Family offices tend to approach AI with three hats on. The three areas do often overlap, but each one brings a different focus.

The first is the running of the office itself: day-to-day operations like other small businesses, how it works, and what it costs to operate. The second is the business or businesses the family owns or controls, where AI shifts the cost structure, the headcount and the competitive position, and where the consequences land directly on the family. The third is the investment portfolio.

Each of these needs a distinct approach. The third is the one to look at today, because over the last few weeks that hat has been busy.

The venture unit of Bernard Arnault’s family office took part in a Series A round for Humanoid, a London-based developer of industrial robots. Bezos’ family office increased its investment in Generalist AI, a software company focused on robotics. And Azim Premji’s family office is reported to be negotiating the lead position in a new funding round for a developer of general-purpose robots with offices in Paris and Silicon Valley. The combined net worth of those three people is about $628.5 billion, and none of those fortunes was made with AI.

It’s showing up everywhere. Luxury goods, retail logistics and Indian software services are buying robots, even though they made their money somewhere completely different.

Bodies, brains and blood draws

The coverage has treated this as a single trade. It is however at least three very different types of investments, and each carries a different risk profile.

The first level is the humanoid, or the general-purpose platform. Humanoid in London belongs here, as does the company Premji’s office is circling. These are bets on a body that can do anything, which is a massive prize to win and a very long way from generating revenue. Barclays puts the humanoid market at $2 to $3 billion today, reaching $200 billion by 2035, a figure that illustrates the enthusiasm around this trend and says very little about the cash flow.

The second level is the software that powers the machines. The Bezos investment in Generalist AI fits here. It is the more tenable position, because a software layer can charge a royalty regardless of whose body ends up winning.

It is the picks-and-shovels model, except the shovels write policies and the picks run training loops.

The third level is vertical automation with a customer already in place. Dan Friedkin’s family office went into RobCo’s Series C alongside a money manager for Italy’s Agnelli family. Alpha Square Group invested last month in Atoms, the industrial robotics company founded by Travis Kalanick. And Harald McPike’s family office backed the Series B for Vitestro, a Dutch company building autonomous blood-drawing systems.

Blood draws are a task. A hospital performs thousands of them a day, using staff it struggles to recruit, at a cost per draw it already knows to the penny. You can build a model around that. You can estimate the hours of labour saved and the throughput gained, and within eighteen months you will know whether you were right.

You only get that at the third level. Worth being sure which level you are actually at, rather than assuming you have diversified your physical AI exposure.

Shopping on sale

Those cheques went out at an interesting time, while the share prices of companies along the AI supply chain were being marked down over concerns about spending. Investors more patient than the rest of the market bought into positions others were leaving, which is the family office structural advantage working in practice.

The demand is widespread. UBS surveyed 307 family offices across more than 30 markets, overseeing $1.3 billion on average for families worth $2.7 billion, and found AI to be the leading investment theme. Some 65% have invested somewhere across the AI value chain, and offices in the US, Asia and Latin America say they intend to allocate more over the coming year. James Whittaker of Deutsche Bank, who runs wealth management for the UK and Nordics, says robotics is coming up frequently in client conversations.

Few investors can hold a position through a decade of execution risk, and family offices are among them. That is only an advantage, though, if you can afford to be wrong for five years and hold anyway, which is a different question from how much you can afford to put in.

Talk to the operator, not the founder

Family offices can do something funds cannot, and it has almost nothing to do with fees.

In that same UBS survey, over three-quarters of owners were still active in their own businesses, mostly in real estate, consumer goods and financial services. Which means warehouses, production lines, a hotel floor, a clinic, a delivery fleet, somewhere a machine either survives reality or does not.

Most people investing in this area are reading somebody else’s pilot study. You can run your own. Put the robot in your own building, on the night shift, working alongside your own staff, and you will learn more in a quarter than a room full of data will tell you in a year.

Access to deployment is due diligence, and almost nobody uses it that way.

So, three things to do with the next one. Get the name of a live deployment, and speak to the operator rather than the founder. Ask what broke at the third site, because the first site always works. And favour the ventures where somebody is already paying for the labour in question.

One more point, which takes us back to the first hat. Being able to deploy this technology across the group’s operations is a claim about competence at home. If your own office cannot say which model performs which task, or what the reporting pipeline costs to run each month, then the deployment advantage is a hypothesis rather than an edge.

Jensen Huang, promoting Nvidia’s own robotics work, put it about as plainly as it can be put: physical AI is here.

About the author

Francois Botha helps design and incubate the family offices of the future through Simple, which pairs high-touch advisory with technology-led solutions to help private wealth owners build offices that last. From formation and strategy through to AI-powered operations, Simple’s team and network of independent advisors provide the knowledge, agentic tools, and governance to move forward with confidence, while keeping the human judgment that matters most at the centre.

He wrote a weekly column on family office strategy for Forbes and speaks regularly at private wealth, family office and technology events.

A barrel of crude went for about $81 last week. On CNBC twelve days ago, Chamath Palihapitiya priced a second kind of barrel: a million tokens, which he calls a barrel of intelligence. On his tally, a good model from Anthropic or OpenAI runs roughly $26 a barrel and the newest one around $56. Elon will sell you a barrel for a dollar. Zuck is heading for a dollar fifty, Google is at a dollar, and the Chinese labs will do it for fifty cents.

The math holds against the current published pricing from these firms. Claude Opus 5 sits at $5 per million tokens in and $25 out, Fable 5 at $10 and $50, GPT-5.5 at $5 and $30. Same input, a spread of roughly a hundred times, depending which pump you happened to pull up to.

And when used correctly, nobody buying should be able to tell you which barrel their office is burning on which job.

The opportunistic position that wasn’t sized

Consider how the same office would handle a commitment of comparable stickiness on the investment side. There would be a memo. A sizing discussion, a view on the exit, a line in the minutes. Somebody would ask what happens if the manager doubles fees or gets acquired.

Here, none of that. Somebody took out a subscription because it demoed well, a workflow formed around it, and within a year the workflow is how the office works. So the office now holds a concentrated, illiquid, unhedged position in a single supplier of a commodity input, and it took that position through the expense account.

The carrying cost is often less visible than the initial pricetag. Since usage is measured, the operational expenses will increase with activity rather than with the value generated, and invoices typically present a single total amount instead of breaking down spending by workflow. If you ask five different people about the monthly cost of running the reporting pipeline, you are likely to receive five puzzled shrugs as no one has established a way to measure it.

Chamath anticipates that a public company might miss earnings expectations by a few cents, which could be traced back to minor expenses that went unnoticed. In contrast, a family office doesn’t have an earnings call to spark such discussions. Instead, it faces a supplier it cannot sever ties with without incurring costs twice, once to dismantle the existing arrangement and again to rebuild it.

Things are opening up faster than before

Three things happened between the 15th and the 24th.

Thinking Machines released Inkling, Mira Murati’s first model, with the full weights published. It is built to be customised rather than to win a leaderboard, and the company earns its money on Tinker, where you fine-tune it on your own material and walk away holding a portable adapter that belongs to you. Bridgewater did precisely that with financial data and ended up with a small model scoring 84.7% on financial reasoning benchmarks, ahead of the leading proprietary alternatives at under a tenth of the cost.

The next day, Moonshot released Kimi K3, a 2.8 trillion parameters open weight model. According to Arena it outperformed the commercial fronteir labs at a fraction of the cost and what’s more, in blind coding tests many developers preferred it to every American flagship.

Then Jensen Huang made the first post of his life on X, and it was not about chips. He shared an open letter arguing that open models strengthen security and enable sovereignty, and he was careful with the framing: the world needs both frontier closed models and frontier open models. Twenty-five signatories became about fifty inside a day.

But pricing moves fast. The barrel you bid on in the spring has been repriced more than once since, and with pricing moving quickly, large firms are expected to release more complicated models alongside less powerful options at cheaper prices. This trend clearly indicates where the market is heading.

Buy the car before choosing the fuel

Here is the sequence, and it actually works the other way around compared to how most offices have gone about it.

Choose the tools first. The workflows, the systems of record, the places where the work actually happens, selected for the job at hand and for where your data is obliged to sit. Then attach intelligence to them as a component rather than as a foundation.

In practice, this comes down to four things:

1. Select tools based on the specific work you need to accomplish, rather than solely on their AI features. Almost every vendor offers similar features, but very few align with your unique workflow.

2. Make sure that each tool will allow you to use a model of your choice, instead of being locked in to the default model it comes with.

3. Keep all your prompts, context, and corrections in external files that you control. This way, your accumulated insights and decisions will remain intact even if you switch suppliers.

4. Map and categorize tasks by complexity: use the less expensive models for standard jobs like reconciliation, extraction, and classification. When there’s a need for power, go for the more costly tools. Many engineers claim that for the majority of tasks, non-frontier models are entirely sufficient. For most offices, the latter usually represents a shorter list than what they’re likely using right now.

And measure. Try to understand the cost per workflow, not per person or org.

Chamath let the useful part slip almost in passing: buy the oil and you still need an engine, and the engine has to make you go faster. Barrels of intelligence are interchangeable by design, just like fuel. But you still need the car to drive, use navigation, cruise control, self driving, climate preferences and more.

Barrels of intelligence get used and reprice. The engine is the part you keep.

One of the big topics in technology circles right now is the one-person billion-dollar company. Sam Altman has talked about a betting pool among tech CEOs on the year when the first company run by a single person, with AI doing the rest, passes a billion dollars in value. A few years ago this was a provocation. Now it is discussed as a matter of time, not of possibility.

The family office industry has been running its own numbers debate in parallel, only ours is about wealth. How many millions does a principal need to justify the existence of their family office? There are plenty of opinions on that, from $500 to $1bn. It was only $100 million a few years ago, but it’s not enough now. And today, people are saying that a family office could be set up even with 50 million. The debate goes on forever and seems to have heated up in recent years. Why? Because the structure of the business has changed. Who drives this conversation is a separate deep dive. But here is the thing: the tech world’s question and ours are, at the bottom, the same question.

Whatever the opinion may be, this is not the right discussion to hold anymore. The wealth threshold was not a law of nature. This was the price of a particular organisational structure, and the price was based on two assumptions. The first is the people. Almost all costs in the budget of a family office go to payroll: people working in a certain role, roles being defined as certain skills bundles and the classic office needed to have the full bundle in-house, ten to twenty people, whether or not they were utilising every single skill. The second assumption was access. A diversified portfolio required institutional ticket sizes in a sufficient number of assets, so it was impossible to have the necessary diversification without a certain scale. The second assumption eroded with time, thanks to ETFs and new investment styles, which don’t require big tickets anymore. Important, but not the key point. It is the people-assumption that eroded most.

Roles become skills

Why does the technology community take the one-person unicorn seriously now? The reason is precise: AI agents deconstruct roles into skills bundles, and with time more and more of these skills can be performed by software under the control of a single person. Role does not disappear. But the assumption that a role must be occupied by a person full-time does.

Family offices are coming to the same understanding from the opposite side. The founder pursuing the one-person unicorn is trying to scale his product to the world without hiring too many people. Family offices were never trying to scale. They had a lot of people because the work required them: portfolios, taxes, compliance, reporting, governance, philanthropy, personal affairs. If technology community is seriously discussing one person orchestrating a billion-dollar company, the claim that one person can orchestrate the operation of substantial private wealth does not look radical anymore.

If I had to do this again, I would hire one person to do this with me and agents and AI for everything else!

The number that changed

A principal of a San Francisco single-family office said it to us quite plainly: “If I had to do this again, I would hire one person to do this with me and agents and AI for everything else”. This sentence is remarkable because of who said it. Not a vendor, not a futurologist, but a founder who built an office the traditional way and understands what can be achieved from a blank sheet.

This structure is quite practical. Human core: the principal, plus one integrator, someone who understands the intent of the principal, can translate it into reality and is fluent with both human experts and AI systems. Around the human core we place a fractional bench: a CIO, a CFO, a compliance officer, each engaging in their tasks a portion of their time rather than occupying a seat in the office. Below it is the agent layer, responsible for reconciliation, document processing, compliance monitoring, research synthesis and reporting, the machinery that used to be the justification for almost the whole organisational structure.

Functionally, this is a family office. It has its purpose, it serves the ultimate client, it uses documented processes and coordinates the family affairs. But it doesn’t have the payroll that was considered as a must-have by the industry. And there is evidence backing this structure. The AI adoption is fastest in small teams of ten or fewer people, especially where individuals have more than one hat to wear—the exact description of a family office. And the family office created now has one additional advantage that cannot be bought in any established family office. Being free from legacy workflows, it can be created with the agent capabilities from day one.

What one person cannot be

The counterargument is very reasonable. One-person family office concentrates the key-person risk to its maximum, and this industry’s true product, trust, continuity and judgment, cannot be delegated to an agent.

But see where the risk lies. In the traditional family office it lies in tenured staff whose departure is the catastrophe that everybody prepares for. In the AI-native family office all processes, context and history are documented in the systems owned by the family. One person is crucial, but replaceable in a way that a CFO of 20 years with all the information in their head is not. Created properly, the office of one person is more immune from institutional risks than the office of 15 built on memory.

And one person is not an endpoint but more a discipline that must be followed. Define the work first, then define what does it and then add people where it is required by the work. Some of these offices will grow to five people, some to ten. The headcount will be an output of the design, not the entrance fee.

Threshold was a symptom

Apply the logic to the end and see that the wealth debate resolves itself. If the office can operate with a principal, one professional and the agent layer, the economics behind the old threshold become irrelevant. A family with 50 million dollars, actively engaged and properly equipped, is, in functional terms, running a family office, and there are far more of them than the industry thinks.

The industry has asked this question for years: How many millions does a family need to justify the existence of a family office? A better question was how many people the job requires. And the answer is clear: one person, well-equipped.

Earlier this month, Alex Karp – CEO of Palantir Technologies – made a splash on CNBC, articulating a message familiar to the company’s enterprise customers: firms pay for worthless tokens and the frontier labs take the most precious asset of their customers – data, workflows and competitive advantage. “Something has gone completely wrong” is how Karp put it. The anchor tried to link this statement to a criticism of OpenAI and Anthropic, but Karp clarified: “No, no, no. This is reporting.”

Something has gone completely wrong.
- Alex Karp

While this is of course self-interest talk, there is a structural issue in Karp’s description which family offices, among all types of institutions, might recognize. This is the same issue they learned to manage for decades in their investment portfolios: the risk of concentration of exposure to a counterparty outside of your control.

Lock-in not priced

Most organizations using AI today make a mistake similar to the one they would never make in relation to their investments. They funnel their most sensitive information, investment memos, family correspondence, deal flow, tax structure, through the API of a single provider, on its terms, within its tooling, without a realistic exit option. The model, the interface, the agent framework and the data pipeline all belong to a single vendor. Any switching would mean starting from scratch.

Karp puts it directly: customers want “control over their compute, their models, their data stack and their alpha.” For family offices, “alpha” is not only investment information. This is the context of the entire family, information that was traditionally kept within the building.

This risk is not a hypothetical scenario. There are three types of risks emerging from ignoring it. The first is pricing power: once you embed your workflow into a particular lab’s ecosystem, you will be a price-taker on each subsequent increase, and frontier pricing is changing quickly in unpredictable ways both upwards and downwards. The second risk is continuity: models become deprecated, terms change and capabilities you’ve built on can be withdrawn or re-packaged. And third, the risk of exposure: each prompt you submit is a disclosure decision, and metered APIs of tokenized models turn your proprietary context into the training data of someone else, or at least into that someone else’s operational understanding of your business.

Separate the model from the harness

This is where our view coincides with Karp’s and practical recommendations start. The value for AI in a family office is not in the model, but in the harness: the tooling, data pipelines, prompts, agent workflows and governance around whatever model is the best this quarter.

The models are becoming interchangeable components. The gap between frontier closed models and open-weight alternatives becomes smaller every day, and for a wide variety of real work – document analysis, reporting, summarization, data extraction – the open-weight models running on infrastructure controlled by yourself are already sufficient. In one of the most widely cited tests of this year, an open-source model inside a well-designed harness proved to be able to perform an enterprise migration task roughly at one-sixteenth of the cost of a frontier model used on its own. Slower? Sure, but this is usually not latency-critical for the family office, and economics compound.

Open-source software is the natural go-to for the harness precisely because so many other builders are building on open-source solutions it gives you no lock-in: an open agent framework or orchestration layer connects to any model over an API: a frontier model where capability matters, an open-weight model running privately where cost or confidentiality matters. The future for most offices is exactly this combination of the two. The point is not open weights everywhere. The point is the ability to choose which model to use at any time, because the layer you invest in belongs to you.

In some cases, the case for open-weights is even stronger: cross-border structures dealing with EU, UK and US jurisdictions simultaneously or offices in the jurisdictions requiring strict data residency rules may find themselves unable to send their data to the frontier labs, and thus require the use of open-weights not as a cost optimisation, but as a requirement of sovereignty.

Where frontier models are still needed

Fronteir models still have an important place, especially for heavy-lifting tasks: codebase-building or code refactoring, building something from scratch, and one-shot prompts. In other words, where capability matters, requires some brute force and no private data needs to be sent outside the perimeter. Route that work to the best lab this month, through your harness.

The picture changes as soon as the task relies on context: a private model is weaker in capability, but give it the full access to the office’s data – documents, history, structures that never leave the perimeter – and it starts with the context no frontier model is allowed to see. This full context is precisely what the frontier providers try to gather because they understand its value is higher than any next increment of raw capability. Kept on the infrastructure controlled by you, it works exactly the other way round: it makes a weak model almost as powerful as the frontier one for structured and recurring work.

Division of labor is straightforward: frontier capability for public one-off heavy-lifting. Private models with full context for anything involving family’s information. The harness allows you to run both, and switch between them as the balance shifts.

What does this mean in reality?

One idea is to approach your AI architecture the same way you use custody: never let the same entity both hold your assets and control the reporting.

This means a couple of things should be considered. First using the open-source tooling that treats models as swappable components, keeping your data pipeline and institutional knowledge in systems under your control, demanding any vendor to be able to point to a credible exit path and importantly, reserving the open-weights for the workflows where confidentiality, data residency or cost make it the right solution.

The families that did best over generations were rarely the ones that picked the single winning stock or just built one business. They were the ones that structured for resilience and kept their options open. AI will be no different. Rent the model. Own the harness.

—

About the author

Francois Botha helps design and incubate the family offices of the future through Simple, which pairs high-touch advisory with technology-led solutions to help private wealth owners build offices that last. From formation and strategy through to AI-powered operations, Simple’s team and network of independent advisors provide the knowledge, agentic tools, and governance to move forward with confidence, while keeping the human judgment that matters most at the centre.

He wrote a weekly column on family office strategy for Forbes and speaks regularly at private wealth, family office and technology events.

Most industries arrived at empathy late. For years, they competed on features and on price, and only recently concluded that how a client feels might be the last durable difference between them. There is now a name for this, the empathy economy, and a growing literature arguing that as machines absorb the thinking, feeling becomes the work that remains for people.

The family office started from a very similar place. A family builds its own office because it not only wants returns and efficiency, but also advice, judgment, discretion, and a relationship that holds across decades, and those are qualities that standardised services rarely provide with any reliability.

Empathy sits right at the centre of all things related to this organisation and that makes the family office a particularly useful place to watch how AI affects the empathy economy, because the shift is arriving from two directions at once.

The machine moves toward empathy

The first direction is the one most people picture. AI is becoming more capable of the human parts of the work, and an entire field now builds systems that read tone, sentiment, and context rather than only process text. Several years ago, people were already presenting technology that could read how people feel as the next commercial frontier. Inside a family office, this shows up as software that remembers everything, anticipates a need before anyone raises it, holds the full history of a relationship, and does so with something close to warmth.

That capability is very possible today and can definitely be useful, so it’s something to keep an eye on.

Attentiveness that a machine can perform becomes attentiveness that is no longer scarce, which is the dynamic that emerged when firms began selling the feeling itself. The subtler point sits underneath it. The memory that makes any of this possible has to live somewhere, and when it lives with a vendor, the relationship it supports becomes something the vendor holds too, and can eventually move.

The human is freed toward empathy

The second direction deserves more attention than it usually gets. Consider what a family office executive’s week has actually contained: reconciling bank accounts, building custom reports that few people read twice, reviewing legal documents and drafting the next ones, and sorting correspondence into folders. For anyone who has spent a career on that kind of work, it represents thousands of hours that produced order without producing relationship or judgment. AI absorbs almost all of it.

So the question worth sitting with is:
What becomes possible once that time returns with abundance?

Efficiency is the short version of the answer. The longer version is presence: knowing the family rather than just the ops and the balance sheet, being available for the difficult conversation instead of being occupied by the paperwork behind it, mentoring the rising generation, and a lot more. The relational work was always the role’s purpose, but the administrative load was the price of doing it, and this is the first technology that credibly lowers that price toward zero.

Which empathy you scale

Both forces push toward the same outcome, and that is what makes it easy to misread. “More empathy” sounds like a clear win. But the two kinds of empathy in play here are competing. A system that performs closeness and a person who finally has the time to offer it can occupy the same role, and only one of them is a real relationship. When the software takes that role, the time that was meant to go back to the family goes somewhere else instead, and the family usually senses it long before anyone admits it.

This is the part that tends to get under-planned, because clearing the admin does not, by itself, bring anyone closer to the family. Admin-zero simply returns the time, and what that time becomes is a separate decision that most offices never consciously make. Left to drift, much of it gets quietly reabsorbed into other work, which is why the offices that handle this well are the ones that deliberately choose to spend the recovered hours on people while keeping the memory and the relationship on infrastructure they own rather than parking both with a vendor. For most industries the empathy economy is a new demand to keep pace with, but for a family office it is much closer to the job it always had, and the real work now is making sure the machine takes over the paperwork and leaves the relationship where it belongs.

About the author:

Francois Botha helps design and incubate the family offices of the future through Simple, which pairs high-touch advisory with technology-led solutions to help private wealth owners build offices that last. From formation and strategy through to AI-powered operations, Simple’s team and network of independent advisors provide the knowledge, agentic tools, and governance to move forward with confidence, while keeping the human judgment that matters most at the centre.

He previously wrote a weekly column on family office strategy for Forbes and speaks regularly at private wealth, family office and technology events.

Manyone is a strategic design consultancy with sixteen offices across Europe and North America. I sit on their expert panel, and last week they hosted a show and tell of their AI products. What was fascinating was to see how they’re approaching AI from an experimentation angle but slowly moving to implementation. They’ve spent the past two years turning their service work into a portfolio of AI-powered products. REAL helps clients generate and validate product concepts. LIVE handles real-time content creation. SIGNS monitors change across hundreds of data sources. RADAR runs futures research. HORIZON captures and organises signals. COLLECTIVE knits company knowledge into one accessible system, with an internal agent called M1 they refer to as their Million Brain. CX VISION builds shared customer vision. COMPETE tracks competitor activity.

Eight products. Each is a fairly basic AI application underneath. The asset isn’t the model. It’s the wrapper: the methodology, the branded interface, the integration with the rest of the work and their services. The same work the agency used to deliver as a service is now packaged, repeatable, and running on AI. They call it productized services.

The pattern matters for family offices for three reasons.

Adjacent industries see it first

Strategic design and consulting firms tend to be a few years ahead of financial services in adopting new operational patterns. Faster cycle times, more competitive pressure, more willingness to experiment in public. Manyone has been working with AI in the open. Family offices have been working in private. Both are service businesses. Both run on accumulated knowledge and high-trust relationships. Both are now figuring out how to use AI without breaking those relationships.

If you only watch your own industry for AI patterns, you’ll be late. The interesting moves are showing up in advertising, design, law, consulting, and product development. The shapes those moves take are the shapes that arrive in family offices next.

The productization lens

A productized service is work that used to require expert hands every time, repackaged so it runs as a product. Same outcome for the client, but the methodology is captured, the steps are repeatable, and the marginal cost of each delivery is low.

This is the same logic behind the building blocks framework applied at the level of whole services rather than individual skills. The wrapper is what holds the value. The model inside can change, the data source can change, the underlying tool can change. The productized service stays.

For family offices, this is the missing frame. Most operational work in a family office is service work delivered bespoke every time. There’s no SOP, no template, no reusable methodology that survives staff turnover. Productizing means writing it down, building the loop, and accepting that the next time can be faster.

What translates directly

Several Manyone products map onto family office work with very little squinting. SIGNS, their data intelligence tool, scans hundreds of sources and surfaces patterns. Swap “industry news” for portfolio company developments, regulatory updates, peer office moves, and family business events, and you have what most analysts currently do manually with a coffee and a newsletter pile.

RADAR and HORIZON handle futures research and signal capture. For a family with multigenerational wealth, those tools translate directly to scenario planning, sector exposure analysis, and long-horizon trend tracking. The output is the same shape: a curated, regularly updated view of what’s shifting.

COLLECTIVE and M1 are the most directly applicable. Manyone built a single agent that holds the firm’s accumulated knowledge: project history, expert profiles, methodologies, prior work. For a family office, the equivalent is the institutional knowledge layer we’ve covered in earlier pieces. The IPS, the deal history, the family principles, the relationship map, the legal positions. All queryable, all maintained, all owned by the office. We’ve discussed why this matters. Manyone shows it’s also buildable as a discrete product.

LIVE produces on-brand content at speed. For family offices, this is beneficiary updates, board briefings, family communications, and routine reporting. This is not bout the strategic decisions, just the recurring outputs that currently consume disproportionate time.

What your office could productize

The exercise is straightforward. List the work the office does every quarter. Quarterly reports. Compliance reviews. Diligence workflows. Family briefings. Capital call processing. Tax season prep. Most offices have at least four of these that happen on a predictable cadence and look broadly similar each time.

Each is a candidate for productization. Write the methodology once. Build the loop. Run it on the office’s infrastructure with its own data and verification standards. The next quarter, the next event, the next briefing all take less time because the wrapper holds.

This isn’t a transformation programme but more a list of recurring tasks and a decision to stop doing each one from scratch.

The disposition is the lesson

The takeaway from Manyone isn’t to copy their products. It’s to copy their disposition. They looked at their own service work and asked which parts were repeatable enough to package and which parts were bespoke enough to leave as services. The family offices that ask the same question will find the answers more concrete than they expect.

And you may even discover that the productized version of your office is already inside there.

For years, “human in the loop” has been the reassuring phrase in AI governance. It meant the AI didn’t run alone. Someone reviewed. Someone signed off. Someone caught the error before it shipped. The loop was a metaphor for accountability, and being in it meant you were the checkpoint.

That phrase is about to mean something completely different.

The most cited line in AI coding this month came from Peter Steinberger: “You shouldn’t be prompting coding agents anymore. You should be designing loops that prompt your agents.” Boris Cherny, who built Claude Code, said the same thing a few days earlier from the other side:

“I don’t prompt Claude anymore. I have loops running. They’re the ones prompting Claude. My job is to write loops.”

The loop they mean isn’t a review checkpoint. It’s a small program that prompts an agent, reads what it produced, decides whether it’s done, and if not, prompts it again. It runs on its own, on a schedule, with budget caps and halting conditions. Matt Van Horn synthesised the discourse and put the shift cleanly: the job moved up an altitude. From writing the code to writing the thing that writes the code.

For all businesses including family offices, the same question now applies. Are you in the loop that approves the work, or the loop that designed the system that approves the work? Both are valid but only one compounds.

People: the role moves up an altitude

The analyst who currently writes the quarterly report becomes the architect of the quarterly-report process. The compliance officer who currently reviews every transaction becomes the designer of the loop that flags anomalies. The COO who currently coordinates the team becomes the operator of an orchestration layer.

Van Horn’s research is clear that this isn’t job replacement. Cherny himself says great engineers matter more, not less. Someone still has to decide what to build, what’s worth doing well, what edge cases need a human. The same applies to a family office. Someone has to decide which loops to build, which tasks deserve them, and which decisions should never leave a human’s hands. The job didn’t vanish. It moved.

Processes: the unit of work changes

Today, work in most organisations including family offices is measured in tasks. Review this memo. Draft this brief. Reconcile this account. With loops, the unit becomes the process itself. The memo-review process runs until done. The brief-drafting process produces the final version. The reconciliation loop catches the variance and routes it for human attention.

The temptation is to call this automation. It isn’t, because automation runs the same steps every time. A loop reads the current state, decides what’s needed, takes action, checks its work, and decides whether to continue. The decision in the body of the loop is what’s new. One engineer in the Van Horn corpus called it cron with a hat on. Loops are cron plus a decision-maker.

This matters for family offices specifically because most operational work doesn’t have stable, predictable steps. Every office likes to work a bit differently. Every deal memo needs slightly different research. Every quarterly report needs slightly different commentary. Every compliance event needs slightly different attention. Loops handle this naturally. Hardcoded automation cannot.

There’s a useful side effect. Most family offices never produced proper SOPs for this work, partly because the steps felt too variable to commit to paper. A loop captures the methodology instead. The structure stays the same, the specifics flex inside it. Writing the loop is finally writing down how the office actually works.

Technology: the asset that compounds

The reusable unit inside a loop is a skill. Van Horn’s sharpest line: a loop without skills is a while-true around a stranger. A loop that calls a library of sharp, tested, named skills is a system that compounds.

This is where the building blocks framework lives. The seven portable layers are exactly what loops read from and write to. Identity gives the agent context. Skills give it tested recipes. Memory gives it state. Verification gives it the check before it acts. Each layer is named, portable, and owned by the office.

Without those layers, a loop is expensive theatre. With them, it’s the operational layer that lets a four-person office produce what eight people used to.

The honest constraint

Loops cost money. Token bills compound. Uber capped its engineers at fifteen hundred dollars per person per tool per month after burning its annual AI budget in four months. Van Horn’s research converges on three hard stops for every serious loop: maximum iteration count, no-progress detection, dollar budget ceiling. Without them, you have what one engineer called a slop machine.

For a family office, those constraints translate into governance. What’s the budget for this loop? When does it escalate to a human? How does it know it failed? These are operational risk questions every office already answers for human-staffed processes. Loops just need them written down and enforced inside the right environment.

This is also where “human in the loop”, the old governance phrase, fits cleanly into the new structure. It’s one of the halting conditions. The loop runs, hits a defined threshold, escalates to a person, waits. The human is still in the loop, just in a more specific place. Not as a general overseer, but as a specific gate the loop knows to call.

But it’s still early days like with many other AI terms and technologies, so not everyone in the discourse agrees with the Steinberger framing. The sharper critique is that a loop built around an agent is “a token-denominated bureaucracy”, and that production-safe loops should be deterministic code with the LLM as a narrator inside strict gates. This means the loop is the controlled process and the agent is what fills the slots.

Which loop are you in?

The shift from tasks to loops is the redesign that AI actually enables. Not faster typing. Not better drafting. A change in the unit of work itself, and through it, a change in what the office’s people, processes, and technology need to look like.

“Human in the loop” used to be a defensive posture: don’t let the AI run unsupervised. The new loops run unsupervised by design, with the supervision built in as gates and verification steps and dollar ceilings. So the question shifts. Are you in the loop that approves the output, or the loop that designed the system that approves the output? The first is a checkpoint. The second is an altitude.

Most offices will need both. The ones that only choose the first will spend the next eighteen months wondering why their AI investments aren’t paying back.

Most articles about AI for family offices focus on which model to pick or which tool to buy. The harder question, the one that determines whether any of those tools will actually work over the long term, is what sits underneath. The infrastructure decision shapes everything: which data the office can use, which audits it can run, which providers it can swap, which costs it actually carries.

There’s no single right answer. But there is a coherent picture of what good looks like, and it has the same shape across most offices that have taken AI seriously.

Three deployment options

The first decision is where the office’s AI work physically runs. There are three coherent options.

On-premises hardware is the maximum-control choice. The office buys or leases its own servers, places them on its premises or in a co-location facility it controls, and runs all AI inference there. No cloud dependency. No shared tenancy. No terms of service that can change without notice. It suits offices with existing IT infrastructure, strict security requirements, or jurisdictions with data residency mandates. The cost sits in upfront capital and ongoing operational overhead.

Private hosted cloud is the more common starting point. The office rents a dedicated environment from AWS, Azure, or GCP with no shared tenancy. Models run in a virtual private cloud with private endpoints, encrypted storage, and no public internet exposure. Lower upfront cost, higher operational predictability, still under the office’s control. The trade-off is that some part of the stack is now operated by someone else, even if the data sits inside the office’s tenancy.

A hybrid architecture splits the difference. Sensitive workloads run locally, lower-risk workloads route to a private cloud layer. A gateway manages what moves between environments. This suits larger offices with distributed operations or specific data classifications that need different treatment.

The layered stack

Below the model, every sovereign architecture has the same set of layers. The names vary by vendor but the functions don’t.

Inbound data is where the office’s information enters the system. Email, calendar, portfolio platforms, accounting books, document stores, meeting transcripts, market feeds. The architecture decides which sources connect, how, and on what schedule.

An ingestion layer parses what comes in. PDFs get OCR’d. Audio gets transcribed. Documents get cleaned and tagged. This layer is unglamorous and does most of the actual work.

A memory layer holds the result. Vector embeddings for semantic search, structured stores for relational queries, document indexes for full-text retrieval. This is where the office’s institutional knowledge lives in a form the AI can actually use.

The local LLM is where inference happens. Open-source models like Llama, Qwen, or Gemma run on the office’s chosen infrastructure. For most family office tasks (summarisation, document analysis, drafting, structured output), modern open models perform at or close to frontier quality.

Agents and workflows sit above the model. They read from the memory layer, call the LLM, take actions through MCP connectors or APIs, write results back. This is where the office’s specific processes get expressed, and where the building blocks framework lives.

An audit and compliance layer logs all of it. Every prompt, every response, every action, with timestamps and authorisation. Without this layer, the architecture isn’t defensible to any auditor, regulator, or principal who asks.

A governed outbound gateway is the last piece. When a task genuinely benefits from frontier capability (extended reasoning, very long context, specific multimodal work), the office can route a sanitised query to Claude, GPT, or Gemini through an API. PII is stripped first. The call is logged. The response comes back into the local environment.

How the models plug in

With the infrastructure defined, model choice becomes a configuration decision. The local LLM is whichever open-source model performs best on the office’s tasks today. Three months later, when a better model lands, the office swaps it. The memory, the agents, the audit trail, and the data flows all stay in place.

The same applies to frontier models through the gateway. Today the office routes complex reasoning to Claude. Next year maybe to Gemini, or a new model from a startup. The gateway changes its target. Nothing else moves.

That’s the practical sense in which the architecture delivers sovereignty. Not because the office never uses commercial models. Because when it does, the office controls the conditions: what gets sent, what gets logged, what gets stripped, and which provider holds the conversation.

The wider point

Most AI infrastructure decisions look like vendor decisions. They aren’t. The vendor question (Claude or GPT or open-source) sits inside the architecture question, which is where the office’s AI work lives and on whose terms.

Offices that invest in the architecture question first end up with something durable: a system that absorbs model changes, vendor changes, and policy changes without being rebuilt every two years. Offices that start with the vendor question end up rebuilding every time the vendor changes its terms, its prices, or its model.

The infrastructure isn’t glamorous. It’s the part of the system that should outlast every individual tool that runs on top of it.

Every family office has a centre of gravity, and for most, it’s the investment book of record. The portfolio management system (or sheet) that tracks what the family owns, how much it costs, and its returns. Beside it sits the accounting book of record: the general ledger, the fund administration platform, the source of truth for reporting. These systems are sticky for good reason. Years of accumulated transactions, historical performance, audit trails, and operational context live inside them. The cost of leaving is enormous. The cost of losing them is unthinkable.

But something is busy changing for all businesses. A recent piece from a16z made the case that in the enterprise world, CRM is becoming just an input. The value is moving from the database to the reasoning layer above it, the system of intelligence that pulls from the record, synthesises across sources, and tells you what to do next. Their analogy is Facebook: the friend graph didn’t disappear, but the newsfeed became the thing people actually used. The graph became infrastructure.

For family offices, this shift arrives with a twist. The data that matters most isn’t all in one system.

The structured and the unstructured

A family office runs on two kinds of data. The structured kind lives in systems: portfolio positions, NAV calculations, capital calls, tax lots, and accounting entries. These are the books of record. Well maintained, well understood, well protected.

Then there’s everything else. Deal memos written in Word documents. Board minutes are stored in shared drives. Investment committee notes captured in email threads. Legal correspondence scattered across inboxes. Family governance principles that exist as a PDF somewhere, or worse, only in the memory of a senior partner who’s been there for decades.

This unstructured layer is where the real institutional intelligence lives. The reasoning behind decisions. The relationships. The family’s preferences and risk tolerances expressed in language rather than numbers. No portfolio management system captures it. No accounting platform indexes it. Until recently, no tool could read across all of it at once.

The intelligence layer

AI changes this. An AI system can pull data simultaneously from the investment book, the accounting system, the document store, the deal pipeline, and the family governance files. It can read a capital call notice, cross-reference it against the current liquidity position, check whether the commitment was approved in the last IC memo, and flag whether the cash is available. All before anyone opens a spreadsheet.

That’s the system of intelligence. It doesn’t replace the books of record. It sits above them and treats them as inputs. The portfolio system still holds the positions. The GL still holds the entries. But the layer that synthesises across them, that tells the CIO what needs attention this morning, that prepares the quarterly report by pulling from six sources instead of requiring manual reconciliation, that’s where the operational value is moving.

The advantage of being small

Family offices have an unexpected edge here. In the enterprise world, the system-of-record vendor is racing to build the intelligence layer itself. Salesforce launched Agentforce. HubSpot is embedding AI across its platform. The risk is that the same vendor who holds the data also owns the reasoning on top of it. Lock-in deepens.

Family offices typically use smaller, more bespoke systems. Portfolio management from one provider. Accounting from another. Documents in a third. This fragmentation, long seen as a disadvantage, becomes an asset when the intelligence layer is something the office builds and owns independently. The office controls the reasoning. The office decides which systems feed into it. The office can swap a provider without losing the intelligence above it.

You don’t want the vendor who stores your investment data also controlling the layer that interprets it. You want that layer to belong to the office, be maintained on the office’s infrastructure, and governed by the office’s rules.

The record stays, the value moves

The books of record aren’t going anywhere. A family office will always need a portfolio management system, an accounting platform, and a secure document store. But the gravitational pull is shifting. The system that matters most won’t be the one that holds the data. It’ll be the one who reasons across all of it.

The ones that wait will find that layer being built for them, by the same vendors who already hold the data. The offices that recognise this early will build it themselves.

The record-keeping now becomes just another input. The intelligence becomes the hire, the process and the asset.

Download the PDF cheatsheet here:

Dont Comply With The Eu Ai Act Simple Signals

155KB ∙ PDF file

Download

Download

The EU AI Act has become a lightning rod. Some say Europe is regulating itself out of innovation. Others argue each member state will interpret it differently, making enforcement patchy at best. The compliance deadlines have already shifted: high-risk obligations originally due August 2026 were pushed to December 2027. The whole thing, some will tell you, is a problem for big tech and not for you.

Six things you can do

Fine. Do not comply with the EU AI Act. Instead, just look at what the Act actually asks for and decide whether a well-run family office should be doing those things regardless.

1. Train your team and appoint a lead

Article 4 of the Act requires AI literacy. That has been in force since February 2025. It means the people in your office who use AI systems should understand what they are doing. Not at a technical level. At a practical one. What goes in, what comes out, what the limits are, and when a human should override the result.

It also means someone should own this. Not a Chief Technology Officer. A point person who tracks which tools the office uses, ensures new staff are briefed, and keeps the skills conversation current as the tools change. You already do this for compliance systems and portfolio platforms. AI is no different. The Act just formalises what good management already looks like.

2. Know what AI systems you are actually running

Before you can manage risk, you need a register. What AI tools are in use? Who approved them? What data do they access? Most offices cannot answer these questions because nobody has made the decision to ask them.

The Act calls this a risk assessment framework. A family office can call it an afternoon. List the tools, map the data flows, note who uses what. That exercise alone reveals more about your operational reality than most strategy documents.

3. Tell people when they are talking to a machine

Article 50 covers transparency. If your AI system interacts with people, those people should know they are dealing with AI. The compliance deadline for this was pushed to December 2026. But the principle is straightforward: do not let someone believe they are communicating with a human when they are not.

For a family office running an internal assistant used only by staff and the principal, the scope is narrow. But if that assistant touches any external communication, if household staff interact with it, or if it generates content that reaches anyone outside the office, the expectation applies. This is not a regulatory burden. It is basic professional conduct.

4. Keep humans in the loop

The Act requires human oversight for high-risk AI systems. But “keep a human in the loop” is not a regulatory invention. It is what every principal expects already. No family office wants an AI system making consequential decisions about investments, staffing, or compliance without a person reviewing the output.

The question is whether your current setup enforces that or merely hopes for it. A verification layer built into your AI workflow, one that requires sign-off before output leaves the system, turns a principle into a process.

5. Know where your data goes

GDPR never went away. The AI Act sits on top of it. If your office uses a hosted AI tool, your data travels to someone else’s infrastructure. Do you know which jurisdiction? Do you know whether it trains the model? Do you know whether you can delete it?

Self-hosting does not reduce the obligations. It does make them significantly easier to evidence. All processing stays local. No cross-border transfer questions. No standard contractual clauses. No adequacy decisions to monitor. For a family office handling sensitive principal data, that clarity is worth the infrastructure investment on its own.

6. Keep records

The Act requires logging and record-keeping for high-risk systems. But an audit trail is not a regulatory imposition. It is what allows you to answer the question: what did the AI do, when, and with what input? If something goes wrong, if output is challenged, if a decision needs revisiting, the logs are what protect the office.

Offices that own their infrastructure can run per-prompt observability. Offices using third-party tools should at minimum know what logging the provider offers and whether they can access it.

The gap is not the regulation

The EU AI Act was drafted over two years ago. The debate about whether it is the right approach will continue. The deadlines will shift again. Member states will interpret provisions differently. None of that changes the fact that every item above, literacy, transparency, inventory, oversight, data governance, record-keeping, is something a professional family office benefits from regardless of jurisdiction.

You do not have to comply with the EU AI Act. But if you run through the list and find your office is already doing these things, you will have better governance, clearer risk management, and a system that earns the trust of the people who depend on it.

Jensen Huang called CEOs who blame AI for layoffs “lazy” this week. Sam Altman admitted his predictions about entry-level job losses were “just off.” And Dario Amodei softened his forecasts.

The reversal of opinion by AI CEOs is partly commercial. OpenAI and Anthropic are both preparing high-profile IPOs that require broad investor confidence but also, the mass unemployment wave has not arrived.

(Source: AI’s Real Job Shock Is the Missing Rung, May 2026)

However, AI-attributed job cuts are real and increasing. It has been reported that 83,387 US job cuts in April 2026, with AI cited for 26%, the top stated reason for the second consecutive month. The mechanism, however, is more specific than the headline suggests.

AI is not eliminating the junior analyst. It is absorbing the work before the analyst has had the chance to understand why it matters.

What the junior role was actually for

The pitchbook deep dives, the research summaries, the first-pass screening memos were never primarily about output. They were about repetition of tasks. The analyst who spent three years on the late night grunt work of investment analysis was developing the pattern recognition that eventually becomes judgment.

You learn what a good business looks like by looking at a hundred businesses. You learn what questions to ask in diligence by asking them badly first. You learn when a number does not add up by spending enough hours in spreadsheets to feel the anomaly before you can articulate it. The global economy has been shaped with an apprenticeship learning model for decades.

Admittedly, it is slow, expensive, and sometimes inefficient but it works in ways that are genuinely difficult to replicate through other means.

Let us get to the numbers, data shows the model is under pressure.

Employment among workers aged 22 to 25 has fallen 13% since late 2022 in the most AI-exposed occupations. Furthermore, recent college graduate underemployment sits at 41.5%. With that in mind, 37% of companies now plan to replace entry-level roles with AI outright.

The work that used to generate the repetitions is being absorbed before the repetitions can do their job.

The crisis nobody is tracking

42% of institutional knowledge exists only in individual heads. In most organisations, this becomes visible when someone leaves. In private capital, it becomes visible differently, when people who should have developed judgment have not had the experiences through which judgment is built.

This is not a story about AI taking jobs. It is a story about the transfer mechanism for knowledge breaking down before anyone has built a replacement.

For a family office, the stakes are specific when the organisation is small and processes are informal. The knowledge of how a principal thinks about a deal, which relationships are strategic, what the family values beyond returns, this lives in the people who have been there long enough to absorb it.

With that in mind, 65% of family offices still rely on manual processes. The founding generation hold an enormous amount of information in their heads while the next generation often does not know what they do not know.

Internal budgets are shifting more towards mid-career hires and managing directors who drive revenue as the organisation optimises for immediate output and quietly erodes its own continuity.

What the agentic future requires

AI agents execute tasks with remarkable consistency, when tasks are documented, processes defined, and context structured.

The DELEGATE-52 benchmark found that agents degrade significantly over extended professional task chains precisely because real professional work is contextual, not procedural. The agent needs the organisation to have made itself legible.

Legibility means writing down not just what gets done, but why (need context).

It means capturing the decision criteria a senior investor applies intuitively after two decades of pattern recognition. Documenting which answers in a diligence process have historically been warning signs. Or treating institutional knowledge as an asset to be maintained, not a byproduct of having the right people in the room.

While tedious, it is a payoff that compounds. An organisation that spends two years systematically capturing institutional knowledge is not two years ahead on model capability. It is two years ahead on organisational intelligence, an advantage that grows with every decision captured, rather than eroding as frontier models equalise.

None of this is what AI vendors are selling. It is harder and slower than deploying a new tool. But it determines whether AI compounds institutional capability or simply processes tasks in a vacuum.

What to do…

The new entry-level job should not be “do what the AI cannot do.” That is too passive, and it misses the point.

The more useful frame: junior roles should be redesigned around supervising AI under human accountability, critiquing outputs, verifying sources, spotting missing assumptions, turning summaries into decisions.

That requires managers to slow down in the right places.

Three practical questions before acting:

1. Before compressing a junior role, map what it taught. If the answer is unclear, the knowledge it carried has not been documented. That is the first problem to solve, not the role to cut.

2. Before automating a research process, document the decision logic the analyst applied. What made a source credible? What triggered a second look? What would have caused escalation? The AI needs this to run reliably. So does the next person in the seat.

3. Before deploying agents into a workflow, verify the workflow is legible. If it cannot be described precisely enough to write a prompt, it cannot be automated safely. The legibility gap is the real bottleneck, not the model.

The second-order risk is a seniority cliff five to ten years out. Seniority is the accumulation of thousands of solved problems and crises averted.

If the current generation of juniors never encounters those problems, because AI resolves them automatically, they may not develop the intuition senior roles require.

The work that remains

The AI job loss story was always too simple. What is happening is more granular: the work through which people developed judgment is being automated before the judgment has been built, and the knowledge those people carried is not being captured before they move on.

Centuries of accumulated practice, contract, and tacit knowledge do not unwind at the speed of a model release.

The doomsday predictions about AI and jobs are opaque. The quieter risk is the erosion of how institutions learn that is still very much underway.

This edition of Field Notes was written by Oliver Yorke, Head of Community & Growth

And what caught our attention this week

A few things the Simple team has been reading, testing, or talking about over the last week

Tools we tried:

• Anthropic - Introducing Claude Opus 4.8: Modest but tangible improvements on its predecessor, with one detail worth noting for anyone using Claude for agentic work: early testers report it is significantly less likely to make unsupported claims or let errors pass unremarked.

• ElevenLabs ships a dubbing model that carries emotion across languages: ElevenLabs’ Dubbing v2 model translates videos across 90+ languages while preserving the original speaker’s tone, performance, and delivery. It’s designed to make dubbing both more authentic and cost-efficient. Watch it in action.

Worth reading:

• Proteins and Zuckerberg: Mark Zuckerberg and Priscilla Chan’s Biohub just released new Evolutionary Scale Models, creating an engine to map, predict, and design proteins with the open system showing results against cancer and immune disease targets.

• Apple jumps in the AI race: Apple’s long-overdue AI Siri finally appears to be taking shape, with Bloomberg giving a glimpse of the revamped assistant, rebuilt on Google Gemini, with a dedicated ChatGPT-style app and support for third-party AI agents.

• The Knowledge Infrastructure Problem - James Word, Medium: A clear articulation of why scaling AI without institutional memory is scaling blindness. An organisation two years into systematic knowledge capture is not two years ahead on model capability. It is two years ahead on organisational intelligence.

Capital signals:

• Anthropic closes $30B raise at $900B valuation: The second $30B round Anthropic has closed inside a single calendar year. Revenue trajectory: $87M run rate in January 2024, $1B by December 2024, $9B by end of 2025, $30B in April 2026.

• Cognition raises $1B at $26B valuation: Fueled by the success of its AI coding assistant, Devin with enterprise customers including Goldman Sachs, Citi, Mercedes-Benz, and the US Army and Navy.

• Kopa.ai raises a tidy €2M seed round: Co-led by XTX Ventures and Practica Capital to build AI agents that autonomously run marketing, creative, budgeting, and store operations for e-commerce brands.

• Modiqo closes as $3 million pre-seed round: Assisting companies in moving from experimental workflows to repeatable and production-ready systems.

The switching has started. Teams that began with ChatGPT moved to Claude when Projects arrived. Others went back when GPT-5.5 landed or now Codex. Some are testing Gemini’s workspace integrations, or running open-source models locally. Every few months, a new capability or a new pricing change pulls people toward a different platform, and each time, the accumulated context and institutional knowledge stays behind.

This movement will accelerate. Anthropic just launched Claude for Financial Services, Legal, and Small Business within a single fortnight, each built on the same architecture: text-based skills, external connectors, structured context. OpenAI, Google, and every serious agentic platform are converging on the same pattern. The tools read the same inputs now. The question is whether the system you are building lives inside a single provider or sits independently, ready to travel with you wherever you go. The practical answer is building blocks.

Seven layers, all portable

A practitioner framework gaining traction in agentic AI circles describes seven layers that sit between a person and any AI tool. Each layer answers a specific question the model would otherwise have to guess at. For family offices, every layer maps to a document type the office already knows.

Identity defines who the AI is working as and who it is working for. A system prompt that establishes the role, the organisation, the communication style, and the boundaries. Think of it as the AI equivalent of a governance charter: the foundational document that says who we are and how we operate.

Context is the institutional knowledge the AI needs to do useful work. Investment mandate, family governance principles, reporting cadence, key relationships, standing decisions. This is policy statements such as the investment policy statement brought to life: the same standing commitments and constraints, structured so the AI can apply them. Most offices have never written this down in one place. That gap is the first one worth closing.

Skills are repeatable instruction sets attached to recurring outputs. A deal screening template. An IC brief structure. A compliance review checklist. These are the standard operating procedures the office has always needed, now written in a format the AI can execute. Each skill ensures the AI produces the same standard of output regardless of who triggers it, which is the difference between individual and institutional AI use.

Memory, connections, verification, automations

Memory is a live document that tracks what is happening now. Active projects, open questions, recent decisions, ongoing diligence. The AI reads it at the start of each session. If the office maintains board minutes or an active deal tracker, the memory file serves a similar function: a running record that keeps the AI current without re-explaining from scratch.

Connections define what the AI can reach. Calendar, email, documents, data sources, external tools. Each connection extends the AI’s ability to act rather than only advise. As the agentic supply chain develops, these connections will extend beyond the office into the systems your service providers operate.

Verification is the layer most people skip, and the one closest to the compliance review protocols the office already runs. It defines what the AI must check before presenting output: source citations, regulatory flags, tone review, factual cross-references. Without verification, you are trusting the model’s confidence. With it, you are building in the sign-off step that every serious process requires.

Automations connect skills into sequences. A new deal arrives, the screening skill runs, a summary is produced, and if it clears the filter, the IC brief skill triggers. These are the workflow procedures the office runs every week, documented in a format where the AI can run them consistently rather than waiting for someone to remember.

Why text files matter

Every layer described above is a text file. A markdown document, a set of instructions, a structured prompt. Claude, ChatGPT, Gemini, an open-source model running locally: they all read the same files. The model becomes interchangeable. The system you built around it stays.

This also means you are building an asset the office owns rather than a dependency on any single platform. The sovereign infrastructure argument applies here at the practical level. Your context, your skills, your memory, your verification rules: these belong to the office, stored on the office’s infrastructure, version-controlled and maintained like any other operational document.

The compounding return

The first agent you build will take time. You will write the identity prompt, assemble the context file, define the first few skills, set up the memory document. It might take a weekend of focused work. The second agent will take an afternoon, because most of the foundation already exists. It shares the same context, the same memory, the same verification layer. You are adding a new role and a few new skills on top of an existing base. The third takes less still.

This is the compounding dynamic that separates offices building systematically from those still treating AI as a series of individual tool decisions. Every building block you create serves every agent you build afterwards. The context file feeds the deal screener, the compliance reviewer, and the communications drafter equally. The investment in documentation pays forward across every workflow.

The tools will keep converging. Every major AI workspace is adding the same capabilities: structured context, persistent memory, reusable skills, external connections. The vendor comparison matters less with each passing quarter. What matters is whether you have built the system underneath, the one that makes any tool work for your specific office, your specific processes, and your specific standards.

The question the room couldn’t answer

We hosted our first thematic breakfast in our Copenhagen office last week, a handful of family offices joined by guest speakers Jonas Boesen Roiy from Microsoft Defense and Thomas Schubbe from European Energy, gathered around a single theme: AI and security.

The conversation covered serious ground. The frontier model race is effectively between the US and China, with most organisations watching from the outside. The pace of change is fast enough that AI will expose human error before most organisations have developed the judgment to catch it.

But the most important moment came near the end, when someone finally asked the question that had been hovering over the whole conversation: does anyone actually have a backup plan? What does a kill switch look like?

(The Simple Office - Table Breakfast Event on 5th May)

Stairs or a metal box

There is a distinction worth making before going further.

An escalator that stops working becomes a staircase, slower, more effort, but functional. You still get where you are going. However, an elevator that stops working leaves you in a metal box, waiting for someone else to fix it.

Most organisations adopting AI are building elevators, and they are doing it with remarkable confidence and almost no attention to what happens when the system fails. Manual workarounds have been quietly let go of and the people who knew how to run the old processes have gone.

The team members with the instinct to notice when something does not look right, the ones who ask the uncomfortable question, who spot the anomaly before it becomes a problem, are often the first to go in a round of AI-driven efficiency.

They are also exactly who you need when the elevator stops.

In a survey of organisations that made AI-driven redundancies, 55% subsequently admitted they had made the wrong decisions. 31% were worse off financially. 1 in 3 lost critical skills they could not replace and 55% discovered that the AI they had brought in to reduce headcount actually required more human oversight than they had initially accounted for.

(The Simple Office - Table Breakfast Event on 5th May)

Every Integration is an Exposure

Most AI security conversations start in the wrong place, with the attacker. The phishing email now indistinguishable from the real thing. The synthetic voice on a wire transfer call. These threats are real and getting worse.

But the risk that surfaced more persistently in Copenhagen was quieter. It is not the adversary who breaks in and it is the system that was never quite right to begin with.

A model trained on incomplete data produces confident outputs on questions it was never equipped to answer. A process that has run without incident for six months will keep running, until the underlying assumptions are wrong in a way that matters.

Model accuracy can degrade by 15% to 40% within three months as real-world data drifts from training conditions, yet most organisations have no systematic process for detecting this.

An accountability structure where the AI recommended and the human approved without really reviewing will hold together perfectly well until the outcome is bad and nobody can clearly say who owned the decision.

None of this requires a malicious actor.

What it requires is what most organisations already have, a system deployed faster than it was understood, into operations that have since come to depend on it.

Three questions that almost nobody can answer

The defence against both categories of risk starts in the same place, not with better technology, but with three questions every principal should be able to answer about their own organisation.

What data is going in, and what is happening to it? If you cannot explain where it goes, who owns the output, and what happens when the output is wrong, you have not adopted AI. You have adopted a dependency.

Who is accountable when the model is wrong? When something goes wrong with an AI-assisted decision, the answer to that question needs to be a person and a process, not a shrug. If it is not immediately clear, the governance layer does not yet exist. For family offices, where nearly half have experienced a cyberattack in the past two years, 78% believe a successful breach would trigger investor panic or withdrawal, the absence of that layer is not an abstract risk.

Do you have a kill switch, and have you ever actually used it? Not a theoretical one, a real process for reverting to manual operation, with people who know how to run it and documentation that guides them.

If the system stopped tomorrow, could your team operate? For how long, and at what standard?

This was never really about AI

The escalator or elevator question is not about technology.

It is about the kind of organisation you want to be when things go wrong, one that has thought through its failure modes, maintains the capacity to operate independently, and knows exactly who is responsible when an automated system produces a bad outcome.

That organisation is not slower or more cautious than its peers, it is more durable.

And in a period where the systems underpinning daily operations are changing faster than the governance structures designed to manage them, durability is the competitive advantage most family offices are not yet building.

The kill switch is not a backup plan. It is evidence that you understood what you were building in the first place.

This edition of Field Notes was written by Oliver Yorke, Head of Community & Growth

And what caught our attention this week:

A few things the Simple team has been reading, watching, and talking about.

Worth watching:

• Why Software Is Becoming Worth $0 In The Age of AI: David Haber, General Partner at a16z, on why vertical AI companies are not selling software — they are replacing labour. TAM is no longer software spend, it is labour spend. 48 minutes, worth it.

• The Briefing: Financial Services: Anthropic’s virtual event for financial services. Directly relevant for family offices thinking about where Claude fits into compliance, reporting, and investment workflows.

Worth reading:

• From System of Record to System of Intelligence - a16z: The reasoning layer above the database is where the next generation of enterprise software is being built. The CRM becomes infrastructure. The AI layer becomes the hub.

• Is Software Losing Its Head? - a16z: Agents may kill muscle memory as a moat, but they do not kill operational logic and context. The companion piece to the system of intelligence article.

Capital signals (A big May):

• Fractile raises $220M: London AI inference chips. 100x faster than existing hardware, 90% lower costs. Anthropic reportedly in talks to buy their chips when production starts in 2027.

• Isomorphic Labs raises $2.1B: DeepMind’s drug discovery spinout. Thrive Capital leading, UK Sovereign AI Fund participating. Vertical AI at its most ambitious.

• Recursive Superintelligence raises $650M at $4.65B: Fewer than 30 staff, no product, backed by Nvidia and GV. Building AI that improves itself. The most speculative bet of the week.

• Stitch raises $25M led by a16z: Cloud-native operating system for financial institutions in the Gulf. a16z’s first GCC investment.

• Embat raises €30M: AI treasury management, 400 European clients, automates up to 80% of manual treasury tasks.

• Wirestock raises $23M: Supplies training data to six of the largest AI labs. $40M ARR. The data supply chain for AI is becoming its own asset class.

• Cala raises $7M pre-seed: Barcelona based, Cala is building infrastructure for explainable AI agents. A big pre-seed for the Iberian market.

For the past two years, the AI conversation in family offices has centred on which model to use. Claude or ChatGPT. Gemini or something open-source. Teams trial one, switch to another, compare outputs side by side. The assumption is that the model is the thing that matters. Choose the right one and the work improves.

That assumption is increasingly incomplete. The AI landscape has evolved through three phases, and most offices are still making decisions as if only the first one counts. In the first phase, everything was about the model: bigger, more capable, better trained. If you wanted better output, you needed a better model. In the second phase, the focus shifted to what you put in front of the model. Prompting, examples, retrieval, structured context. The same model could behave completely differently depending on what it was given to work with. We are now in a third phase, and the shift is significant. The question is no longer which model to choose or what to tell it. It is what environment the model operates in.

Where the idea came from

The concept has a specific origin. In February 2026, Mitchell Hashimoto, the founder of HashiCorp and co-creator of Terraform, published a post describing a habit he had developed while working with AI agents: every time an agent made a mistake, he engineered a permanent fix into the agent’s environment rather than trying to improve the model. He called this practice “engineering the harness.” Within weeks, both OpenAI and Anthropic published follow-up articles expanding on the idea.

In the development world, that environment is called a harness: the runtime that wraps around a model and gives it persistent memory, reusable skills, execution tools, error handling, and approval gates. The model stays the same. The harness determines how reliably it can act.

The data here is striking. A recent benchmark from Endor Labs tested the same models across different environments and found that the results were dramatically different.

GPT-5.5 scored 61.5% on functionality in its native environment but 87.2% in a different harness. Same model, same week, two completely different results.

OpenAI CEO Sam Altman acknowledged this directly in a recent interview, saying he no longer thinks of the model and the environment around it as separable things. And research from Adnan Masood found that 65% of enterprise AI failures trace back to defects in the operating environment, not the model’s reasoning. Context drift, schema misalignment, state degradation. The model was fine. The system around it was not. McKinsey’s research reinforces the point from a different angle: 80% of agentic AI implementation time is consumed by data engineering and governance, not by the model itself.

Beyond code: the guidance layer

The harness concept originated in software development, but the principle applies wherever AI agents are doing work. In a workspace context, the operating environment is the guidance layer you build around the model: the system instructions that define how it should behave, the rules that govern what it can and cannot do, the context files it draws from, the memory it retains between sessions, and the verification steps that catch mistakes before they reach anyone else.

When a family office uses an AI tool to draft a client communication, the quality of that draft depends far less on whether the model is Claude or ChatGPT than on whether the tool knows the family’s communication preferences, understands the relationship history, follows the office’s tone, and checks the output against the right standards before presenting it. That configuration, the institutional knowledge and operating principles wrapped around the model, is the guidance layer. It is functionally the same concept as the developer’s harness, applied to knowledge work.

Without it, every AI interaction starts from zero. The model produces generic, context-free output that requires heavy editing. With it, the model operates within the boundaries and with the knowledge that make its output genuinely useful for your specific situation.

Why this reframes the vendor question

Most family offices evaluate AI tools by comparing the models underneath them. That comparison will become less meaningful over time, because every major tool is converging on the same set of capabilities. They all support structured context, memory, connections to external systems, and reusable instructions. The interfaces differ. The features arrive at different speeds. But the underlying architecture is increasingly similar.

88% of AI agent pilots never reach production. The top blockers are evaluation gaps, governance friction, and reliability issues, all problems in the operating environment, not the model. (Source: Forrester/Anaconda 2026)

The offices that will get the most from AI are the ones that invest in the layer that actually determines quality: the guidance they build around whichever model they choose. That means defining how the AI should communicate, what context it should draw from, what rules it must follow, and what checks it should pass before output reaches anyone. The model provides the capability. The environment you build around it determines whether that capability translates into something your office can trust.

The model is not the whole story. It may be the least important part of it.

From what Sequoia have been preaching about services as the new software, this week, OpenAI and Anthropic separately structured joint ventures with private equity firms to acquire engineering and consulting firms, not to build better AI, but to help businesses use the AI that already exists.

OpenAI’s vehicle, The Deployment Company, is raising roughly $4 billion from 19 investors including TPG, Brookfield, and Bain Capital, valued at $10 billion. Anthropic’s joint venture is finalising $1.5 billion with Blackstone, Goldman Sachs, and Hellman and Friedman. Both will embed engineers directly inside businesses to redesign workflows around AI.

The comparison drawn was to Palantir, years of engineers embedded inside clients, not because the software was incapable, but because organisations were not structured to absorb it without help. OpenAI and Anthropic have reached the same conclusion about their own products a decade later.

The numbers confirm what they already knew. Deloitte’s 2026 State of AI survey found 60% of employees now have access to AI tools, fewer than 60% use them regularly. Only one in four organisations has converted more than 40% of its AI pilots into production. Separately, 85% of enterprises are pursuing AI but only 39% have deployed it at scale.

(Source: Current approach to transformation with AI. Source: Deloitte State of AI 2026 - deloitte.com)

The consulting industry priced this in before the AI labs did.

BCG generated $3.6 billion from AI work in 2025, 25% of total revenue. McKinsey says 40% of its projects are AI-related. Accenture has more than 85,000 AI and data professionals.

The services market around AI deployment already exists and is enormous with OpenAI and Anthropic simply entering it, not creating it.

The bottleneck was never the model as it was always the organisations institutional deployment.

Why This Is Not Your Problem In The Way It Is Theirs

The natural response for a family office principal is either relief or anxiety but both miss the point.

The Deployment Company exists because large enterprises have large, specific problems, procurement cycles lasting months, data in legacy systems nobody fully understands, thousands of employees with different workflows, governance structures requiring multiple sign-offs before anything moves. A genuine labour-intensive problem.

A family office is a different kind of organisation entirely. Eight to twelve people. A principal who can describe what good output looks like. A defined set of recurring processes, IC prep, deal screening, client reporting, compliance review, on a predictable cadence. No procurement committee and no IT department blocking installation.

The structural advantage is not that the technology is simpler. It is that the organisation is smaller and more legible to the people running it.

The gap that requires an army of consultants at enterprise scale can be closed by one or two people with genuine intent.

Most family offices have not recognised this because the framing of AI adoption has been dominated by enterprise stories.

Building the System Underneath

The most useful reframe comes from a concept gaining traction in AI practitioner circles: the personal agentic operating system. The argument is simple. Most people treat AI as a series of tool decisions, which model, which app, which subscription. The tool is the least important decision and what matters is the system built underneath it.

For a family office context, four of them matter most right now.

Context: An AI with no genuine understanding of how your office works will produce generic output, because you have given it generic input. The fix is a documented, maintained description of your context: investment mandate, recurring processes, standing decisions, key relationships. Most people have never written this down and that is the first gap.

Memory: A live document covering active projects, ongoing diligence, open questions, and decisions made. The AI reads it at the start of each session. The difference between an AI that knows your current deal pipeline and one that does not is the difference between a tool and a team member.

Skills: Repeatable instruction sets attached to recurring outputs, an IC brief to the same standard every time, a screening summary that follows the same logic regardless of who runs it. Skills are how individual AI use becomes institutional AI use.

Processes: Skills connected into sequences. Inbound deal arrives, first-pass filter runs, screening summary produced, IC brief triggered if it clears. The process is what makes AI run consistently rather than only when someone remembers to use it.

Every AI tool is converging on the same surface capabilities where the model matters less over time and the system built underneath is what compounds.

The Structural Point

OpenAI and Anthropic are spending billions because large enterprises cannot build that system without external help. The context is too distributed, the processes too fragmented, the organisation too complex to assemble quickly without dedicated resources.

A family office can build it in weeks. The context exists in the principal’s head. The processes already run, they just have not been written down and handed to an AI. The skills can be built one recurring output at a time.

The families that navigate this period well will not be the ones that waited for a deployment company. They will be the ones that spent the next quarter building something simple, specific, and repeatable.

Three Questions Worth Asking this Week

The commercial case for closing this gap is now measurable. Grant Thornton’s 2026 AI Impact Survey found organisations with fully integrated AI are nearly four times more likely to report revenue growth than those still piloting, 58% vs 15%. The difference is not the technology. It is accountability and process.

The three questions below are a way to locate where your office sits on that spectrum.

1. Which recurring output in your office varies most depending on who produces it? That is the first skill to build, not the most complex process, the most unreliable one.

2. Does your AI have genuine context about your organisation, or are you re-explaining from scratch every session?

3. If you were out of the office for two weeks, could your team produce the same AI-assisted outputs to the same standard? If not, you have individual AI.

The gap between individual AI and institutional AI is what context, memory, skills, and processes are designed to close.

This edition of Field Notes was written by Oliver Yorke, Head of Community & Growth

Work with Simple

Simple works with family offices on AI adoption, building the context, memory, skills, and process layers that turn individual AI use into something the whole organisation can rely on.

If the questions above have uncomfortable answers, the conversation is worth having. hi@andsimple.co

And what caught our attention this week:

A few things the Simple team has been reading, testing, or talking about.

Tools we tried or even built:

• Mike: A lightweight personal workflow tool that makes AI use persistent and consistent across sessions.

• ProgramBench: Challenges AI agents to recreate software executables using only documentation and experimentation, without access to source code. A useful stress test of what agents can actually do independently.

• Harvey’s Legal Agent Benchmark: Open-source tool for evaluating AI agents on legal tasks. Worth watching as the first serious attempt to benchmark agents on domain-specific professional work rather than general capability.

Worth reading:

• Writer’s 2026 Enterprise AI Adoption Survey: 1,200 executives, 1,200 employees. The numbers that matter: AI super-users save 9 hours per week, 4.5 times more than laggards. Only 29% see meaningful ROI. 75% of executives admit their AI strategy is more for show than internal guidance. The deployment gap is structural.

• Google is writing a licensing agreement, not building a consultancy: While OpenAI and Anthropic acquire engineers, Google is in talks with Blackstone, KKR, and EQT to license Gemini across their entire portfolio. No embedded engineers, mo services arm and distribution speed over implementation revenue. The contrarian position on how to win enterprise AI and worth reading alongside this week’s article.

Capital signals:

• Ethos raises $22.75 million Series A: AI consulting startup co-founded by DeepMind scientist led by new investor Andreessen Horowitz.

• China to invest $50 billion in Deepseek: Talks to raise money from China’s National Artificial Intelligence Industry Investment Fund, a one-year-old government-backed fund with around $8.8 billion in capital.

• Fazeshift raises $17M Series A: AI agents automating accounts receivable for finance operations teams. Vertical AI going deep into one specific, high-friction back-office workflow. The pattern of purpose-built agents targeting a single function rather than horizontal tools continues.

• Tessera Labs raises Series A led by a16z: AI-native platform automating enterprise software transformations and system integrations. Capital moving into the layer that makes AI deployments actually connect to existing systems.

• Panthalassa raises $140 million for wave-powered offshore AI data centres: Peter Thiel backing offshore compute nodes powered by ocean energy, data transmitted to land via satellite. The most unusual infrastructure bet of the week.

• Pit raises $16 million seed led by a16z: Stockholm-based enterprise AI platform out of stealth, building customised AI solutions for fragmented business workflows. Pit is led by the co-founders of European scooter giant Voi, including Voi CEO Fredrik Hjelm. a16z leading two separate AI workflow rounds this week, Tessera and Pit, is a directional signal worth noting.

• And Elevenlabs lines up Hollywood: The voice AI company hits $500M ARR and expands its enterprise footprint with including BlackRock, Nvidia, Jamie Foxx, and Eva Longoria

There is a natural progression happening with AI that most people don’t see until they’re inside it. You start with a chat. You ask questions, draft documents, summarise things. It feels useful but limited. Then you go deeper. You discover that the tool can work with files, write code, build interfaces, connect to databases. Suddenly you’re not just asking questions. You’re building something. A dashboard for your portfolio. A tool that automates a reporting workflow. A client-facing application that pulls from your own data. The capabilities are real, and the speed is extraordinary. What would have taken weeks with a development team can take an afternoon.

This is where the builder camp lives. And increasingly, it’s where people from the other two camps are arriving, whether they planned to or not. The tools have become good enough that someone with a clear idea of what they need can prototype it themselves. No engineering background required. Andrej Karpathy, one of the founding researchers at OpenAI, coined the term “vibe coding” in early 2025 to describe this shift: building software by describing what you want and letting AI write the code. It has scaled faster than almost any previous developer tooling category.

By Q1 2026, vibe coding tools represent roughly $4.7 billion in annualized spend with 92% developer adoption. Non-technical user adoption grew 520% year-on-year. The consistent finding across multiple sources: 90% of AI agent demos would not survive 24 hours of real production traffic.

The 80% moment

The first 80 to 90% of a build happens remarkably fast. The AI generates working code, structures a database, designs an interface, connects to an API. For someone who has never built software before, this feels like magic. And in many ways it is. The prototype works. It does the thing you described. It handles the use case you had in mind.

This is the moment where many people believe the project is nearly done. It looks right. It functions. The demo is impressive. But the distance between a working prototype and a production-ready tool is where most builds stall. That gap between the pitch and the proof applies to your own internal builds just as much as it applies to vendor demos.

Temporal’s Production AI Stack Report found that 62% of teams lose measurable time or revenue to reliability issues in AI systems. Their summary of the problem: today’s AI tools make it easy to build dazzling demos, but harder to build systems that survive in production.

The last 10 to 20% is where the difficulty concentrates. Edge cases the prototype never encountered. Error handling that was never designed. Data validation that doesn’t exist. Authentication that isn’t secure. A database schema that works for ten records but won’t scale to ten thousand. These are the things that separate something you can show from something you can rely on.

What the last mile actually requires

Having spent a decade running a design and build company for the web, I’ve watched this pattern from the other side. Building a proper product involves a sequence of decisions that most first-time builders don’t know they need to make. Planning, architecture, build, testing, iteration, QA, delivery. Each stage has its own discipline. Each stage catches problems the previous one missed.

When you leave those decisions to an AI agent, the agent will make them for you. It will choose a database structure, design API routes, decide how authentication works. The problem is that the agent doesn’t know your business requirements well enough to make the right choices. It makes plausible choices. Technically sound choices. Choices that work in the demo. Whether they work for how your office actually operates, how your data actually flows, and how your compliance requirements actually function is a different question entirely.

From an enterprise perspective, the gap widens further. Security hardening, access controls, audit trails, data encryption at rest and in transit: these are decisions that require deliberate design, not defaults. For a family office handling sensitive data across jurisdictions, the difference between a prototype’s security posture and a production application’s is the difference between a locked door and a closed one.

According to the Taskade State of Vibe Coding 2026 report, 63% of vibe coding users are non-developers who often lack the knowledge to identify missing authentication, exposed API keys, or insecure database configurations.

A new kind of client relationship

Here is what makes this moment interesting for service providers and consultants. For a decade, client engagements started from scratch. Requirements gathering, scoping, design, build, delivery. Very few clients arrived with a half-built product they had made themselves. When they did, it was usually from another developer or agency, which is a different situation entirely, because professionals build to recognisable standards.

That pattern is changing. Clients are now arriving with prototypes they built using AI. And this is genuinely useful. The client has already translated their business requirements into a working model. They know what they want because they’ve already tried to make it. The thinking has been done. The prototype is the brief.

The opportunity sits in the last mile: taking what the client built and making it production-ready. Fixing the architecture. Hardening the security. Handling the edge cases. Making it reliable enough to run as part of the office’s stack rather than sit on someone’s laptop as a demo.

The first 80% may not need outside help at all. The value shifts to completion, polish, and durability. For service providers, the question is whether they’re set up to receive a client’s AI-built prototype and finish it, rather than insisting on starting over.

What this means now

The tools will keep getting better. The 80% will creep toward 85, then 90. But the last mile involves judgment, experience, and design discipline that the current generation of AI handles unevenly. For offices that are building, the honest question is whether the prototype on your screen is something you can actually put into production, or something that needs a different kind of attention to get there.

The distance is shorter than it used to be. But it still needs crossing.

A Different Kind of AI Question

Most family offices are asking the wrong AI question first.

The internal debate centres on tools, which ones to adopt, which processes to change, how fast to move. These are reasonable starting points but they are not the questions with generational consequences.

The more important question is where AI is creating value outside the organisation, and whether your capital is positioned to reach it before the window closes.

Anthropic secondaries are rumoured to trade at valuations exceeding $1 trillion, with annual recurring revenue reported above $30 billion. OpenAI sits at approximately $850 billion on around $25 billion in revenue, growing at triple-digit rates.

More than 80 AI-native companies have already surpassed $100 million in annual recurring revenue, a pace that has no precedent in venture capital history.

(Source: Number of AI companies with $100M+ ARR, Sapphire internal analysis, Feb 2026)

All of this has been created in private, in growth rounds, in secondaries, in the vehicles that family offices with the right access and the right conviction have been building positions in for the past three years.

The question most family offices are debating is operational but the one with real wealth implications is not.

Where the Value is and Where Most Portfolios Are Not

The AI infrastructure hyperscale buildout is legitimate and here to stay. The top five are projected to spend roughly $600 billion in capex in 2026, up around 36% year‑on‑year.

But the families capturing the most significant returns from AI are not doing so through listed technology equities. They are doing it through private market positions, growth equity, venture, secondaries, in the companies building the intelligence layer itself.

The J.P. Morgan 2026 Global Family Office Report surveyed 333 offices across 30 countries. 65% named AI as their top investment priority and more than 79% had 0% infrastructure exposure despite its role as the physical backbone of AI.

The stated priority and the actual portfolio are pointing in different directions. That is not a construction problem but is a knowledge and access problem.

The companies at the centre of this cycle were not built overnight. They were funded across multiple private rounds year on year and the investors who captured most of the value creation were in the room at Series B and Series C, not at the IPO.

Public markets have historically valued late-stage technology companies more conservatively than the private rounds preceding them, which means investors arriving at listing day are more likely to see compression than expansion.

The recent trajectory of AI infrastructure companies trading publicly is consistent with that pattern.

History is consistent on what happens to investors who arrive at the IPO rather than before it.

Across 147 venture‑backed technology companies that went public between 2010 and 2021, pre‑IPO investors achieved an average gross multiple of roughly 20-40x on invested capital, while IPO‑day investors earned near‑zero or negative returns within six months.

In a separate analysis of more than 721 US IPOs between 2012 and 2016, the median six‑month post‑IPO return was negative around 7-8%.

The companies driving the current AI‑valuation surge are not exceptions to this pattern as the most extreme version of it.

The Knowledge Constraint

Access is always a constraint. But the primary constraint for family offices is the knowledge required to act with conviction.

SpaceX at $1.75 trillion on approximately $16 billion in revenue implies a sales multiple approaching 100 times. OpenAI at $1 trillion on $25 billion in revenue is 40 times and the company faces $14 billion in projected losses for 2026.

These are not simple investment decisions. They require a realistic view on revenue quality, the path to profitability, and what the regulatory environment means for long-term holders within the context of the AI paradigm.

Anthropic derives an estimated 80% of revenue from enterprise customers, historically the stickiest, most predictable revenue base and the kind public markets assign higher multiples to. OpenAI has a larger consumer base but significantly higher infrastructure costs.

Although these are different businesses have different risk profiles, being able to distinguish comes from a genuine familiarity with how AI businesses actually work. That originates from engagement with the technology, not simply reading a pitch deck.

The families that built meaningful private equity positions in earlier cycles did not wait for consensus. They developed genuine conviction about how the asset class worked, where the returns came from, what to ask when something did not add up. It was knowledge that compounded overtime. By the time the late arrivals showed up, the entry point had moved.

AI is at that moment now.

The Connection Most Family Offices Have Not Made

Most family offices have adopted AI the same way they adopted cloud software a decade ago. They bought the subscriptions and told the team to use it but that is it.

Although that approach produced some efficiency gains then, it still difficult to produce immediate investment conviction.

The families building genuine internal AI capability are doing something different. They are not just using AI, they are developing the institutional judgment to evaluate it. Building workflows that depend on genuine comprehension rather than surface familiarity with a product that always sounds confident.

That distinction matters operationally and it matters more as an investment capability.

The same judgment required to know when an AI system inside your organisation is producing something unreliable is the same judgment required to assess whether an AI company’s revenue is as defensible as its pitch deck suggests.

It is the same skill applied in two directions.

Internal adoption, done with real seriousness, is not just an operational improvement. It is an investment in the institutional knowledge that determines whether you can participate in the most significant private market opportunity of the past two decades with genuine conviction.

The families that will be on the right side of this cycle are not the ones that had the most AI subscriptions.

They are the ones that built enough genuine understanding to know what they were buying when it mattered.

And what caught our attention this week:

A few things the Simple team has been reading, testing, or talking about.

Tools we tried or even built:

• Claude for Creative Work: Anthropic has launched Claude for creative work, enhancing its ability to support writing, design, and ideation with better tone, style, and iteration.

Worth reading:

• Darwinian Specialization in AI: The inference market is fragmenting as workloads diverge, across latency, multimodal, and edge, creating distinct infrastructure needs and room for multiple winners.

Capital signals:

• Reliable Robotics raises $160 million: Autonomous aircraft systems, led by Nimble Partners. Physical AI is the theme threading through the largest non-frontier-lab rounds of 2026.

• Omni raises $120 million Series C at $1.5 billion: AI-enabled analytics platform backed by Iconiq Growth.

• Legora hits $5.6 billion: Head to head with Harvey, Nvidia’s CVC arm NVentures, has backed Legora as part of a $50 million Series D extension. Legora is tiltering closer to Harvey’s $11 billion valuation, the battle of legal AI is on, which celebrity cover is next…

Work with Simple

Simple works with family offices at the intersection of AI adoption and AI investment helping them build capabilities to navigate both with genuine conviction.

If the questions above are ones your team is working through, we would like to be part of that conversation.

Reply to this email or reach out to us at hi@andsimple.co

Every family office we speak to sits in one of three technology camps. The first runs on spreadsheets. Excel is the operating system. Data lives in files, communication happens through email, and processes are manual, familiar, and surprisingly resilient. The second has adopted products. Portfolio management platforms, data aggregation tools, CRM systems, accounting software. Some use big-ticket enterprise solutions, others have found more budget-friendly alternatives. The third builds. These offices connect various systems, manage their own integrations, and would be the first to jump on an AI agent platform. They treat technology as something they own rather than subscribe to.

Each camp has a different relationship with AI. But the message for all three is the same: there is no single AI tool that does everything your office needs. And the sooner that becomes clear, the better your decisions will be.

The tool is not the stack

What most offices haven’t yet recognised is that their technology is already a stack, regardless of what they use. The Excel office has a stack: spreadsheets, email, a shared drive, a phone, maybe a portal from the custodian. The product office has a stack: the portfolio platform, the accounting system, the CRM, the document store, the reporting layer. The builder has a stack by design.

Previously, the focus was on point solutions. A tool for data aggregation. A tool for reporting. A tool for portfolio management. Each solved a specific problem in isolation. That thinking is dead. The surface area for AI is far broader: email, calendar, knowledge management, research, drafting, communication, compliance monitoring, decision support. Intelligence that’s useful requires more tools connected together, not one tool doing everything. AI has made the connections between these functions visible and valuable in a way they weren’t before.

Why one tool will never be enough

The temptation is understandable. ChatGPT or Claude can draft a memo, summarise a document, analyse a dataset, and answer a research question. For an individual user, it can feel like a single tool that does everything. But AI chat tools aren’t built to be your operating system. Not all of the office’s work lives inside them, and for most offices it shouldn’t. The portfolio data sits elsewhere. The compliance logic sits elsewhere. The family’s communication history, governance documents, and relationship context sit elsewhere. Portfolio management and reporting platforms aren’t built to handle this either. They solve their own problem well, but they were never designed to hold the full breadth of an office’s knowledge.

The real issue is that institutional knowledge needs to flow across tools, not live inside any single one. The office’s context, its reasoning patterns, its accumulated understanding of the family and the portfolio, has to be accessible wherever work happens. That means a knowledge layer the office controls, independent of any individual product.

The sovereign AI vision takes this further: full control over infrastructure, models, and data, with nothing leaving the building. That’s the utopian promise, and for offices that can resource it, it’s increasingly achievable. But even offices that aren’t ready for full sovereignty benefit from thinking in layers. A knowledge layer the office owns. A model layer that connects to whichever provider it chooses. An integration layer that pulls from existing tools. The stack becomes modular rather than monolithic.

The trust question cuts across all three camps

Within each camp, there is a further split. Some offices are enthusiastic about AI and already experimenting broadly. Others are cautious, sceptical, or actively distrustful. Both positions are legitimate, and neither changes the stack question.

The enthusiast still needs to think about where context lives, which tools connect to what, and what happens when a provider changes its terms. The sceptic still needs to recognise that AI is already embedded in the tools they use, often without explicit disclosure, and that understanding the stack is a prerequisite for managing the risk they’re concerned about.

Trust is earned through architecture, not enthusiasm. The offices that design their technology as a stack, with clear boundaries between what they own and what they rent, will be better positioned regardless of how quickly they adopt AI itself.

The view from each camp

The Excel office has an advantage it may not recognise: simplicity. Fewer dependencies, fewer integration points, fewer places where data can leak. The transition to AI can be deliberate and controlled. The risk is staying still too long and finding that the gap between capability and what’s possible has become too wide to close incrementally.

The product office faces a different challenge. The stack already exists, but the office may not see it as one. Each tool was chosen independently, often by different people at different times. AI arrives as another product to add, rather than a layer that should connect what’s already there. The risk is accumulating tools without architecture.

The builder might think they have the clearest view of the stack, but the reality is more demanding than it appears. The target state is a constantly moving target. What you built last quarter may already need rethinking. And building means taking on product management on top of security, maintenance, and every integration update. Every connection is the office’s responsibility. The advantage is control. The cost is that control requires sustained attention, and the landscape won’t wait for you to catch up.

[IMAGE: Deloitte Figure 12 — Level of preparedness for AI adoption. Strategy 42% highly prepared, technology infrastructure 43%, data management 40%, risk and governance 30%, talent 20%. Infrastructure and talent declining year-on-year. Source: Deloitte, State of AI in the Enterprise, January 2026. N=3,235]

Designing the stack you need

None of these camps is wrong. The question is whether the office has looked at its technology as a connected whole and made deliberate choices about what sits where. That means understanding which tools your suppliers use, which systems hold your most sensitive data, and what happens when you need to change a piece.

The offices that do this work now will have the flexibility to adapt as the technology shifts. The ones that treat each tool as a standalone decision will keep solving the same integration problem over and over.

There is no single AI. There is only the stack you design around it.

In the 1840s, Britain built the most extensive railway network in the world in under a decade.

Most companies simply moved their existing cargo onto trains, same routes, same schedules, same commercial relationships, faster delivery.

But the firms that used railways to redesign their operations entirely are the ones that built lasting advantages. And for the rest, they just had faster horses.

Writing for a16z, George Sivulka makes the same observation about AI in Institutional AI vs Individual AI that we are in this same moment of the past where the technology is not the constraint.

The question is whether organisations have redesigned themselves around it or simply replaced the horse with a train and kept everything else.

For family offices navigating AI adoption, that distinction is the whole question.

Most family offices that have adopted AI have given their team members subscriptions. Individuals write better first drafts, summarise faster, research more efficiently.

At the level of the organisation, almost nothing has changed. The investment process looks the same. The IC meeting runs the same way. The weekly report takes the same amount of time.

Sivulka identifies seven factors that separate what he calls institutional AI from individual AI with three landing with particular force across the functions of a family office.

Pillar 01: Coordination - Individual AI Creates Chaos

When every team member uses AI differently, the organisation does not become more productive. It becomes harder to run.

The pattern repeats across every function of a family office. Analysts draft investment memos in their own way, with no shared view of what the output should contain. While two people in operations handle the same administrative process with two different AI-assisted approaches that nobody has sat down to compare.

(Source: Productive individuals (or agents) row in different directions alone. If left uncoordinated, chaos can follow)

The issue is not that people are using AI badly. It is that nobody has decided what good looks like at the level of the organisation. That decision has been left to individuals, and individuals have made it differently.

Where to start

Pick one process more than one person touches, an IC brief, a compliance review, a client report. Define a single AI-assisted workflow for it. One template, one output, one review step and start there.

Pillar 02: Signal - Individual AI Creates Noise

AI has made it easier to produce polished output across every function but this is also the problem.

In investment management, deal flow arrives AI-polished and in higher volume. The quality of presentation is uniformly high. The quality of the underlying opportunity is not. In client services, families receive more communication than ever but more is not the same as better.

Whether they contain what they should is a different question and one the AI will not ask on your behalf.

Defining what good looks like before generation starts is the constraint now.

(Source: AI slop from individual productivity tools is proliferating at a high rate. Humans cannot alone sort through the noise and a class of new AI products is needed to do so)

Where to start

Pick one recurring output, a screening summary, a client quarterly, a board pack. Write down what it must contain before AI touches it and build the workflow around that definition. The filter protects the judgment of the people who use it.

Pillar 03: Bias - Individual AI Confirms, It Does Not Challenge

AI agrees with whoever is asking. Ask it to make the case and it will. Ask it to find the gaps and it will do that instead. The output follows the question and most people ask questions that confirm what they already think.

Every organisation has structures designed to counteract exactly this tendency. Review processes, sign-off layers, external scrutiny, these exist because individual judgment, left unchallenged, drifts toward confirmation.

They were designed for a world where producing a polished argument took time and effort. That friction was doing useful work. AI has removed it across every function simultaneously.

(Source: Individual AI echo chambers fuel division, drawing two humans apart, a dynamic which at scale creating factions in an otherwise coherent organization)

Where to start

In whichever function concerns you most, run one session with a single instruction: find everything wrong with how we currently do this. Bring the output into the next relevant meeting. That is AI being used to challenge the institution rather than confirm it.

What rebuilding the factory actually means

Institutional AI is a design decision, not a product category. It means embedding AI into the process itself, not making it available to individuals who choose to use it.

Across a family office, this looks different by function but the test is the same: does the process produce a consistent output regardless of who runs it?

A compliance review that depends on one person’s prompting habits is individual AI. One built around a defined workflow that anyone can run is institutional AI. The same test applies to client reporting, governance documentation, operational processes, and investment analysis.

To enable these advances, it requires someone willing to take two or three recurring processes and redesign them rather than add AI on top of what already exists.

The tool will be the same for every family office but the process will not. That is where the gap opens and where it compounds.

And what caught our attention this week:

A few things the Simple team has been reading, testing, or talking about.

Tools we tried:

• AI Mode in Chrome: Google’s new way to explore the web with AI built directly into the browser. A step beyond search: the browser interprets, synthesises, and responds rather than returning links. Early days, but the direction of travel is clear.

• Claude Design: Anthropic launched Claude Design this week, a new product that lets you collaborate with Claude to produce visual work: prototypes, slides, one-pagers, marketing collateral. Listen to a helpful guide here presented by the AI Daily Brief.

Worth reading:

• The hourly cost of AI agents is approaching human rates (Toby Ord): The time horizon of tasks agents can complete has grown exponentially. The cost to reach those horizons has grown at the same rate. For some models, hourly costs now rival human equivalents. It is worth tracking for anyone deciding where agentic AI makes real commercial sense.

• Ramp built every employee their own AI coworker: 99% adoption, then most people were stuck. The fix was institutional: shared skills, memory, coordination layers, automations running overnight. The most important thing they learned: the people who got the most value were not the ones who attended training. They were the ones who got a result on day one. A concrete example of what rebuilding the factory looks like in practice. What out for more from Ramp!

• Sam Altman calls Anthropic’s Mythos campaign “fear-based marketing”: OpenAI’s CEO argued this week that presenting a model as too dangerous to release publicly is a convenient way to keep AI in the hands of a small elite. He has used similar framing himself. Worth reading alongside last week’s Field Notes, the gap between what Mythos can actually do and how it has been presented is a thread that is not going away.

Capital signals (Deals Deals Deals):

• Cursor in talks to raise $2 billion at $50 billion valuation: Thrive and Andreessen Horowitz expected to lead. Cursor is targeting over $6B in annualised revenue by end of 2026. A coding tool at a $50B valuation is a signal about where enterprise software spend is heading.

• Cognition AI negotiating funding at $25 billion valuation: The company behind Devin, the autonomous software engineering agent. The valuation reflects continued appetite for agentic AI ahead of revenue scale.

• Era raises $11 million to build software for AI gadgets: $9 million seed led by Abstract Ventures and BoxGroup. Building the platform layer for AI-native hardware. Early, but signals that serious capital is moving into the AI gadget category.

• Loop raises $95 million Series C: AI data platform for logistics backed by Valor, Founders Fund, Index Ventures, and J.P. Morgan Growth Equity. Tackles enterprise data fragmentation across supply chain systems. Vertical AI going deep into a messy domain rather than building general tools. The template for what purpose-built looks like.

• AcuityMD raises $80 million Series C at $955 million valuation: Commercial intelligence for MedTech with a continuously updated knowledge graph mapping physicians, facilities, procedures, and reimbursement dynamics. The data exists but no single company can assemble it alone. AcuityMD does.

• NeoCognition raises $40 million seed: Self-learning AI agents that specialise rather than generalise, backed by Cambium, Walden Catalyst, Vista, and angels including Intel CEO Lip-Bu Tan.

This edition of Field Notes was written by Oliver Yorke, Head of Community & Growth & Isak Mosfeldt Mejlhede, Marketing & Events.

Work with Simple

Simple’s Explore engagement is a structured four to six week process to identify which workflows should be redesigned around AI and to build the institutional layer that makes the difference last.

If the coordination, or bias problem maps onto how your team is operating right now, we would like to talk.

Reply to this email or reach us at hi@andsimple.co

Walk through Times Square this month and you’ll see something new. A legal AI platform called Legora has taken over the billboards with an actor, a knowing smile, and the tagline “Law just got more attractive.” The campaign runs across airports, city centres, and digital screens in New York, London, and Scandinavia. It was shot by an Oscar-winning cinematographer. The creative director described the concept as a dumb idea that turned out to be the smartest strategy.

This is not a product demo. There are no screenshots, no feature lists, no efficiency claims. It is a lifestyle campaign for an AI tool. And it is not the only one. Legora’s direct competitor Harvey has signed an actor best known for playing the fictional lawyer Harvey Specter on television, alongside sponsorship deals with football clubs in Paris and London. These companies are marketing legal AI the way fashion houses market fragrance.

The approach most AI companies still take

For the past few years, most AI tools have marketed themselves on features and functional benefits. Faster drafting. Smarter search. Automated workflows. Take the day off and let AI handle it. The promises have been bold, the demonstrations polished, and the results, for the most part, have left users with high expectations and largely wanting. The gap between what was promised and what was delivered became one of the defining patterns of AI adoption so far.

That approach hasn’t disappeared. But something is shifting alongside it. A subset of AI companies, particularly those with large funding rounds behind them, are moving away from capability claims entirely. They are spending on brand, on celebrity, on aspiration. The pitch is no longer “look what this tool can do.” It is “look who you become when you use it.”

When the demo can't do the work, the brand has to.

What the shift tells you

When software companies start marketing like luxury brands, it usually signals something specific: the product alone can no longer differentiate. The underlying technology has become accessible enough that multiple competitors offer broadly similar capabilities. When the demo can’t do the work, the brand has to.

Legora raised $550 million at a $5.5 billion valuation last month. That level of capital needs to find customers fast, and brand marketing at this scale is designed to create preference where the product hasn’t yet created loyalty. It is a strategy borrowed from consumer goods: build the emotional association first, let the product catch up later.

For the legal AI market specifically, this is accelerating. Firms are being courted not just by sales teams but by campaigns designed to make adoption feel aspirational rather than operational. The question is whether the offices being marketed to can tell the difference.

Why family offices should recognise this playbook

They already know it. Wealth management, private banking, and luxury goods have used aspiration and identity to sell for decades. The imagery changes but the mechanism stays the same: associate the product with a feeling, and let the feeling close the sale.

Family offices have spent years developing the instinct to look past this. The same principal who can distinguish between a wealth manager’s brand and their actual performance can apply that instinct here. The billboard is not the product. The celebrity endorsement tells you about the marketing budget, not the capability.

The question remains the same one that applies to every AI vendor conversation: what has this tool proven it can do, in a context like yours, with data like yours? Has the office seen it work on its own documents, its own workflows, its own compliance requirements? Or has it seen a polished demo and a famous face?

The question worth asking

The marketing is getting better. Campaigns are more sophisticated, the production values are higher, and the emotional pull is stronger. None of that tells you whether the product underneath has improved at the same rate. For family offices evaluating AI tools, the arrival of lifestyle marketing in this space is actually a useful signal. It means the market is maturing, competition is intensifying, and vendors are looking for new ways to create preference. That’s a healthy development for buyers, as long as the office keeps its evaluation grounded in evidence rather than affect.

The billboards will keep getting better. The question is whether you’re buying what’s on the screen or what’s behind it.