The Founder-Multiplier Trap

OneThe log that didn't change

For about two years my corpus has been getting better every month. Not vaguely better — measurably denser, with more of it joined up. Every strategic conversation I have with a model now starts inside a world instead of on a blank page. It reaches back into work I did for a different client in a different year for a different reason, finds the sibling case, tests the new idea against the old contradiction, and hands it back to me sharper than I gave it. It is astonishing, every single turn, and I have said so out loud more than once.

Somewhere in the middle of that two years I did something that spoiled the mood. I went and looked at the escalation log.

Not the volume — the volume had gone up, because the business had. I looked at the shape. Specifically: of the hard calls that came to me this quarter, how many were shapes I had already resolved before? A pricing structure I had ruled on. A delivery boundary I had drawn and explained. A class of client objection I had answered, written up, and filed. The kind of case where the right answer already existed somewhere in the substrate, and the only reason it arrived at my desk was that arriving at my desk was faster than finding it.

The density had not moved. Two years of compounding, and the proportion of already-solved shapes coming back to me was the same as when I started.

That is a strange finding, and the reason it is strange is worth sitting with. Nothing had gone wrong. Client work was better. Turnaround was faster. My own judgment on genuinely novel problems had improved, because I could now reach a much wider field of prior thinking in the seconds when it mattered. If you had audited the firm on any measure a firm normally keeps — revenue, delivery quality, client satisfaction, corpus size, retrieval accuracy — you would have written a glowing report.

And the one number that describes whether the business can survive without me had flatlined.

The trap has a name and a very specific shape

I have written before that succession fails when it is treated as an archive problem — that document counts rise while escalations stay exactly where they were, and that this is not a failure of effort but a failure of object. What I had not accounted for was what happens to that failure when you point a genuinely good AI apparatus at it. The answer is not that the apparatus fixes it. The answer is that the apparatus can make it worse while making everything else better, and can do so invisibly, because the instruments a firm keeps are all pointed the wrong way.

Founder augmentation improves current earnings. Judgment transfer improves terminal value. AI lets them move in opposite directions.

That is the whole argument, and I want to be precise about who it is for. If your firm's value is capacity, or a licence, or a platform, this is not your problem. If your firm's value is judgment — if what clients buy is that a specific person, or a specific small group of people, knows which way to call a situation that looks like six other situations and isn't — then this is the failure mode that will find you, and it will find you in a good year rather than a bad one.

It will find you in a good year because that is the only kind of year it produces.

What this piece owes you

Three things, in order. First, why the visible part of expert work — the questions, the frameworks, the language — was never the asset, and what actually is. Second, the causal chain from "my AI made me better" to "my firm is worth less than it looks", stated as seven steps you can check individually against your own records. Third, a test: a four-arm ablation on temporally frozen cases that returns a verdict on which of four assets you own, with the verdict written down in advance so that neither of us can move the goalposts after the result arrives.

I am not going to re-teach the role-migration ladder or the month-of-absence trial; those are built and published and linked, and this piece assumes them rather than repeating them. Nor is this the design of the ongoing client relationship — that is a separate apparatus with its own dependency gradient, and it names this trap explicitly as a different question requiring a different instrument. What this piece adds is the thing those instruments were not designed for: the dynamic that AI introduces, where the founder's capability and the firm's transferability come apart, and the measurement that separates them.

TwoThe question was never the asset

Start with the thing I used to believe, because it was flattering and it was half right.

The belief goes like this. Someone could sit in on my client sessions, write down every question I ask, and walk away with nothing much. They would have the probes but not the altitude. They would hear the answers and not know which ones mattered. I have said versions of this out loud: even if they duplicated the questions, it wouldn't be meaningful, because they couldn't interpret the results.

That is true. It is also, read carefully, an alarming thing to say about a business you intend to be worth something one day.

The question is only the probe.

Watch the mechanism, because it is more specific than "experience matters". A team finishes a project. I sit with the delivery lead and the client sponsor separately, and I ask something ordinary: what did you value at the end that you did not think you were buying at the beginning? Anyone can ask that. It is eleven words. Write it on a card.

The answer comes back: honestly, the biggest thing was that we resolved all the disagreement about what the numbers meant before anyone started building.

One listener hears: better discovery. Note it, move on, put it in the lessons-learned deck.

What I hear is a collision. That answer runs straight into a body of prior work about selling certainty as a product before selling the implementation it de-risks; into the idea that the friction between two parties' definitions is itself a rent someone is currently collecting; into a question about whether the pre-build reconciliation should be priced and scoped as its own bounded commitment rather than absorbed into a project; into a pattern I have seen fail in a specific way when the client's internal disagreement was political rather than definitional. And then the next question I ask is different from the one I would have asked otherwise, because the answer moved my model.

Same eleven words. Same reply. Two entirely different pieces of work.

This has a name, and it is thirty-six years old

The mechanism is not mine and it is not new. In 1990, Wesley Cohen and Daniel Levinthal named it in a paper that has been cited ever since: the ability of a firm to recognise the value of new external information, assimilate it, and apply it to commercial ends is, they argued, "largely a function of the firm's level of prior related knowledge".¹ They called it absorptive capacity. What you can extract from a signal depends on what you already had.

So the copying argument holds. The screens reveal the promoted answer, never the evaluation function that made promoting it rational — the rejected shapes that looked plausible and failed economically, the abstraction refused because it forked under real conditions, the exception that stayed local when it looked general. Someone can steal the code. They cannot steal the year. Two people can hold the same model, the same transcript and the same list of questions and arrive at radically different useful insight. They have the same compute and the same raw observation. They do not have the same evaluation function.

Good. Now turn it around and point it at yourself.

The uncomfortable inversion

If the questions are not the asset, and the interpretation is what matters, then everything of value in that exchange happened after the answer arrived, inside one person's head, in a step that produced no artefact anyone else could use.

I have just described a firm with no transferable asset in it.

That is the inversion the copying argument sets up and never resolves on its own. "They can't copy me" and "my firm cannot function without me" are the same sentence with different lighting. The first one is what I say to a competitor. The second is what a buyer, a lender, a bank, an insurer or a future partner hears — and unlike the competitor, they are the ones who put a number on it.

Cohen and Levinthal saw this too, and they were blunt about it in language that reads like it was written for exactly this article. They were describing the "gatekeeper" — the individual who serves as the organisation's interface to outside knowledge — and they wrote: "Even when a gatekeeper is important, his or her individual absorptive capacity does not constitute the absorptive capacity of his or her unit within the firm."¹ And they added the warning that lands hardest under conditions like the present ones: when change is rapid and it is not clear where in the firm a piece of outside knowledge is best applied, "a centralized gatekeeper may not provide an effective link to the environment".¹

They wrote that before anybody had a language model. The whole argument I am making is what happens to a centralised gatekeeper when you hand them a very good one.

ThreeFour things we call expertise

Here is the move that makes the rest of this tractable. "Our expertise" is one phrase covering four different assets with wildly different properties, and until you separate them you cannot have an honest conversation about what your firm is worth.

The scarce thing is not a question and it is not a body of knowledge. It is a calibrated ability to interpret an answer under history, reject plausible wrong branches, recognise when the evidence is insufficient, and update after consequences arrive. Note that this is four verbs, not one quality — and each of them can be scored separately, which is what makes the test in Section Seven possible at all.

Split on that definition and you get four layers.

Only the fourth is clearly a terminal-value asset. The first two can support a formidable boutique while leaving the enterprise almost entirely dependent on one person.

Why the third layer is where everyone lies to themselves

Layers one, two and four are easy to be honest about. You know whether your questions are public. You know whether you are the one making the hard calls. You know whether anyone else is producing paid decisions without you, because you would have noticed the free time.

Layer three is where the self-deception lives, because a compiled kernel and an elaborate personal archive are indistinguishable from the inside. Both are large. Both are well-organised. Both produce excellent answers when you drive them. The difference between them only appears when someone else takes the wheel, and most firms never run that experiment, because the founder is faster and there is work to do this week.

I hold myself to this. I have published an evidence ledger for my own firm's estate, and the row for the compiled corpus reads: implemented, internally used as a capability substrate — with the promotion condition being reuse by someone other than the author, and the kill condition being that the whole thing turns out to be embalming rather than transfer. That row is not promoted. I have not earned it. The category error it guards against is the one this whole section is about: a founder's powerful knowledge system is not evidence that capability transfers across a bench.

The economists' version, which is the same table

If you prefer this in the language of the research literature: absorptive capacity is usually split into a potential half — acquiring and assimilating knowledge — and a realised half — transforming and exploiting it, with the ratio between them treated as an efficiency factor. A firm can score extremely well on the potential half and near zero on the realised half. It acquires everything. It assimilates beautifully. And nobody except the founder can transform any of it into a decision.

That ratio is the founder-multiplier trap written as arithmetic, and it is worth noticing that better AI improves the numerator far more easily than the denominator. Capture got cheap. Transformation did not.

FourThe chain

Now the mechanism, in the order it actually runs. Seven steps. None of them is a mistake. Each one is individually defensible and locally correct, which is precisely why the sequence is so hard to see from inside.

Step one: the kernel raises the founder's altitude. A compiled worldview means work starts oriented instead of cold. The prompt becomes a pointer into a compiled field rather than a fresh specification, and the class of problem you can reliably take on goes up. This is real and it is good and I would not undo it. There is even evidence that the effect strengthens with operator tenure: people using the same assistant for six months or more show a measurably higher success rate in their conversations, an association not explained by which tasks they chose.² Read that carefully, though: it is a learning curve on the operator. It is not a learning curve on the organisation, and the gap between the two is the subject of this article.

Step two: harder decisions get routed to you. This is the step people miss, and it is the load-bearing one. It is not that you got faster at the work you already did. It is that the firm's routing rule changed. When a senior person is weighing whether to fight a difficult call through themselves or send it up, the calculation includes how long it takes you to dispose of it. Lower that number and more calls go up. Nobody decides this; the queue discovers it. Every individual routing choice is efficient, and the aggregate is that the hardest 15% of the firm's decisions have quietly become the hardest 25%, all landing in the same inbox.

Step three: AI raises how many such decisions you can personally absorb. Which means the routing change does not produce a visible bottleneck. It produces a busy but coping founder. The field evidence on this is stronger than intuition: in a pre-registered experiment with 776 professionals at a global consumer-goods company, individuals working with AI matched the performance of two-person teams working without it.³ An individual who performs like a team is exactly the condition under which routing everything to one individual stops looking irrational.

Step four: clients renew for access to your discrimination. Not for your methodology, which they could read. Not for your capacity, which they can buy cheaper. For the specific thing that happens when you look at their situation. This is lovely, it is high-margin, and it is the revenue line most tightly coupled to your continued presence in the world.

Step five: the exhaust records your interventions without making them reproducible. This is the quiet one. You are generating more record than ever — transcripts, decision notes, session logs, a growing wiki. Every one of those artefacts records what you decided. Almost none of them records the discrimination that made the decision: the branch you killed in the first ninety seconds and never mentioned, the resemblance you rejected, the piece of evidence whose absence was the actual signal. A firm can build the world's best institutional memory for operating yesterday's version of itself and change nothing about what it does.

Step six: revenue grows around a more productive key person. Which is growth, and which everybody celebrates, and which means that each additional dollar of revenue is more tightly coupled to one biography than the dollar before it. Concentration is not a static property you either have or don't. It is a rate, and in this configuration the rate is positive.

Step seven: terminal value stays constrained, or key-person concentration actively increases. The firm is worth more this year and no more transferable. Possibly less.

That is not a failure of AI augmentation. It is a failure of asset migration.

The correction I have to make to my own version of this

There is a version of the founder-multiplier story that is intuitive, popular and — as far as I can tell from the evidence — wrong. It says: AI is a multiplier, the founder is the highest multiplicand, therefore the founder gains most and pulls further ahead of the bench.

The field experiments say close to the opposite. In the pre-registered study run with 758 consultants at a global management consultancy, on tasks inside the capability frontier, consultants below the average performance threshold improved by 43% against 17% for those above it.⁴ In a staggered rollout across 5,179 customer-support agents, productivity rose 14% on average — but 34% for novice and low-skilled workers, with "minimal impact on experienced and highly skilled workers".⁵ Inside the frontier, AI compresses the skill distribution. It does not stretch it.

So the mechanism cannot be that I gain more than my people do on ordinary work. On ordinary work they gain more than I do, and that is excellent news for transfer.

The mechanism has to be the other two things, and they are the two things the same studies also measured. First, the frontier is jagged: on a task selected to sit outside the capability frontier, consultants using AI were 19 percentage points less likely to produce correct solutions.⁴ Outside the frontier, the model is not a lift; it is a confident wrong answer with a persuasive structure, and discrimination is the only thing standing between it and the client. That residual is where founder judgment still decides — and it is exactly the part that does not transfer by being written down.

Second, routing. Compression at the bottom and a jagged edge at the top together produce a very specific organisational outcome: the ordinary work gets absorbed by capable people with good tools, the residue routes to the person with the discrimination, and the residue is now a larger share of what the firm sells. The bench gets better at the work that was already leaving. The founder becomes the sole supplier of the work that remains.

I find this more troubling than the naive version, not less. The naive version had an obvious remedy — spread the tooling. The real one says the tooling is already spreading, and the concentration is happening anyway, in the layer above it.

FiveWhat you are currently accepting as evidence

If the chain is real, why doesn't anyone notice? Because every piece of evidence a firm currently uses to reassure itself about transfer is evidence of something else. Let me go through them in the order they usually get offered, including the one I offered myself.

"Look how much of our own IP the system pulls into every answer"

This is the one I was proudest of, so it goes first. The reference density in my own strategic conversations is genuinely remarkable — the system reaches back across years of unrelated work and brings the relevant piece forward without being told it exists.

Here is what that demonstrates: orientation. The system can enter a large accumulated world rather than beginning cold, and can join the immediate question to work produced at other times for other purposes. That is a real asset and I am not taking it back.

Here is what it does not demonstrate. Four thresholds remain, and none of them is cleared by a dense set of references:

1. Does the cited material actually support the inference? Retrieval proves relevance was found, not that the argument built on it holds.
2. Did the inference change a decision, or decorate an answer? A great many beautifully sourced paragraphs have never altered anybody's behaviour.
3. Did the eventual outcome support the decision? Which you cannot know at the time, which is why it has to be scheduled rather than assumed.
4. Can somebody else reproduce the quality without you reconstructing the reasoning? This is the only one that speaks to transfer at all, and it is the only one nobody measures.

The honest formulation is narrow and I have had to live with it: the references demonstrate the potential existence of orientation capital. Accuracy, calibration, transfer and economic value still require separate receipts.

A citation-rich response can still be a beautifully sourced sealed mirror.

"Our people say the answers are good"

They are good. That is not the question. The question is whether an answer that is good when you are in the room is evidence about a state of the world in which you are not.

And there is a sharper problem underneath, which is that felt productivity has now been measured against actual productivity and found badly miscalibrated in at least one rigorous setting. A randomised controlled trial of 16 experienced open-source developers across 246 tasks — people with an average of five years' experience on the repositories in question — found that allowing AI tools increased completion time by 19%. The developers had forecast a 24% speed-up beforehand. After finishing, they estimated they had been sped up 20%. They were slower, and they believed the opposite by roughly forty percentage points.⁶

I want to be careful with that result rather than weaponise it. Sixteen developers, one domain, one tool generation, and the authors themselves say experimental artefacts cannot be entirely ruled out; the same group has since revised the experimental design. It is not proof that AI makes experts slower. It is proof of something narrower and, for this argument, more useful: the gap between felt and measured productivity under AI is large enough to be wrong in sign. If that gap can invert a measurement as concrete as elapsed time on a coding task, treat your intuitions about a much softer quantity — whether judgment has moved into your firm — as worth exactly nothing until instrumented.

"The wiki is comprehensive now"

A wiki preserves conclusions, tests and patterns. It does not automatically create calibrated judgment: the capacity to make the call when the pattern doesn't quite fit, and to know which kind of not-quite-fitting matters. Knowledge transfers by reading. Judgment forms by adjudicated repetition with consequences.

Which raises the obvious follow-up: where does adjudicated repetition come from now? Historically it was a by-product of the pyramid — juniors saw hundreds of ordinary cases as a side effect of billing, and built a map of the real world without anyone paying for the mapping. That source is being dismantled in public data. Administrative payroll data covering the period after generative AI became widely adopted shows early-career workers aged 22 to 25 in AI-exposed occupations experiencing a 16% relative employment decline, with employment for experienced workers remaining stable.⁷ A labour-market research institute studying the same shift put the consequence in one sentence: "companies still need human expertise. They just don't have as much need for the entry-level roles where people have traditionally developed that expertise."⁸ Surveyed professionals expect the effect on judgment formation directly: in a 2026 survey of 736 legal professionals across 46 countries, 71% said early-career people still need structured support from experienced colleagues to develop the judgment AI risks displacing, and anticipated it taking a year longer for them to reach trusted independent judgment.⁹

Put those together and you have the accompanying failure to the one this article is about. Your firm's future judges were being manufactured as a by-product of work that is now being done by machines. The founder-multiplier trap says your judgment isn't moving into the firm. The apprenticeship collapse says nobody else's is being formed either. Same decade, same cause, different half of the balance sheet.

"We'd be fine — I'm insured"

The foundational US guidance on valuing closely held businesses has been saying otherwise since 1959, and in language so close to this argument that it is slightly eerie. Revenue Ruling 59-60 instructs that "the loss of the manager of a so-called 'one-man' business may have a depressing effect upon the value of the stock of such business, particularly if there is a lack of trained personnel capable of succeeding to the management of the enterprise", and directs the valuer to consider "the absence of management-succession potentialities".¹⁰

The same ruling does allow insurance as an offsetting factor — "the loss may be adequately covered by life insurance, or competent management might be employed on the basis of the consideration paid for the former manager's services".¹⁰ Read the second clause. The offset is not the money. The offset is that competent management might be employed — that the role is fillable at the price of the role. Insurance covers the transition of a firm whose judgment is fillable. It does nothing for a firm whose judgment isn't, because there is nothing to hire into.

How much is at stake, empirically? The cleanest natural experiment I could find is a study of managerial deaths across roughly 13,000 Danish firms, which is an ablation on the founder conducted by mortality rather than by design. Industry-adjusted operating return on assets fell 0.9 percentage points over a two-year window around the death of a CEO or an immediate family member — "equivalent to an 11% decrease in OROA".¹¹ The same study found no robust equivalent effect for the deaths of individual board members.¹¹ The authors note they cannot separate expected from unexpected deaths, and that this means they are likely underestimating the effect.

Two things follow. One: the discipline of removing a person to find out what they were worth is not an exotic idea I invented; it is how this gets measured when it can be measured at all. Two: an 11% profitability effect is what happens on average across all firms in an economy, most of which are far less person-shaped than yours.

SixThe uncomfortable half

Everything so far has treated absorptive capacity as your advantage. It is time to say the other half out loud, because a version of this argument that only cuts in my favour is not worth publishing.

The client has absorptive capacity too. They own internal history, proprietary evidence, relationships, operating constraints and authority that I do not possess and never will. They know which executive killed the last version of this idea and why. They know which number in the pack nobody believes. They have the situated knowledge, and their AI is improving on exactly the same release schedule as mine.

If prior related knowledge determines what you can extract from a signal, then the client's stack of prior related knowledge about their own business is, in the domains that matter to them, deeper than mine. What I have is a wider comparative base across many businesses. What they have is depth in the one that pays them. Those are different assets and it is an empirical question which one wins on a given decision.

The direction of travel is not encouraging for suppliers. When enterprises stop chatting with models and start wiring them into their own operations, the usage pattern changes shape entirely: 77% of API transcripts show automation patterns, dominated by full task delegation, against just 12% augmentation.¹² Full task delegation is the shape of work that used to be bought from outside. I have argued at length elsewhere that the professional-services unit compresses from both sides as the share of work a client chooses to externalise falls, and I am not going to re-run that argument here. What matters for this piece is narrower and more personal: it means that one of the four asset layers can be beaten not by a competitor but by the client, and you will not see it in your pipeline until it is done.

So the test in the next section has an arm that most advisers would never publish: the client's own team, using their own AI and their own context, on the same frozen cases as me. If they match my kernel on cost, that is not an embarrassment to be suppressed. It is the single most valuable finding the test can return, because it arrives a year or more before the equivalent finding shows up in revenue.

SevenThe test

Everything up to here is a diagnosis. Here is the instrument.

The reason to build one at all is that the alternative — another demonstration — cannot answer the question. If I run a live session and produce something impressive, we learn that the joint system of me-plus-kernel-plus-model works. We learn nothing about which component did the work, which is the only thing a buyer, a partner, a successor or an insurer actually needs to know. The cleanest proof programme separates the sources of advantage rather than producing another impressive answer.

That is an ablation: remove one component at a time and see what the result does. It is standard in machine learning, and — as the Danish mortality study shows — it is how economists infer the value of a person when they cannot run an experiment. Here it is, specified end to end for a firm.

Step one — freeze the cases

Take six to twelve real decisions from your own history. For each one, reconstruct the evidence package as it stood at the decision date and nothing after it. The proposal as it was written, not as it was later amended. The data as it was understood, including the parts that turned out to be wrong. The organisational context as it was described at the time, including the assumptions everyone was carrying that nobody wrote down.

Then remove the outcome. Everything downstream of the decision — what happened, what it cost, who was proved right — goes into a sealed envelope that only the scorer opens.

This is not a formality; it is the whole methodological load. There is a well-developed literature on what happens when a test is run on data that already contains its own answer. In finance the canonical treatment is backtest overfitting: run enough strategy configurations against history and excellent simulated performance is trivially achievable, and because analysts "rarely report the number of configurations tried", nobody can evaluate how overfit any particular result is.¹³

Transfer that to your own situation and it stings. Every demo you have ever given was a configuration you chose. The cases were selected — not dishonestly, just naturally, by the same taste that built the kernel. You have run an unreported number of configurations, and the number is part of the evidence. Which is why the case set has to be fixed and sealed before anybody runs anything, and why the honest write-up says how many cases were considered and how they were selected.

Step two — build the interpretation forks

At least a third of your case set should be forks: pairs of cases with near-identical surfaces and different underlying histories, requiring different correct actions.

Two clients ask for the same reporting capability. On the face of it the requests are interchangeable — same words, same shape, similar size, similar stated urgency. In one, the request is a symptom of a definitional disagreement between two divisions that has never been adjudicated, and building the thing will hard-code the disagreement into infrastructure and make it permanent. In the other, the definitions were settled two years ago in a governance forum nobody mentions any more, and the request is exactly what it appears to be. The correct action in the first case is to refuse the build and sell the adjudication. The correct action in the second is to build it, quickly and cheaply, and not be precious about it.

An arm that has read your published frameworks can produce a beautiful answer to either case. Only an arm that can tell them apart has discrimination. That is what a fork measures — not knowledge, but the resolution at which the world is being seen.

Step three — the four arms

Same frozen cases, four conditions, run independently and without sight of each other.

Two design notes that decide whether the result means anything.

Arm 2 gets a fair run. The most common way this test is rigged — usually unconsciously — is to hand another practitioner the kernel cold, give them an afternoon, and conclude from their mediocre output that the kernel is inseparable from its author. The evidence says operator tenure with a tool matters and compounds. If arm 1 has two years of driving practice and arm 2 has an afternoon, you have measured practice, not portability. Give arm 2 a real ramp and record how long it was, because "how much ramp does a competent stranger need" is itself one of the most useful numbers the test produces.

Nobody scores their own arm. Clinical trial reporting standards have spent decades formalising this and the vocabulary is worth borrowing wholesale — the CONSORT 2025 statement, published simultaneously across five major medical journals in April 2025, separates allocation concealment from the blinding of participants, care providers, outcome assessors and data analysts as four distinct things to get right.¹⁵ You do not need the full apparatus. You need the one discipline it exists to enforce: the person marking the answers must not know which arm produced them, or the founder's arm wins on reputation before anyone reads a word.

Step four — score four things, not one

Resist the urge to score "quality". Score the four verbs from the definition of discrimination, separately, because they come apart and where they come apart is the finding.

• Diagnosis. What did the arm conclude was actually going on? On a fork, this is where the two cases either separate or don't.
• Recommended action. What would they do, specifically enough to be executed by someone else on Monday? "Engage stakeholders" is not an action; it is a way of not having one.
• Calibration. How confident were they, and were they right to be? An arm that is right 70% of the time and says so is more valuable than an arm that is right 80% of the time and is certain every time. Confidence that does not track accuracy is the single most expensive property an adviser can have.
• Refusal and escalation quality. The one everybody leaves out, and often the one that decides the money. Did the arm recognise a case where the evidence was insufficient and say so? Did it identify what it would need in order to decide? An arm that confidently answers a question it should have refused has failed that case, no matter how good the answer reads.

Score against the sealed outcomes where an outcome exists, and against independent review where it does not — some of the most valuable cases are ones where the right call was to refuse, and refusal has no outcome to check.

The design principle underneath all of this is the one forecasting tournaments established: fix the questions in advance, resolve the outcomes later, and score everyone on the same rule. What makes that work is not the sophistication of the scoring but the fact that nobody gets to choose their questions after seeing how they did.

What honest reporting looks like

Hold the goal fixed, vary the condition, report shape rather than fake precision — this is the same method I have used before for measuring what a compiled worldview is worth, and its most important section is the list of forbidden metrics. For this test the forbidden list is: any percentage improvement without a stated denominator; any multiplier; any claim of transfer based on a case set you chose after seeing the results; any aggregate score that hides which arm won which fork. Say "on four of the six forks, arm 2 separated the pair correctly; on the two it missed, both were configuration-history cases" rather than "arm 2 scored 78%". The first sentence tells you what to build next. The second one tells you nothing and sounds better, which is why it is the one people write.

EightReading the result

Write this table down before you run anything. That is the entire point of it — a verdict you commit to in advance cannot be renegotiated by whoever is most disappointed on the day.

Translated into commercial terms, the whole table collapses into three outcomes and it is worth holding all three in view, because the third one is the one nobody names:

If the kernel materially improves decisions while reducing routine demands on me, I have built a successor asset. If clients still need me at the same density, I have augmented the founder. If they no longer need me but will not pay for the kernel, I have successfully disintermediated myself without completing the value migration.

That third outcome is a real and reachable place. It is what happens when you do the transfer work honestly and never build a commercial object around the thing you transferred into. You end up with a client who can operate beautifully without you and no reason to send you money — which is a better ethical position than the trap and a worse commercial one.

A court ran a version of this in 1987

If the ablation framing feels academic, consider how the key-person discount is actually adjudicated. The valuation body coursework that credentialed appraisers are taught from lists what a court has accepted: a 10% discount in one estate case, 10% again in another — and then a case where the discount was rejected, "since the decedent's sons were managing the company both before and after the father's death", with the court noting that one son "had been groomed for the position, and that succession planning had taken place".¹⁶

Read what made the difference there. Not that more had been documented. Not that the founder had been less capable. The discount vanished because another person was demonstrably operating the business before and after — an observed condition, in the world, with and without the key person. That is arm 2 of this test, run by circumstance and adjudicated by a court, four decades ago.

NineWhat I can claim, and what I can't

I have designed this test. I have not completed it. Everything above is a protocol, not a result, and I want that stated in the same breath as the protocol rather than buried at the end where nobody reads it.

This is the same posture the rest of my work in this area carries. The role-migration ladder is a designed programme with exit criteria, not a reported completed run. The succession metrics are proposed instruments with no observed production baselines. The month-of-absence trial is an acceptance-test design; no completed trial is reported. And the broader claim that a compiled kernel makes an ordinary bench materially more capable remains, by my own published admission, the unresolved proof — for the industry and for me. The ablation exists because that question is open, not because I know the answer and am staging a reveal.

The figures I went looking for and did not use

Researching this piece I went hunting for the number every reader wants: how much does founder dependence actually cost you at sale? There is a figure in wide circulation — that owner-dependent businesses transact at a discount of roughly one to two turns of EBITDA — attributed, when it is attributed at all, to the business-broker association's market survey. I downloaded that survey, extracted the full text and searched it for "owner", "depend" and "key person". It reports multiples by deal size and advisor sentiment. It contains no owner-dependence discount figure of any kind. The number is broker folk wisdom wearing a citation.

The same thing happened with a widely quoted statistic that entry-level roles in consulting fell from 41% to 26% of postings between 2018 and 2024, attributed to a named research institute. I downloaded that institute's report. The direction is stated in the text and I have quoted it above. Those specific figures are not in it; they appear to be readings taken off a chart and then repeated until they acquired the authority of a measurement. A related claim — that entry-level jobs in law, consulting and investment banking fell 35% since 2023 — I could not trace to any primary source at all. I have used neither.

And on the valuation side more broadly: I could not find a published standard from any international valuation standards body specifying a methodology or range for key-person discounts. What exists is authoritative recognition that the risk is real, a factor list appraisers are taught to weigh, and a professional-body admission that the impact "is often incorporated into the future benefit stream computations or the company-specific discount and/or capitalization rate development".¹⁶ That sentence is worth reading twice, because it says the discount is not computed — it is folded into judgment.

Which is, I think, a better finding than the number would have been. Founder dependence is real, it is priced, and nobody will show you the arithmetic. That is precisely the condition under which you should stop trying to argue about the discount and start producing the evidence that removes it — which, per the 1987 case above, is not documentation. It is a second person visibly operating the business.

Why the test has to be repeated, not run once

One more reason a single run is insufficient, and it is the strongest objection to this whole article: maybe the divergence is a stage rather than a trap. A founder might have to reach a higher altitude before there is anything worth transferring, in which case a period where capability outruns transferability is the healthy early phase of a good programme, not a pathology.

I think that is genuinely possible, and it is exactly why the instrument is a repeated measurement. A stage shows movement between runs — arm 2 closing on arm 1, the ramp shortening, the forks separating more often. A trap shows a flat line while the corpus grows. One run cannot tell them apart. Two runs, a few quarters apart, can. The anecdote never can, which is why the anecdote has been winning for years.

There is a discipline point hiding in that, and it belongs to the wider argument rather than to this piece: you cannot date the moment your client's substitute becomes good enough, so the only controllable clock is how fast you close questions with evidence. A repeated ablation is a falsification instrument pointed at your own asset base. Its value is not the reassurance it might provide; it is the rate at which it retires beliefs you would otherwise carry into a valuation conversation unexamined.

There is a second reason to repeat, and it is a moving target rather than a fixed one. Prepared machinery gets a disproportionate uplift when models improve — a casual user gets a better chatbot, an operator with accumulated kernels and evaluation machinery gets a lift to the whole production function. But the same release also improves the general model that arms 3 and 4 are using. So the quantity that actually matters is a net one: the dividend to your maintained kernel, minus the client's gain in substitution capability, minus the commoditisation of your outputs.

That is a bookkeeping identity, not a measured quantity, and I will not pretend otherwise. But it is a bookkeeping identity with a nasty implication: a static corpus can have a negative net position even while its interface gets better every year, because the public model keeps getting better at reconstructing the same answer while the corpus ages. Re-running the frozen cases after a model upgrade is the cheapest way to find out which side of zero you are on — and it costs almost nothing, because the cases are already built and sealed.

There is also a failure mode inside the kernel itself worth naming, because building a sharper one is not automatically progress. A kernel sharp enough to be a moat is also sharp about what it fails to value, and an optimiser pointed at a scoring function will find that function's weaknesses. Widening it dilutes the moat, because the kernel is the moat. The only thing that keeps that trade-off from running away is that proposals must meet deployed reality and what comes back must be allowed to change the evaluation function. Which is also why "updates from real outcomes" sits inside the definition of layer four rather than being an operational nicety attached to it.

TenWhat changes on Monday

The ablation is a quarter of work. Here is the part you can do this week, which will tell you whether the quarter is warranted.

Read the escalation log by shape

Pull the last two comparable windows of hard calls that reached you — two quarters, or two halves, whatever gives you an honest mix of ordinary work. For each one, tag two fields and nothing else:

• Shape: is this a class of case that has been resolved before, or genuinely new?
• Channel: did it arrive through a system anyone can see, or through your mobile?

Then compute one ratio: already-solved shapes, per hundred cases. Not the raw count — growth will flatter a raw count into meaninglessness. The rate.

Two rules make this honest rather than comforting. First, if you cannot see the informal channel, you cannot claim the number; a metric that ignores the mobile and the hallway is a vanity instrument, and the fastest way to improve a badly designed escalation metric is to drive the work underground. Second, do not invent a baseline. Your first honest window is the baseline. Any rate you reconstruct from memory to make the trend look better will destroy the credibility of the whole exercise the moment real instrumentation arrives.

What you are looking for is very specific. If the already-solved rate is falling across the two windows while the corpus grew, transfer is happening and this article is a warning you have already heeded. If it is flat or rising while the corpus grew, you have the trap, and you have it as a number rather than a feeling.

Then read the reason column on the repeats, because it tells you which repair you need. A repeat because the answer was never written down is an extraction problem. A repeat because the answer exists but nobody could find it at the moment of work is a findability problem. A repeat because somebody found it, read it, and escalated anyway is the interesting one: it means the written answer does not carry enough of the discrimination to be actionable by the person holding it, which is exactly the layer-3 failure the ablation is designed to confirm.

The question underneath all three repairs is the compounding test, and it is worth asking out loud at the end of every engagement: did we leave merely another deliverable, or a sharper corpus that makes the next piece of work easier and better? A loop that piles findings into chat history is activity. A loop whose outputs become substrate the next loop reads as priors is an asset. The escalation log is simply the cheapest available way to tell which one you have been running.

Pick six cases and seal them

While the log read is running, start building the case set. Six is enough to be informative and small enough to actually finish. Two of them should be forks. At least four should be cases whose outcomes never became public, for the leakage reason above.

The work here is unglamorous — reconstructing what was actually known on a given date is fiddly, and you will discover that some of your own decisions were made on less evidence than you remember. That discovery is free and it is worth the afternoon on its own.

Expect this to be the step that slips, and know why. Every other output of a project close pays out immediately — the invoice this month, the case study next quarter, the reusable component on the very next build. A record kept so that a later decision can be scored against it has no customer at all: no client asks for it, nobody is measured on it, and it produces nothing this quarter. So it will always be the thing you start properly next year. Put the six cases in someone's actual objectives, or accept that you have decided not to know.

Say what would change your mind

Before the test runs, write down the result that would make you abandon the position you currently hold. If you believe the kernel transfers, name the arm-2 result that would prove it doesn't. If you believe your questions are not the asset, name the arm-3 result that would prove they are. If you believe the client cannot do this themselves, name the arm-4 result that would settle it.

Put those in writing, dated, before anyone runs a case. Otherwise the result will be interpreted by whoever most wants a particular answer, and in a founder-led firm that person is always available and always in the room.

CloseWhat you can actually earn

I opened with the log that didn't change, and I should be clear about what I think it means. It does not mean the compiled corpus was a mistake. It was the best capital investment I have made, and the astonishment is genuine. It means the corpus did the job I asked of it — raise the altitude of what I can reliably take on — and did not do a different job I had quietly assumed came bundled.

Those two jobs feel the same from the inside because, before AI, they moved together. The only way to get more capability into a firm used to be more people getting better at things, which meant capability and transferability rose as one quantity. AI decouples them. Once decoupled, every instrument built to measure the joint quantity — revenue, utilisation, client satisfaction, corpus size, retrieval quality — reports the wrong answer with complete confidence.

So the claim I am not making is that your firm is doomed, and the claim I am not making is that I have solved this. The claim I am making is narrower and I think harder to escape: you do not currently know which of the four assets you own, the instruments you have cannot tell you, and there is a test that can.

And the strongest version of the claim you can eventually earn — the one worth building toward, and the one I have not yet earned — is conditional, three-part and testable. Terminal value survives when personal judgment has been compiled into a calibrated, portable evaluation capability: clients can operate yesterday's answer independently, ordinary practitioners can carry known cases without the founder, and the firm remains worth paying for because deployment keeps moving its frontier faster than copying can.

Three conditions. All of them observable. None of them satisfied by how impressive the founder has become.