The Learning Subsidy

Two firms close the first delivery of a brand-new AI-native offer in the same week. The engagements rhyme almost exactly: a fixed commercial envelope, two senior people, something like eight weeks, a client who is genuinely pleased and says so in writing. Same line at the bottom of both — contribution a long way under target, close enough to break-even that nobody wants to read the number aloud in the portfolio review.

One of those firms has just made an excellent investment. The other has just lost money.

Nothing in either firm’s management accounts can tell you which. Not the contribution line, not the utilisation report, not the satisfaction score, and certainly not the case study, which in both firms is already in draft and will be excellent.

The sentence that survives everything

Firm B has a sentence available to it, and the sentence is very hard to argue with. We were investing in learning. It is often true — it was true, in part, for Firm A as well. It is also the most durable piece of unaudited rhetoric in professional services, because no instrument has ever existed that could contradict it.

“We’re investing in learning” becomes the oldest consultancy excuse in the book for losing money.

Notice where that line comes from. It is not an outsider’s objection; it arrives from the same conversation that proposed the idea in the first place, one paragraph after proposing it. Any argument for a deliberate early loss that does not carry its own refutation this closely is not an argument, it is an appetite.

A verdict is not a control

The governing rule for all of this is already published, in the service architecture this book sits underneath.

A first engagement at break-even or a loss is rational if the loss is purchasing reusable capability. It is irrational if it is purchasing heroics.

That rule is correct, and on its own it cannot be used. Read it again as an instrument rather than as a sentence: it is a verdict. It is delivered after the fact, by whoever is in the room, about an engagement that has already been staffed, run and invoiced. It names no approver. It sets no limit. It carries no date. It requires no list.

Everything that makes a decision a control is missing from it. And that is not a criticism of the rule — the same chapter says so itself, and hands the job on:

How that subsidy is funded, accounted for and killed if it fails is a distinct piece of work with its own treatment.

This is that treatment.

Two ways to be wrong, and firms specialise

Without the instrument, a firm does not get one error. It gets whichever of two errors its culture prefers, repeatedly.

The first belongs to the finance partner. A capability investment is, from where they sit, indistinguishable from margin leakage — same line, same sign, same absence of an asset — so the safe move is to kill it, and killing it is nearly always defensible. The second belongs to the delivery partner. “Strategic account” is a phrase with no ceiling and no end date in it, and once it has been said about an engagement it can be said about the next one, and the one after that, until the firm is running a permanent subsidy that has never bought anything anybody can name.

Both errors are made by competent people acting reasonably on the information available, which is the tell that the problem is structural. You do not fix it with better judgement. You fix it with a record that makes the two cases look different before anyone has an interest in the answer.

What this is not

Five fences, and they are load-bearing rather than decorative, because every one of them is a thing a reader could otherwise take from this book and use badly.

• Not permission to under-deliver. The client-value ledger is drawn first and is never the source of the subsidy (Chapter 4).
• Not permission to use customers as undisclosed research subjects. The exchange is written down, in advance, in terms the client has agreed to.
• Not training on client data. What crosses is shape, not content — and there is a gate, with an owner (Chapter 6).
• Not a claim to reuse rights the contract does not grant. The rights position is a column in the ledger, not a hope held at renewal.
• Not ordinary discounting, loss-leader pricing, or land-and-expand. Those aim outward at a market position; this aims inward at your own cost curve (Chapter 13).

Part I builds the instrument. Part II runs it once on a single two-engagement cohort, every figure labelled as the illustrative parameter it is, then breaks it four ways. Part III holds the line where that is hardest: in the statutory accounts, in the promotion committee, and across a portfolio competing for one budget.

The claim underneath all of it is small enough for a Post-it and hard enough to have taken a book. Engagement one may buy the machine. Engagement two has to prove you bought one.

The last chapter ended on a claim about instruments. This one is a claim about arithmetic, and it is the reason the instrument has to be a second ledger rather than a better estimate.

An early engagement of a productised offer creates two returns. The first is the one the customer bought. The second is supplier capability return — something that makes future engagements cheaper, faster, safer, or possible at all.

Hold onto that phrasing, because most capability claims fail on it immediately. Ask a delivery lead what the engagement produced and you will get a list. Ask what the next engagement can now do that this one could not, and the list gets very short, very quickly — which is useful, because the short version is the only one worth funding.

So why is it always the second return that vanishes?

Because a contribution line has room for exactly one, and the one with an invoice attached wins every time. But the deeper reason is not arithmetic at all. It is ownership.

The first question — did we keep the customer’s promise? — has an owner in every firm on earth: the engagement lead, the account partner, the delivery manager, someone whose bonus depends on it. The second question has no owner at all. As the architecture above this book puts it, it is nobody’s deliverable, it appears in nobody’s utilisation, and it is therefore the first thing to disappear when a project runs late.

An unowned return does not get measured badly. It gets claimed — and a claim cannot be capped, audited or killed, which is precisely why claims survive so well.

Capital formation, and the concession that buys it

Name what is actually happening when a first engagement builds an exception taxonomy and an acceptance harness inside the delivery envelope. That is capital formation in the delivery system. You are applying project accounting to something that behaves partly like R&D or platform investment — and it is worth stating the concession in the same breath rather than waiting to be caught by it: not necessarily capitalised as an accounting asset. The word “investment” is being used economically.

That concession is not a retreat. It is the reason the rest of the book exists, and Chapter 11 shows that the accounting standards agree with this book more thoroughly than a finance partner might expect. For now, the point is narrower: if the second return is real and the statutory accounts will never carry it, then the firm needs a record that will.

The bench changes the size, not the structure

The argument is usually made at scale, and the scale version is genuinely compelling.

You’ve got a 150-consultant bench, and you sell a fixed-price project with one or two consultants in it for six to eight weeks. You can tolerate a marginal gain or loss, because it’s building the product the rest of the team is going to use.

Correct, and it invites a conclusion that has to be blocked now rather than later: that this is enterprise doctrine, and a small firm can ignore it. The opposite is true. A large bench changes the size of the second return, not the structure of the decision. Two consultants spending six weeks on something 150 people will load is a spectacular multiple; the same six weeks in a four-person firm is a smaller multiple and a much larger share of everything the firm has.

A four-person practice has less room for an unappropriated loss than a global one, not more. It gets one shot at the subsidy, it will feel every point of it, and it cannot absorb a bad outcome inside a portfolio. Everything in this book applies at four people. Only the arithmetic scales.

A subsidy denominated in hours saved can never be audited

There is a standing temptation to justify the second return in the currency the firm already uses: hours saved. Refuse it, for a mechanical reason rather than a stylistic one. A subsidy denominated in hours saved cannot be audited, because the denominator is the thing being consumed — every hour removed appears as good news twice, once as efficiency and once as released capacity, and never once as anything the firm can point at afterwards.

Count what the engagement adds in kind instead: decisions that got better because the context was finally in the room; institutional knowledge recovered; questions answered that nobody previously knew to ask; the density of the substrate; reuse across futures that did not exist when the asset was built. That is a capability-denominated return, and its rows survive a sceptical finance partner precisely because none of them can be inflated by a self-reported baseline.

The industry is counting the wrong thing, and knows it

For the argument. A sector running at roughly a third of consultant time unbilled which still describes capability-building as unaffordable does not have a capacity problem. It has a denomination problem: utilisation counts hours sold, not assets built, so the unbilled third registers as failure rather than as available investment. And look at the distance between 37.7% project margin and 9.9% firm EBITDA. Everything between those two numbers is overhead — which is exactly where reusable-asset work lives, and exactly what gets cut first.

Against it, and this matters more. At around ten per cent EBITDA, an uncapped subsidy is genuinely dangerous. A firm with that margin structure cannot absorb an open-ended investment thesis, and a finance partner who resists one is not being small-minded. That is the whole reason the next chapter exists, and why the ceiling is not an optional field.

There is a second reason this argument has become urgent rather than merely interesting. When cheap cognition starts compressing the unit the invoice counts, productivity applied to that unit accelerates its depreciation without giving the firm ownership of whatever replaces it — the case made in detail elsewhere in this series. Building the successor unit’s machinery is therefore worth paying for. What it is not worth is paying for it by accident.

That is the sentence a CFO has to either accept or refuse. Accepting it does not commit anyone to capitalising anything; it commits them to the view that one line item contained two different kinds of spending. Refusing it commits them to the view that everything a first engagement spends is delivery cost, including the harness that will run on every engagement afterwards.

The rest of this book is the machinery that makes refusing it hard — and, just as importantly, the machinery that makes accepting it safe. Because the second return cannot appropriate itself. Somebody has to do that, in advance, on one page.

Here is the page. It takes about an hour, it is boring to fill in, and every field on it is useless if it is completed after the engagement rather than before. That is not a procedural nicety — it is the entire mechanism, and everything else in this chapter is a consequence of it.

The omission column is doing the work. The shape is borrowed deliberately from the anatomy of a commercial falsifier — observation, window, caller, accepted results, all four or it is decoration — because a subsidy is a bet, and a bet without a stated losing condition is a preference.

The expiry, and where cohorts come from

Expiry is the field most often left blank, and the one with the sharpest origin. The original formulation bounds it in the same breath as it proposes the loss:

It’s not like we’re going to run at a loss for five years, or two, or even one. A loss or a minimal gain on the first two engagements — maybe per industry.

Two things are hiding in that sentence. The first is the bound: the subsidy has a horizon, and the horizon is short. The second is “maybe per industry”, which is where the subsidy cohort comes from — the unit of appropriation is not one engagement, it is a small set of engagements sharing one offer, one band and one repayment test. Chapter 13 supplies the depreciation evidence that makes a short horizon a requirement rather than a preference; here it is enough that the field exists and has a value in it.

Why the approver and the assessor are different people

This looks like governance theatre and is the single best-evidenced design choice in the chapter.

Escalation of commitment — continuing to fund a losing course of action — is described in the management literature as one of the most robust and costly decision errors in the organisational sciences, and the meta-analytic work finds that one of its most powerful drivers is whether the decision maker faces a strong ego threat, with time already invested among the strongest antecedents.² A subsidy is, by construction, a decision with a named owner’s judgement attached and a growing pile of invested time behind it. Asking that owner to mark their own receipt is asking them to do the one thing the evidence says they will find hardest.

The same work carries a finding that argues against the obvious alternative: sharing the decision authority makes escalation worse, not better. A committee is not a control. One named approver, and a separate assessor with no authorship of the decision, is.

The structural half of this has been known in product development for decades. Stage-gate practice exists because, in Cooper and Edgett’s blunt formulation, “once a project begins, there is very little chance that it will ever be killed”, and gates with genuine Go/Kill authority are strongly correlated with the profitability of new-product efforts.³ Cooper’s own phrase for the fix is worth stealing outright: gates with teeth, or investment decision points. That is exactly what a subsidy approval record is — a resource commitment approved, a ceiling set, and a dated next gate agreed.

“But you can’t know in advance what you’ll learn”

This is the objection that arrives first in every room, and it is half right. You cannot predict what you will discover. You can absolutely name the class of artefact the engagement must leave behind, and the test each one has to pass.

Take the pair concretely. “A typed exception class, with a detection condition and a route” is a class: it can be demanded in advance, and its acceptance test can be written before anyone knows what the exception will be — the condition fires automatically on a case nobody flagged by hand. “The source-disagreement class” is a discovery: nobody could have named it, and nobody needed to. The record commits to the first and stays silent about the second. An eval harness is foreseeable; which cases go into it is not.

That is the chapter’s real argument, and it is worth being specific about the mechanism. A team that has pre-committed to leaving a taxonomy behind logs exceptions as they occur, because the log is a deliverable rather than an afterthought. A team that has pre-committed to a harness builds it during the work, when the cases are live, instead of reconstructing it in the final week from memory. A team that has pre-committed to a corrected band driver records the driver at the moment it turns out to be wrong, rather than remembering afterwards that the band was roughly fine.

None of those behaviours are expensive. All of them are nearly impossible to retrofit. The record is a design instrument aimed at engagement one, wearing the clothes of a finance control.

Retrospective relabelling

Three noes and it is an overrun. This is stated without hedging because hedging it is how the instrument dies: any version of the rule containing the words “in practice, of course” can be applied retrospectively to anything. Chapter 10 works the counterfeit; Chapter 11 shows that the accounting standards forbid the identical move in almost the same words.

Not a typed reserve

One distinction, because the two instruments look alike on a dashboard. A typed reserve absorbs pre-declared surprise inside the perimeter — access delayed, exception density above the band’s assumption — drawing down visibly against a published rule, decided by the delivery lead, logged rather than negotiated. A learning subsidy is a pre-approved shortfall in firm contribution, aimed at the firm’s own future cost curve, decided by a named approver before the engagement begins.

Different meter, different owner, different payoff. A reserve protects the promise; a subsidy buys the machine. Confusing them produces the worst of both: a reserve nobody can see, funding assets nobody named.

A thesis that cannot fail is not a thesis. The ceiling, the expiry and the kill conditions are there so that this one can — and the next two chapters are the two sides the approved amount gets spent against. The order in which they are drawn turns out to be a control in its own right.

The approval record names an amount. This chapter and the next are the two sides it gets spent against, and the first thing to settle is not what goes on them but which one is drawn first.

The client-value ledger is drawn first. Always. It is drawn before anyone opens the capability side, before the subsidy is reconciled, and before anybody says the words “on balance”. That sequence is not tidiness. It is the mechanism that keeps the subsidy paid out of firm contribution rather than out of client outcome — because a firm that draws the capability side first will, under margin pressure, find the difference on the client side.

Five rows

The third column is where the argument lives. Each of those failures is a subsidy taken from the wrong party, and none of them looks like under-delivery from inside the firm.

Three ways the client side gets shaved

An exception granted rather than resolved. Something in the client’s environment does not fit the offer. Resolving it costs the firm a week; granting an exception costs nothing today and moves the consequence onto the client’s side of the line, where it will be discovered in six months by somebody else. The ledger’s requirement — record the exception and what it cost — is what makes the choice visible while it is still a choice.

Evidence assembled at the end. Producing evidence as the work goes is more expensive for the supplier and much more valuable to the buyer, because it can be checked against the work rather than against a memory of it. Assembling it in the final week is cheaper and produces a document. The client bought the first thing and, in the absence of a row that distinguishes them, routinely receives the second.

An obligation carried out with no date. Something remains open at close. If it has an owner and a date, it is a commitment. If it has neither, it is a debt the client is now holding without having agreed to hold it — and it will be settled, eventually, at the client’s cost or at the relationship’s.

What the acceptance row records is the acceptance that actually happened: a complete evidence chain, and no negative observation triggered. The semantics of that belong to the verification work alongside this book and are not re-derived here.

The floor this whole instrument stands on

The customer receives useful delivery now, with explicit learning rights and clear IP and de-identification boundaries. Anything else is undisclosed vendor research funded by somebody who thought they were buying a service. That is the floor. Chapter 6 prices what happens when an asset cannot legitimately cross it; this chapter simply refuses to let the floor move.

“Every engagement has some give and take”

It does, and the instrument does not forbid generosity. It forbids funding the firm’s learning out of the client’s outcome. Give-and-take that is recorded, costed and inside the promise is ordinary professional practice and always has been. The same give-and-take, unrecorded, is the exact mechanism by which a subsidy quietly changes payer — and the only difference between the two is a row in a table.

This is the shortest chapter in the book, and that is deliberate rather than neglectful. The ledger is five rows because a longer one invites negotiation, and the rule admits no elaboration: a subsidy that touches the client’s outcome is not a subsidy at all.

The undiscountable side is now drawn. The other side is the one that has to be earned, and it needs six columns rather than five — because every line on it is a claim the firm is making about itself.

Every line on this side is a claim the firm is making about itself, so the ledger opens with the question those claims have to answer. Not what did we make? but:

What can engagement N+1 now do that engagement N could not?

A capability ledger is not an inventory. It is a set of dated claims, each with a named owner, a rights position and a test that could fail.

Six columns

The rights column sits third rather than last on purpose. If the answer there is refused, the rest of the row is a cost rather than an asset, and the ledger should discover that early — which is Chapter 6’s subject.

The owner column is not administrative either. “The team” cannot hold a rights position, cannot answer for a reuse hypothesis, and cannot be asked in nine months whether the thing was ever loaded. The discipline is inherited: disposed means terminated in exactly one disposition, with a named owner and a date — and “maybe later” is not a disposition, it is a deferred decision that must carry a revisit trigger.

Six classes, each with a test that could fail

Two items on the source list are the ones firms most reliably lose, and both belong here: the rejected architecture and why it failed, and the marketing claim that is now supported by evidence. The first is class six. The second usually has no home at all, which is why the marketing team keeps asking delivery for anecdotes.

One acceptance condition

Every class above has its own test, and all of them are instances of one rule, inherited from the compounding work: storing something is insufficient; it compounds only when the residue changes the starting state of the next run.

That bar is higher than “it exists” and lower than “it saved money”, and it has the great advantage of being checkable by somebody who was not there. It also disqualifies, at a stroke, most of what firms currently call reusable: the folder, the prompt library, the deck of worked examples, the well-attended internal briefing. None of those are loaded by the next run. They are waiting to be remembered, which is a different thing and a much weaker one.

Promised and accepted are two columns

The approval record named a list. The ledger marks what actually landed, and the gap between them is not a failure to be explained away — it is the most honest signal the firm produces about its own delivery. Three shapes, and each says something different.

• Promised and missing. The work went somewhere else. Usually this is information about the offer’s real cost drivers rather than about the team — the harness did not get built because the exception load was three times what the band assumed, which is itself worth knowing.
• Accepted but unplanned. Serendipity. The ledger must have a line for it, or the instrument punishes the thing it exists to encourage.
• Promised and accepted, matching every time. The most dangerous shape, and the one that looks best in a portfolio pack. A firm whose ratio is always 1.0 is promising only what it had already built, and the subsidy is buying nothing it did not have.

What if the best asset was the one nobody predicted?

This is the strongest objection to the whole instrument, and it deserves a real answer rather than a reassurance. The claim is that rigid pre-declaration will kill the best learning, because the best learning is always the learning nobody saw coming.

True of the prediction. False of the ledger. An unplanned asset that is nominated, disposed and owned sits on the ledger exactly like any other — it simply arrives in the “accepted but unplanned” column and is marked there. What the instrument refuses is the unplanned asset that exists only as a story. Serendipity survives. Anonymity does not.

There is a related worry about administrative weight, and the honest answer is that this is six columns and, in practice, four to seven rows. If a delivery lead cannot list the engagement’s promotable assets and their owners inside an hour, that is not evidence the ledger is too heavy. It is evidence the capability ledger does not exist — and that is the finding, not the overhead.

Three columns into every row, the ledger asks whether the asset may cross at all. That question has an owner, a gate — and, unlike anywhere else it has been asked, a price.

The asset that fails is rarely the trivial one. It is the expensive one: three days of a principal’s time spent working out how one obligation interacts with another, in one jurisdiction, for one class of entity. Genuinely clever, genuinely hard-won — and specific, which is exactly what the gate refuses.

The gate itself is not this book’s to rebuild. Three territories — client truth, engagement learning, promotable capability. One gate with five conditions in order: de-identification, abstraction, transferability, clearance, human approval; a candidate that fails any one of them does not cross. Five dispositions, exactly one of them mandatory per candidate.

What that architecture did not have to do — because it was answering a different question — is price a refusal. That is this chapter.

A refusal is a line, not an absence

Walk the sequence, because the money moves in a place most firms never look. The asset was named in the approval record. Part of the subsidy was approved on its account — the ceiling was set in the expectation that this thing would exist and would be usable. The engagement produced it. It reaches the gate. It fails.

At that moment three things happen simultaneously, and only the first is normally recorded. The capability row is marked refused, with the reason preserved. The learning stays where it was made, which is correct. And the corresponding subsidy line is marked unrepaid — permanently. The money does not come back, and no later engagement will repay it, because there is nothing there to load.

Recording the refusal as a cost is not bookkeeping fastidiousness. It is the only thing that keeps the gate honest in a bad quarter, because under margin pressure the cheapest way to make a subsidy look repaid is to loosen the gate slightly — and a ledger that shows refusals makes the loosening visible as an anomaly rather than invisible as a mood. A capability ledger that has never once shown a refusal is not running a gate. It is being written to look good.

The clause most firms are relying on does not cover this

There is a legal edge here that almost nobody in delivery has noticed, and it inverts the intuition completely.

Most services contracts carry a residuals clause, and most firms assume it is the thing that lets them reuse what they learn. Residuals clauses permit a party to use information retained in its employees’ unaided memory, even where that information was disclosed in confidence. But the scope of use, as practitioners set it out, “generally excludes tangible materials.”⁴

Read that against the previous chapter. Everything the capability ledger demands — a written taxonomy, a runnable harness, a decision tree, a regression test — is a tangible material. The moment learning becomes reusable in the sense this book means, it steps outside the clause the firm was quietly relying on.

That is worth sitting with, because it explains a pathology that has always looked like culture and turns out to be partly legal architecture. The person who keeps the knowledge in their head is on the safest contractual ground available. The person who writes it down, de-identifies it and makes it loadable has created an artefact that requires an express right. The legally easiest path is the economically worst one, and no amount of encouragement fixes an incentive that is structured that way.

The fix is in the ledger’s third column. Rights positions have vocabulary already: background IP is what existed beforehand and is needed to perform the work; foreground is what the work creates. The observed pathology is that most agreements treat this as a tick-box rather than a commercial decision — clauses become theoretical, and nobody discovers the answer until a renewal.⁵ Naming the asset before the engagement forces the question into the negotiation, where it costs a conversation instead of an asset. Where the mechanism fits, a feedback licence is often the cleaner instrument: it grants use of what was submitted rather than depending on what someone remembers.

And the honesty test that sits underneath all of it: if your compounding genuinely requires the client’s confidential content, say so out loud, to yourself first. You have a data business rather than a service architecture, and it should be priced, contracted and governed as one. In this book’s terms the consequence is sharper still — the subsidy is buying something you cannot own, and the ledger will never balance.

The bargain, and the version of it that is actually documented

The analogy people reach for is the inference intermediary that exchanges compute for improvement data, and it is worth stating precisely rather than loosely, because the loose version asserts things about companies that cannot be sourced.

The documented instance is Google’s Gemini API, where the exchange appears as a row in the pricing table itself — “Used to improve our products”: yes on the free tier, no on the paid tier — and the terms confirm it, adding an instruction that is the whole point: “Do not submit sensitive, confidential, or personal information to the Unpaid Services.”⁶

Three observations, and then the analogy stops earning its keep. The provider prices the learning right — it is worth the entire fee. The provider refuses the confidentiality risk rather than managing it, which a services firm cannot do; a firm has to run a membrane instead, and that is strictly harder work than a disclaimer. And the provider’s learning goes into weights, where it is not inspectable, not reversible and not attributable, while a firm’s goes into an external substrate where it is all three. Same shape of bargain. Entirely different substrate, and the difference is the ethics.

“Our clients would be uncomfortable if they knew”

They would be uncomfortable if the learning were undisclosed, unbounded, or paid for out of their outcome. None of those is what this is. The client-value ledger is drawn first and cannot be discounted; the gate governs what may cross and who approves it; refusals are recorded. The client conversation has a shape, and it is short: what stays, what may leave and in what form, who approves it, what the client gets in return, and what happens to refusals. The wording is the lawyers’ job; the structure is the architect’s.

A firm that cannot say those five things out loud to a client is not running a subsidy. It is running an extraction — and a firm that cannot show you its refusals has a substrate that is accumulating by accident.

The ledger can now name an asset, own it, check its rights and price its refusal. What it still cannot do is prove any of it — and proof, it turns out, has a staffing requirement.

Somebody is allocating people to the second sale of the offer. The client asked for the two who did the first one. They happen to be between engagements. They know the product. Every reason is good, and none of them is about the subsidy — and the decision has just determined whether anything in the previous six chapters can be tested.

Transfer means ordinary capable staff lead materially more of engagement two because shared infrastructure changed, not because the same heroes worked late again. Put the originating pair back on it and the comparison measures their personal learning curve — which is real, and which leaves when they do.

If engagement two still requires the same heroes at the same density, you didn’t invest in learning. You subsidised the customer.

Note the counterparty in that sentence. Not wasted — transferred, to a party who did not ask for it and would not have paid for it. The firm made a gift, and then wrote a case study about the gift.

Eight rows, and what moves each one

Two of those rows are marked because they are the seam. The published measurement set for a service flywheel — estimate error, production variance absorbed, human disposition load, exception classes, acceptance evidence, reusable artefacts, recurring escalation rate, scarce-expert density — asks whether the flywheel is turning. Cost-to-serve and delivery margin ask whether the appropriation was repaid.

Those are not the same question, and conflating them is how a firm ends up pleased with an offer that is still underwater. A flywheel can be turning while the subsidy remains unpaid: the recurring parts got cheaper, but not by enough, or not yet, or not in the places the ledger named. Point at the flywheel table for mechanism; keep the two money rows for the receipt.

The row that makes firms quit

Engagement two puts a disposition surface in front of the team, and suddenly there are forty of them with names attached. At that moment the firm faces a choice with nothing technical in it: accept that the number rose because the measurement got honest, or conclude that the new way of working “created overhead” and quietly stop counting. Firms that pick the second option choose narrative over control — and having chosen it once, they never get a real number again.

The second shift

There is a piece of outside evidence for the staffing rule that is better than anything this corpus could generate, and it arrived by accident.

Economists studying an automobile assembly plant found that each ten-fold increase in cumulative production halved average defect rates — an ordinary learning-curve result. The interesting part came when the plant started a second shift, staffed with different workers. The learning process did not restart. “Instead, the shift actually begins at average defect rates that are below the first shift’s rates.”⁷

The researchers’ conclusion is the engagement-two rule written in economics: “much of what is learned in the plant becomes embodied very quickly in the physical or broader organizational capital of the plant, rather than remaining only with workers.”

Do not over-read the analogy — a car plant is not a consultancy, and the magnitudes do not transfer. What transfers is the identification strategy: change the humans and see whether the improvement survives. It is cheap to run, which means declining to run it is a decision rather than an oversight. And note the study’s own control — no major production-technology change during the period observed. That is the same control a two-engagement comparison needs, and the one firms break most often.

Four controls, named here and used in Chapter 8

• Same offer, same band. Comparing a small engagement with a large one and calling the difference learning is the oldest way to fake this.
• Different client. A second engagement with the same client compounds relationship knowledge, not machinery, and the two are easy to confuse.
• No model or tooling generation change. This is the one firms break silently. A materially better model arrives between the two engagements, everything improves, and the improvement is attributed to the substrate.
• Ordinary capable staff. Discussed above, and it is the control most often broken for the best reasons.

The quantitative honesty check on all of this already exists: paid bounded units divided by scarce expert dispositions, with strict discipline on both numerator and denominator. It says whether leverage improved. The receipt asks something narrower and harder: whether this appropriation, of this size, against these named assets, was repaid. Use both.

And be clear about what the receipt is not. It is not a satisfaction score, not a case study, and not a retrospective. A retrospective produces a document; a disposition produces a change the next run reads without being asked. If the honest answer at the end of an engagement is “the team learned a lot”, you held a meeting.

One objection has been circling since Chapter 2 and deserves a straight answer here, even though its evidence arrives later. You have just described a learning curve; every business has one. True — but a learning curve lives in people and leaves when they do, while this lives in artefacts an ordinary consultant loads on day one. The test that separates them is the staffing assumption above. Chapter 13 supplies the evidence on how weak, variable and confounded service learning curves actually are, which is the reason the receipt has to be measured rather than assumed.

Part I has now specified an instrument: an approval record, two ledgers, a price for refusal, and a receipt. None of it has been run. The next two chapters run it once, on one cohort, with every figure labelled for exactly what it is.

The setting is generic on purpose. A first fixed-commitment engagement of a new AI-native offer: a bounded decision product for a mid-sized knowledge business, priced in a band, delivered inside a fixed commercial envelope, two senior people, roughly eight weeks.

Every magnitude below is an index number. Target contribution = 100. No currency appears anywhere in this chapter or the next, because the moment an index point looks like a dollar, a reader starts benchmarking against their own firm and the model stops being a model.

The approval record, filled in

The three kill conditions, written before the work:

1. By week four, no exception has been typed with a detection condition. The taxonomy is not being built; it is being intended.
2. At the point of first acceptance, the harness exists only as a document.
3. Consumed effort passes the ceiling before the promise is met — at which point the shortfall has stopped buying assets and started buying delivery.

What makes those usable is not their wording. It is that each is a lookup rather than an argument, and each can fire while there is still something to decide. A kill condition that can only be evaluated after the engagement closes is a post-mortem finding wearing a governance costume.

Five assets, five tests

What the forty-five points actually paid for

The subsidy is not consumed by one thing. In this cohort it goes three ways, and the split is the point rather than the proportions.

Now the observation this chapter exists to earn. Only the third of those is buying anything. The first two are the cost of finding out — real, unavoidable, and not an asset by anybody’s definition. The third is the cost of not having to find out again.

That distinction is why a subsidy is not simply a discount with a story. A firm that cannot separate the three at all — that cannot say which portion of a shortfall bought something loadable — is running an overrun, whatever its record says. And a firm whose asset-construction share is close to zero has just discovered something important about its first engagement while there is still time to act on it.

The client-value ledger, drawn first and clean

The subsidy did not touch this side. That sentence is only worth writing because the two granted exceptions are recorded with their cost — which is what makes it checkable rather than reassuring. A client-value ledger with no exceptions on it is either an unusually clean engagement or an unusually forgetful one, and the reader cannot tell which.

Nomination is not promotion

At close, six candidates are nominated: the five planned, plus one that nobody predicted. A principal answered a novel architecture question and, this time, wrote down a de-identified decision tree, a mandatory evidence list and a regression test alongside the answer.

Recurrence nominates. Humans promote. None of the six is promoted in this chapter, and that restraint is the chapter’s discipline rather than an omission. Engagement one can only nominate. Promotion is a gate decision, and repayment is a receipt, and both of those belong to engagement two.

One row that gives the specimen away

Senior disposition load in engagement one is recorded as not instrumented. Not a low number — no number.

That is the honest entry, and it is also the tell that separates a specimen from a brochure. Every first engagement of a new offer has an unmeasured disposition load, because the surface that would have measured it did not exist yet. A model that reports a flattering figure for something nobody counted has invented data in the one place a sceptical reader will check. The next chapter shows what happens when it is measured for the first time, and why the result surprises people in precisely the wrong direction.

So the assessor marks the record: five of five planned assets nominated, one extra, both ledgers drawn, subsidy consumed at 48 points against a ceiling of 60. Everything in that sentence is a claim. Nothing in it is yet true.

The next chapter changes the people, runs the offer again, and marks it.

The previous chapter ended with claims. This one begins with a staffing sheet, because the receipt is decided by the design before it is decided by the work. The same illustrative-model banner applies throughout: index numbers, stated assumptions, no measured client data.

The cohort ledger

Every row has a mechanism beside it because numbers without a mechanism are numerology. A reader who disagrees with a figure can still test the mechanism, and the mechanism is the transferable part.

The asset that does not cross

Asset A6 — the unplanned one, the de-identified decision tree — passed easily. The one that failed was A4’s neighbour: a client-specific regulatory exception worked out during the first engagement. A rule about how one obligation interacts with a second, in one jurisdiction, for one class of entity. Most of three days of a principal’s time. Genuinely clever, and genuinely the most sophisticated single thing the engagement produced.

A candidate that fails any one condition does not cross, and this one failed the second. Note what did not happen: nobody re-wrote it at a higher level of abstraction until it passed. That move is available in every firm, it takes an afternoon, and it produces a pattern that is either attributable or untrue.

The arithmetic of the receipt

Forty-eight points were drawn. Engagement two came in at 88 against a target of 100, so the visible movement is twelve points of contribution recovered relative to the first engagement’s 52. Twelve is not forty-eight, and the difference is the whole reason this section exists.

Take the repaid column first, because it contains the trap. Only twelve of those twenty-two points appear in this engagement’s contribution line. The other ten are senior hours that were not spent here — the principal who took four escalations instead of nineteen spent the difference somewhere else in the firm, and that value shows up on a different offer’s P&L or on no P&L at all. A subsidy ledger that only reads the originating offer’s contribution line will systematically under-count its own repayment, which is a strange way for an instrument to fail and a common one.

The carried column is an estimate about engagements that have not happened, and it should be written as one. Fourteen points is the assessor’s judgement that a taxonomy and a routing rule will still be loading on engagement four. If the offer is demoted before then, that fourteen becomes a write-off, retrospectively — and the ledger should be able to say so without anyone feeling they have been caught out.

The written-off column is the honest one. The refused asset is gone for good. So is the unrecoverable share of unpriced discovery, which was never an asset in anybody’s accounting — a point Chapter 11 shows the standards make far more bluntly than this book does.

Twenty-two, fourteen and twelve sum to forty-eight, but only because two of the three are estimates and one is a judgement about the future. A repayment schedule that balances exactly is usually a schedule that was fitted to the answer. The useful discipline is not arithmetic tidiness — it is that each column is marked with what kind of number it is: observed, estimated, or written off.

What did not move, and what each flat row means

Authority-dependent dispositions did not fall, and are not supposed to. If that row ever collapses towards zero, the correct response is alarm rather than celebration — something consequential is being decided without anybody owning it.

Two genuinely novel exception classes appeared, from a constraint surface the first client did not have. That is not the flywheel failing. That is the flywheel doing its actual job: making the recurring parts cheap so that scarce attention lands on the new edge.

Anything resting on a principal’s relationship transferred not at all. The parts of engagement one that went smoothly because somebody had credibility with a particular executive had no equivalent in engagement two, and no artefact could have carried them. That is the cleanest available demonstration of the difference between a substrate and a hero.

The preserved rejection is invisible in every row. Nobody spent a week on an architecture that had already failed, and a week not spent appears nowhere in any measurement set ever devised — which is precisely why rejections have to be preserved deliberately rather than expected to justify themselves.

Forty dispositions, and the moment of choice

Engagement one: not instrumented. Engagement two: forty dispositions with names attached. A firm reading its own ledger for the first time sees that as the new way of working creating overhead. It is the measurement getting honest, and the assessor’s entry against that row is the important one — it is recorded as a baseline, not as an improvement or a deterioration.

Engagement two often produces the first real number rather than a better one, and a firm that cannot record a baseline without arguing about it will never get a second data point.

The verdict, written as a disposition

Subsidy substantially repaid. One line written off. Expiry not reached. Ordinary staff led. The offer moves from first client to second-engagement transfer on the lifecycle, and the next decision the portfolio faces is scale, repair or demote — where demotion counts as a success whenever it stops a zombie product consuming senior time forever.

And the counterfactual, in one sentence, because it is the most useful thing in the chapter: had the originating pair run engagement two, every row above could have moved in the right direction and the verdict would still read fail — because the control was broken and the comparison would have measured two people rather than a substrate. Chapter 10 works that case, along with three others.

What the assessor actually writes is four lines, and they are worth reproducing because they are the whole output of the instrument: subsidy 48 of a 60 ceiling; repaid 22 observed, 14 carried to engagements three and four, 12 written off; one asset refused at the abstraction condition and recorded as unrepayable; delivered by ordinary staff, so the receipt stands. That is a paragraph a finance partner can argue with, which is more than any case study has ever offered them.

One last thing, stated plainly at the close. n = 2 is not a portfolio. This chapter shows what you would have to record; it establishes nothing at all about your offer. One pair of engagements cannot establish a curve, and Chapter 13 has the evidence on how thoroughly that is true.

Rewind to the start of engagement one and let it fail. Four ways, each of which happens constantly, each of which produces a defensible-sounding portfolio entry, and all four of which share exactly one property: nothing in them could have failed. That is the detection rule, and it is why the checklist at the end of this chapter is six questions where no is decisive.

The third ledger, denominated in people

There is a fifth failure that is not a counterfeit subsidy so much as a hidden one. A first engagement can come in green because staff paid for it personally: nights and weekends normalised, more people pulled into each issue, review notes getting shorter, unresolved questions carried forward, and a growing reliance on one or two load-bearing specialists. That is green by heroics — status produced by discretionary human effort the design did not fund.

The point for this book is narrow and sharp. There is a third ledger, it is denominated in people, and almost nobody keeps it. A subsidy repaid out of staff resilience has not been repaid. It has been refinanced, at a rate nobody wrote down, from a lender who will eventually leave.

Part II has now run the instrument once and broken it four ways. Part III holds the line in the three places where holding it is hardest: in the statutory accounts, in the promotion committee, and across a portfolio of offers competing for one budget.

It takes a finance partner about ninety seconds to arrive at the only question that matters from where they sit: does any of this go on the balance sheet?

No. Almost certainly not, and the more genuinely novel the learning, the more certainly not. Concede it immediately and completely, because the concession is what makes everything else credible — and because this was never a capitalisation argument. It is a management control, and its entire value is that it can be exceeded, be late, and fail. No narrative can do any of those.

IAS 38, walked

Research-phase expenditure is expensed as it is incurred, with no discretion: “No intangible asset arising from research (or from the research phase of an internal project) shall be recognised.”⁸ Development expenditure can be capitalised, but only when six criteria are met simultaneously — technical feasibility, intention, ability to use or sell, probable future benefits, adequate resources, and reliable measurement of the attributable cost.

Then the paragraph that settles the question for a first engagement before the argument starts: “If an entity cannot distinguish the research phase from the development phase of an internal project to create an intangible asset, the entity treats the expenditure on that project as if it were incurred in the research phase only.”

Read that against the thing a generative first engagement actually is. Discovery and construction are interleaved by design — that is what makes the interior productive and what makes the band hard to set. By default, therefore, all of it is research, and the burden of separating the phases sits on the firm. A firm that cannot carry that burden has the answer decided for it.

The control test is the rights gate, wearing accounting clothes

Here is the part that ought to change how a practice leader feels about the whole exercise. IAS 38 says an entity “usually has insufficient control over the expected future economic benefits arising from a team of skilled staff and from training” for those benefits to meet the definition of an asset. Better people are not an asset. Neither is a smarter team, a wiser principal, or a delivery group that now “knows the domain”.

But the same standard names the condition under which knowledge is controlled: an entity controls the benefits “if, for example, the knowledge is protected by legal rights such as copyrights, a restraint of trade agreement (where permitted) or by a legal duty on employees to maintain confidentiality.”

That convergence is not a coincidence, and it is worth using in the room. The rights column in the capability ledger, the refusal that costs money in Chapter 6, and the demand that an asset be named rather than described are all instances of the same underlying question the standard is asking — can you actually control the benefit, or do you merely hope to?

The relabel is not merely disreputable. It is impossible.

Two paragraphs finish the argument that started in Chapter 1. First, once expensed, always expensed: “Expenditure on an intangible item that was initially recognised as an expense shall not be recognised as part of the cost of an intangible asset at a later date.” Second, on what can never be part of an asset’s cost at all:

…identified inefficiencies and initial operating losses incurred before the asset achieves planned performance.

An overrun is an overrun. The standard-setters got to the book’s sharpest edge first, and they got there for the same reason: because the alternative is a rule that can be applied retrospectively to anything.

The US position, and a phrase worth stealing

The American treatment is moving, and not in the firm’s favour for novel work. FASB’s ASU 2025-06, issued in September 2025 and effective for annual periods beginning after 15 December 2027, replaced the old sequential project-stage model for internal-use software with a probable-to-complete threshold — and attached a novelty gate to it:

If significant development uncertainty exists, the probable-to-complete recognition threshold… is not met until that significant development uncertainty has been resolved.¹⁰

Uncertainty is defined to include novel or unproven features, and performance requirements that are still being substantially revised. That is a precise description of the engagement a learning subsidy exists to fund. The accounting will not help you there, but the vocabulary will: significant development uncertainty is a better name for the state a first engagement is in than anything this book had, and it gives the subsidy a clean resolution test. The bet is resolved when the uncertainty is, and not before.

The tension, staged rather than dodged

Investors have been complaining about this asymmetry for years. More than 70 per cent of respondents to a CFA Institute survey agreed that for many companies the most valuable assets do not appear on the balance sheet, and that expensing internally generated intangibles front-loads costs while the benefits arrive later, distorting profitability and valuation.¹¹

And from the same source, the objection that has kept the line where it is: “Greater flexibility in capitalization could be abused to manage earnings.”

Both are true, and the second one is the reason the first has not been fixed. Which lands the argument exactly where this book wants it: every discipline in the capability ledger — named, owned, rights-checked, dated, engagement-two-tested — is the price of being taken seriously by the person who raised the abuse objection. A firm that wants credit for capability investment without any of that machinery is asking to be trusted, and the whole reason the standards are strict is that trust did not work.

“If it’s all expense anyway, why bother?”

Because the accounts answer a different question from the one a principal is asking. Statutory accounts ask what may be recognised. The subsidy asks what was bought, by whom, and was it repaid. A firm that only runs the first is structurally incapable of telling an investment from an overrun — which is where Chapter 1 came in, with two firms and one line.

So: keep the subsidy visible as a management line inside delivery contribution, not as a capitalised asset. Report it beside the capability ledger, so that the amount and what it bought are read in the same minute by the same people. And do not let a conversation about capitalisation become a reason to abandon the control, because the standards are not the obstacle here. They are the reason the control is necessary.

The accounts cannot see the second return. Neither, it turns out, can the promotion committee — and that is the harder problem, because that is where the people are.

The first is a list of hard problems solved: architecture calls taken, escalations answered, engagements rescued. It has a number attached, and the number is in the system.

The second is a class of problem that stopped arriving. It has no number, no field, and no place on the form — and it is worth considerably more.

This is not a values problem and it will not yield to exhortation. It is a counting problem.

Why asking nicely has never worked

Within a company, you still try to keep your own intellectual property to yourself.

That is an observation about people rather than a complaint about them. If you are the person who knows the obscure thing, your utilisation rises, your indispensability rises, and your promotion case writes itself. Asking that person to document everything for the benefit of the organisation is asking them to reduce their own leverage, and the answer has been the same for thirty years: “People were encouraged to share, but it didn’t really work. Everyone wanted to be the hero.”

The status inversion — that the highest-status act should be solving the hard case once and making sure nobody needs you for that class again — is already named in the architecture above this book, along with the condition attached to it: it only works if the firm actually measures the second thing. This chapter is the measurement.

Fossil credit

Escalation classes retired: attributed to a person, dated, and reported in the same review as utilisation. Four properties make it a ledger line rather than a values statement, and each one prevents a specific way the idea normally dies.

The underlying rule is inherited and not re-argued here: every escalation should reduce the probability of the next equivalent escalation, and a fossil is a framework, a playbook, a decision rule, a code pattern, a test, an eval, an anti-pattern, an example or a mandatory escalation trigger. What fossil credit adds is the ledger entry.

One unit of it, from the cohort

Chapter 8 had an unplanned nomination. A principal answered a novel architecture question and this time, instead of answering it and moving on, wrote down a de-identified decision tree, a mandatory evidence list and a regression test alongside the answer. In Chapter 9, a different team resolved that class without going anywhere near them.

That is one unit of fossil credit, and it is the only kind that can be audited — because the evidence is the escalation that did not happen. Note the cost, because it matters to whether anyone will do it: roughly a day of writing, on top of solving the problem. That is not free, and pretending it is free is how firms get compliance instead of fossils.

The counter-incentive, taken seriously

An expert whose escalation rate falls has just reduced their own indispensability. That is a real cost to a real person, and it is not answered by a slide about culture.

And here is the honest limit. This book cannot show that fossil credit survives a partnership’s promotion process. The mechanism is sound, the measurement is auditable, and none of that guarantees it will outlast one bad quarter and one partner who wants a utilisation number defended. It goes on the falsifier list in the next chapter, where it belongs — not in a footnote.

“They’ll just write documents nobody needs”

They would, if the artefact were the evidence. It is not. The evidence is the falling escalation rate for a named class, which means a fossil nobody’s next call routes through scores exactly nothing. The metric is deliberately unpleasant to game: you have to be right about which class actually recurs, and being right about that is most of the skill.

What the expert becomes, and what happens to everyone else

The role on the other side of this is not smaller. The senior expert becomes a canon author, a product improver, an exception resolver — rather than the human API every project must call. Their scarce time migrates towards genuinely novel edges, which is both more interesting and considerably more defensible than being the person everyone books.

For those who never fossilise anything, the consequence is not punishment. It is reclassification. They are delivery capacity, they are priced and staffed as delivery capacity, and the firm stops describing a full calendar as a capability strategy. That needs to be exactly fair: it is not an accusation of laziness, and some of the best deliverers in any firm are not canon authors. A practice needs both. What it cannot afford is to confuse them.

The industry data suggests the current arrangement is failing the seniors as much as the firm. SPI Research’s own critique of its most famous metric notes that firms exceeding 80 per cent utilisation suffer higher burnout and attrition, and that in lower-maturity firms senior consultants are often overburdened through a lack of delegation, which makes them increasingly expensive to replace.¹³ The problem is documented. The fix proposed here is not, and that asymmetry should stay visible.

One person’s ledger scales to one offer’s ledger. But a firm does not run one offer, and the last chapter is the portfolio: three candidates, one budget, and the observations that would show all of this to be ceremony.

Thursday. Three candidate offers, one subsidy budget, and a decision that cannot be deferred again. Offer A has a receipt from a completed cohort and wants to scale. Offer B is halfway through its first engagement and is already asking about a second subsidy. Offer C has a sponsor, a good story and no first client.

A firm does not approve a subsidy. It allocates a subsidy budget — and the allocation is governed by something almost nobody has priced: the asset a subsidy buys depreciates.

Four allocation rules

• Cap per offer, not per engagement. An offer that needs a second subsidy for its second engagement has told you something important. Per-engagement caps hide exactly that signal by resetting it. Offer B’s question answers itself under this rule.
• Stagger expiries so two cohorts are never both unproven. Otherwise the firm is holding two unfalsified bets simultaneously, and it will resolve them together, in the same meeting, under the same mood.
• Never fund a second subsidy for an offer whose first receipt is unpaid. This is the escalation brake, and it is the rule that will be argued with hardest in the room — always by someone who genuinely believes the next one will be different.
• Reserve capacity for repair. Repair is the most common correct answer after a failed receipt and the least funded one. A portfolio with no repair budget has only two moves, scale and kill, and a firm with only those two moves chooses scale.

Why expiry is a field and not a formality

Organisational knowledge depreciates, and the rates in the literature are startling. One study of pizza franchises found knowledge depreciating at 17 per cent per week — “roughly one half of the stock of knowledge at the beginning of month would remain at the end of the month”. Wartime Liberty ship construction showed 25 per cent per month, with a follow-up study finding a slower but still substantial 3.6 to 5.7 per cent.¹⁴ Rates vary by setting and by method. The direction does not.

The causes read like a description of what a promotion membrane exists to prevent: technological change making past knowledge irrelevant, individual forgetting, failure to codify knowledge in organisational memory systems, ineffective knowledge management, and staff turnover. And the economists have the image that carries it — competition with learning and forgetting is “akin to racing down an upward-moving escalator”: keep selling and marginal cost falls; slow down and the firm slides back up its own learning curve.¹⁵

The curve, honestly

It is tempting to reach for learning-curve theory and let it do the arguing. Do not — the literature does not say what everyone assumes it says.

Three things follow, and each one bites. Variation is enormous, and it appears “not only across industries, products, and processes, but also for subsequent runs of the same product within the same plant” — which is precisely the comparison a cohort ledger is making. Services learn slower than manufacturing, by a wide margin. And the whole log-linear picture is confounded, because “in any demand curve with a constant elasticity of demand, it can be shown constant exponential progress yields the same log-linear relationship predicted by a learning curve”; in the one clean natural experiment, experience explained between 40 and 67 per cent of the observed cost reduction.¹⁶

The limit that bites hardest for an AI-native offer is the last one in that source: the evidence applies best where there is a standard production process to iterate on. “If we need to completely change the method of manufacture or the structure of the technology — well, I don’t think we should count on learning by doing to deliver that.” Which is a fair description of an offer whose method changes every time a model does.

Therefore: engagement two must be measured to be cheaper. It is not cheaper by theory. A firm that reaches for Wright’s Law to justify a subsidy has substituted a citation for a receipt.

The first-of-a-kind gift

One industry has been arguing about this for fifty years and has produced a definition worth stealing outright. In nuclear construction, non-recurring costs — design, testing, licensing, supply-chain qualification — can add roughly 30 to 35 per cent to the construction cost of a first reactor, and are meant to be amortised across subsequent units. Then the OECD Nuclear Energy Agency had to add a caveat that reads like it was written for this book: some reactors are still counted as first-of-a-kind even when they are not literally first, “in the absence of a governance model to ensure the effective transfer of learning from one reactor to another.”¹⁷

Without a governance model for transferring learning, the second one is still a first one. Engagement two without a promotion membrane is engagement one again, at full price. The same source adds the warning that completes it: “some nonrecurring cost may become recurring if normal learning process dynamics are altered” — which in this vocabulary means that if the perimeter mutates between the two engagements, there is no series effect left to harvest, and the change-control instrument that keeps a perimeter stable is a separate discipline with its own treatment.

And one more from the same literature, because it is the reason a ceiling exists at all: a review of 44 large engineering projects found final construction costs were often twice the initial estimates, with one named cause being “reporting low-cost estimates as a strategic decision to secure early support from stakeholders.” Optimism about what a first engagement will cost is documented across industries. It is not a personal failing of your delivery lead.

This is not discounting

The separator is worth stating precisely, using the legal tests as vocabulary rather than as compliance advice. Predatory pricing is defined by its target and its payoff mechanism. In the United States, a plaintiff must show prices below an appropriate measure of cost plus “a reasonable prospect, or, under §2 of the Sherman Act, a dangerous probability, of recouping its investment in below cost prices”.¹⁸ In Australia, the regulator describes predatory pricing as occurring “when a firm substantially reduces its prices below its own cost of supply for a sustained period” in ways that damage or deter competitors.¹⁹

Both point outward, at a rival’s viability, with the payoff arriving through restored market power. A learning subsidy points inward: its target is the firm’s own future cost curve and its payoff is a named asset that lowers the cost of engagement two. If a principal cannot say which one they are doing, they have a discount, not a thesis.

Two limits, stated plainly. Neither authority says ordinary below-margin pricing by a non-dominant services firm is unlawful — both require market power or below-cost pricing plus recoupment. And this is not legal advice. The tests are being borrowed for their clarity about purpose, which is the thing a subsidy record documents contemporaneously and a discount never does.

What would falsify this book

And the evidence gap, owned rather than buried: there is no published paired first-versus-second engagement margin dataset for professional services. The industry numbers that exist — utilisation, project margin, overrun rates — are cross-sectional aggregates, not paired observations. Everything in Chapters 8 and 9 is a measurement design, and the literature above says that even many pairs, badly controlled, would not settle it either.

Monday

Engagement one may buy the machine. Engagement two has to prove you bought one.