy to selectively interpret data, too easy to ignore contrary evidence. And for founders AI means you can raise more, build and ship products faster, demo better. Which all means you can fool yourself faster than ever.

Working forwards - laying out the path

In order to have something meaningful to work on in the premortem phase we need the forward looking plan. In essence, what needs to be true for that gangbuster Series A, B, C? Achieving X revenue, with Y margin by Z time.

I test this with three questions:

1. What do you need to learn? For example, how customers buy, how you will sell, how you will deliver value.

2. What do you need to prove? That you can nail an ICP, that you can achieve repeatability, that you can nail expansion?

3. What capabilities do you need to build? A tight loop of feedback between customers, sales and product? The ability to move fast while dealing with burgeoning complexity?

These questions force specificity, starting us down a more open-minded path.

Working backwards - embracing the premortem.

Now imagine it’s all gone wrong.

1. You said you’d have $10m ARR. You have $1m.

2. You said you’d have 120% NRR, but instead you are churning customers.

3. You said customers would receive incredible value. Maybe they are but you can’t prove it.

That stellar Series A, B, C is a distant memory, instead you are fast approaching the end of the runway. A major down round or shuttering the business are now the only options.

Why, what went wrong?

This sometimes takes some coaxing. I’ve had founders say to me, “the only risk I’m worried about is not taking the whole market.” I get it, but humour me. What went wrong? What mistakes did you make? What did you miss entirely? What assumptions are you basing your theory on that could be flawed? Play the game and imagine your business is dead in the water. Why?

Importantly don’t think about this in context of the unavoidables or uncontrollables - such as a competitor raising more money or the market conditions changed. Focus on what’s in your control.

Again, what, exactly, broke?

• We failed because our ICP was too broad. Or we failed because we went too narrow too soon.

• We failed because of a total lack of focus on repeatability.

• We failed because the sales cycles were too long.

• We failed because the time to value was too long and the value was inconsequential.

• We failed because we were misaligned as a team.

• We failed because we hired too many people with the wrong skills - too senior, too junior, no hustle.

• We failed because I as a founder did not adapt.

Over the last five companies I have worked - all AI native - three themes come up repeatedly: the time to sell; the time to onboard; and the ability to prove value. No different from the conversations I have been having for years. Selling. Onboarding. Proving value.

Name the risks and put them into sharp context. If expansion is your ultimate signal and your ideal customers are not expanding, your premortem should start there.

Then make it practical. Name the white hot risks. The ones that can kill the company.

For each one:

• Estimate likelihood and impact.

• Decide the trade-offs.

• Agree mitigations with clear owners.

• Define leading indicators that tell you early if it is happening.

• Share it with the team, this is nothing to be ashamed of.

This is not about being pessimistic. The goal is to avoid wishful thinking, not to lose belief.

A final thought.

VC-backed startups are, by definition, attempting something that has not been done before. That has always been true, but in the AI era where the possibilities and advances seem endless and speed is unprecedented, the risk of fooling ourselves is unprecedented.

So never forget the first principle. Don’t fool yourself.

We talk about the importance of data, integrations, strategic partnerships as core elements of our strategy and how they grow our defenceability.

My problem with moats and defenceability is that they are, well, defensive. They are about not losing. Think about it in terms of a castle with huge walls, turrets, archers, boiling oil and of course a massive body of water surrounding us. Sure that makes us harder to beat. But it doesn’t stop our enemies doing their damndest; beating down the walls or starving us into submission.

Taking it back to the current times, perhaps they make it harder for a customer to swap us out, but is that the message we want to convey.

Power is something different altogether.

If I’m powerful, sure I have my defences, but I’m confident I will not lose the castle, because my enemies won’t even think of attacking. If I’m powerful I can expand my territories, seize more natural resources and build stronger allies. And become more powerful.

Power is what matters in this age. There is almost no differential value in product, AI is eroding that. The strategic decisions we make to build power is what will sustain companies, ensure they continue to raise the best capital, hire the best employees and, most importantly, win the best customers while doing so at premium high-margin prices.

I’m going to be digging more into this topic as a key guiding principle in the coming months, interviewing founders and executives, bringing more and more examples to strenghen my case, and build my power in defence of power.

In this article (somewhat left-field) I’m focusing on the role of human authenticity.

TLDR:

Philip K. Dick asked what counts as real in a world of replicas. Founders face the same question in the AI Age: when machines can mimic everything - ideas, design, words - how do we create something truly human?

• Authenticity: Like Philip K. Dick’s question of what is “real” in a world of replicas, founders in any era must build companies that create genuine emotional impact, not just left-brain logic.

• Humans are the spark: AI can optimise but it is backward-looking and does not (cannot?) provide the original creative leap, counterintuitive insight, or instinct that define breakthrough products.

• Storytelling: Logic doesn’t move people, stories do. Founders must make their vision desirable and memorable, not just rational. We all know it; we make big decisions emotionally and justify them rationally.

• It’s personal: Intuition, judgment, empathy, and imagination, qualities AI cannot replace, are essential in building enduring companies.

• Magic > mechanics: In an age of infinite optimisation, what separates exceptional founders isn’t just speed but the scarce, irreducibly human essence of creativity, storytelling, and belief.

What counts as real in an age of AI-replicas

Philip K Dick was a brilliant and prolific author, with an incredible ability to imagine futures that must have seemed so distant when written and yet now seem so close.

If you have watched The Man in The High Castle (but not read the book), or watched Bladerunner, but not read “Do Androids Dream of Electric Sheep”, the book the film was based on, then I’d encourage you to do so.

Neither much reflect the TV series or films they inspired, but the glimpses of a future they offer is very evocative. There’s a wonderful piece in ‘The Man in the High Castle’, towards the end of the book, that I hope I’m not ‘misrembering’ where one of the characters, Juliana I think, walking down a street in San Francisco, sees glimpses of a parallel world in which the Axis of Evil did not win. She sees it briefly and then confused walks on.

Do Androids Dream of Electric Sheep is similarly thought provoking. Rick Eckard, the bounty hunter, dreams of owning a real sheep, but has to settle for an electric one that he cares for just as passionately.

The question “Do androids dream of electric sheep?” is one of authenticity; in a world of convincing replicas, what counts as real? Do androids dream, do they feel, are they less than human? As a bounty hunter Rick is tasked with tracking down and ‘retiring andys’, employing a test that measures empathy. Answering no to these questions has allowed Rick to kill them without compunction, though his conviction fades as the story unfolds.

It seems to me we’re facing the same dilemma in the AI era.

So much we read is machine-generated pap for machines to parse and read (please don’t get my colleague Kamil Mieczakowski started on this!).

AI tools can simulate intelligence, empathy, and even creativity. For founders, the challenge is to build platforms that create authentic value for their customers, on a rational and emotional level, and that build deep levels of defensibility and market power to stand the test of time.

They can only do this by embracing the fact that company-building is not just a logical, mathematical endeavour. It’s deeply human:

1. From the point of creation, when the founder dreams of a better world (perhaps not a sheep on the roof!),

2. to the intuition and gut instinct that guides their decisions,

3. to their ability to make big counterintuitive bets,

4. to the story they tell that binds the company together’

5. to their respect for their customers and, of course,

6. to their care for their colleagues.

Everything left brain is necessary but insufficient, as is everything right brain, it’s the combination that works. We’re emotional beings. We decide for emotional reasons, we reinforce with rationality. This is proven and evidence-backed. The book, Descartes Error by Anonio Damasio is a great primer point on this. The premise is that emotion and reason are not separate or opposing forces, but deeply intertwined and interdependent. Damasio would argue that feelings are just as important as logic in building an AI-native company. One hot take for me, from reading the book a few years ago, is that decision making is a ‘whole body and brain’ process.

There is no doubt in my mind that AI can enhance creative thinking, when used as a collaborative tool. When writing these articles I readily turn to chatgpt to challenge my thinking and broaden my perspective, but the words written here are my own and deeply felt.

We have to take real care with assumptions, context and data quality, but using AI to challenge those assumptions, to play devil’s advocate, to interview us, challenge us and provoke us to think more deeply is incredibly powerful. And, full disclosure, something I do when writing these articles.

Human intuition

There is so much I could write on the topic of intuition or gut instinct; every founder I have spoken to has mentioned, often indirectly, the right and left brain, the rational and emotional aspects of company building and directly mentioned the role of intuition and instinct.

In 2024, for Notion Capital’s Cloud Champions, I interviewed more than 20 European CEOs and Founders, each with more than $100m revenue and the idea of instinct as a founder superpower came out repeatedly.

This quote from Avi Meir, CEO of TravelPerk summed it up best.

If I had to offer advice based on my experience, it would be to trust your gut instincts. As a founder, you’ll often have moments where you feel strongly about a decision, even if it doesn’t seem entirely rational or if others advise against it. While it’s important to back up these feelings with data, there’s something to be said for following your intuition. I’ve learned to listen to that inner voice, recognizing that it’s an unconscious processing tool based on past experiences, education, and possibly even genetics.”

The most successful leaders understand this deeply. Intuition is the learned insight gleaned from doing the work over years and years and years. It’s something to cherish and to hone. It allows to make fast decisions, built on slow learning.

It goes beyond founders of course, Matt Welle, CEO of Mews prioritises systems builders and first principles thinkers across his business:

“We are actively identifying and nurturing people with natural learning skills and adaptation capabilities. [While embracing AI], we must not lose sight of the need for instinct and insight.

AI doesn’t understand the depth of a customer problem in the way a founder or a product leader does, developing often counter-intuitive solutions. Yes, AI can spin up an application, but does it know how to solve that problem in a way that delivers rational and emotional value for customers and deepens defensibility? The founder and their teams understand the problem they are looking to solve and without their input AI cannot build the right product.

Matt continues: The answer must not be just making people more efficient. Sure discovery is faster, but the validation and the intuition still has to happen with humans. There’s no replacing that.”

Ciaran O’Mara, co-founder of Protex.AI emphasises the importance of “that gut feeling that tells you ‘I don’t think we should do it this way’.”

His conclusion is clear: “We don’t build products our customers ask us for. We’re in the business of doing something nobody’s ever done before. That needs deep thinking, it needs creativity, it needs collaboration, it needs lots of things to connect together... AI can help along the way, but it’s not going to drive it.”

What’s possible and what’s valuable sits at the heart of every AI-native startup.

What truly separates the exceptional founders isn’t just their ability to move fast or optimise execution or the intuition. It’s the creative leap. It’s the ability to imagine and deliver products no one asked for, but everyone ends up needing.

Nick Cheetham, former CRO at Currencycloud and now advising multiple SaaS and Native-AI companies captures this perfectly: “If you’re doing something that’s never been done before, then where does that come from? It comes from human creativity and the instinct to be able to take a counterintuitive leap. The instinct to make a decision that flies in the face of the data and the information. I know the data says go this way, but we’re going to go that way.”

Don’t miss the magic

It’s easy to fall into the trap of thinking that company-building in the AI Era is a mechanical process. But that misses the magic.

Richard Valtr, Founder of Mews, puts it well: “The original IP, that’s the art, that’s the alchemy. AI can replicate execution and remix existing ideas, but it cannot generate the spark. In an age of infinite optimisation, the scarce resources are creativity, storytelling, and judgment.”

Jeroen van Hautte, Co-Founder of Techwolf offers a crucial insight:

“There are two ways you can use AI. You can use AI to avoid thinking, and you can use it to enhance thinking.” He continues: “We need to expand our thinking while leaving room for creativity in terms of the counter intuitive leap that says, ‘you know, we’re going to do something that the AI is not telling us to do’.”

Storytelling and controlling the narrative.

Telling stories in a way that connects at a human and emotional level and not simply objective and rational is critical. Valtr continues: “I’ve got ten seconds with a customer, or a VC, or an employee. If I can’t make it obvious what we’re doing and why it matters, it doesn’t matter how good the tech is.”

My premise is that our industry is over-reliant on rational, product-led marketing and under-invested in creativity, narrative and right-brain storytelling. Storytelling is the most natural way for humans to communicate, developed over hundreds of thousands of years. Dave Peterson, former CMO of Celonis and Author of Playbigger explains: “storytelling is the essential foundation of all business communication and in this AI-era of endless noise, ever more important.”

Take Homer’s The Odyssey as an example of the power of storytelling.. It was formed in an era of spoken storytelling, passed down for generations, and was only put to paper approximately 400 years later. Imagine the power of stories that can last through generations! Something I often talk to founders about is the importance of their customers (and users) telling stories about their company or product. Do you know what they say, I ask. Imagine if we gave our customers the stories to describe the transcendental impact of our solutions.

We’ve built entire go-to-market engines around rational choice and behaviour, but humans don’t work like that. Emotional triggers, stories, and identity drive action far more than features or benefits.

We created a logical, product-led, left-brain system and assumed that’s how people decide. But people don’t behave like spreadsheets, even if AI does. Storytelling, identity, and behavioural cues are what actually move them. Narrative isn’t the icing on the cake, it’s the delivery mechanism for your value.

That’s why right-brain thinking, story, aesthetics, emotional logic, is essential. Not as a substitute for data or systems, but as an amplifier, a megahorn. It makes what you’ve built desirable.

AI can do many things. But it can’t believe. It can’t dream of electric sheep. It can’t create something from nothing. That still belongs to the founder and the exceptional people on the journey with them.

Don’t be a replica!

Building in the AI era demands more than mastery of the tech; it demands mastery of what makes us human.

AI can simulate intelligence and pseudo-empathy, and even creativity, but it cannot originate. It cannot dream up an idea, believe in the possibility of an unlikely outcome, or care deeply for your colleagues or customers.

The very best AI-founders, those that go on to build legendary companies will bridge that gap; combining left-brain logic and science with right-brain imagination, intuition, and storytelling to create something truly authentic and transformational.

The most exceptional leaders understand this. Like Ciaran, Jeroen, Matt and Richard, they use AI to enhance their thinking, not avoid it. They learn slowly, but make fast, intuitive judgment. They trust the gut that’s been earned through experience. They build companies that connect on an emotional as well as rational level. And they tell great stories.

AI can replicate execution, but not the creative leap. It can scale belief, but it cannot create it.

That spark, human, imaginative, irrational, and deeply authentic, remains the true source of power.