The Learning Subsidy

Engagement one buys the machine. Engagement two proves you bought one.

When should you knowingly accept poor margin on the first delivery of a new AI-native offer — and how do you stop “we learned a lot” becoming an excuse? Scott Farrell, LeverageAI.

Two firms close the first delivery of a brand-new AI-native offer in the same week. Same shape of engagement: a fixed commercial envelope, two senior people, about eight weeks, a client who is genuinely pleased. Same line at the bottom: contribution came in a long way under target, close enough to break-even that nobody wants to say the number out loud in a portfolio review.

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

Nothing in either firm's management accounts can tell you which is which. That is the whole problem, and it is not an accounting problem — it is a design problem, because the two firms were never running the same instrument. One of them decided, before the engagement started, what the shortfall was buying, how large it was allowed to get, and what observation would end it. The other decided afterwards that it had been learning.

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 already published, in the service-architecture work that sits above this piece.¹ It is correct, and on its own it is not usable, because it is a verdict rather than a control. It tells you how to judge the engagement once it is over. It does not tell you who approves the shortfall, up to what limit, against what named list, by when, or what ends it. The same book says so explicitly 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.”¹

This is that treatment. The instrument is a learning subsidy: a pre-declared, capped, expiring appropriation of the difference between target contribution and approved first-engagement contribution, spent against a named list of reusable assets, and repaid — or not — by ordinary staff in engagement two.

Why one ledger cannot hold two returns

An early engagement of a productised offer creates two things at once. The customer gets the thing they bought. The firm may also get something that makes the next engagement cheaper, faster, safer or possible at all: an exception taxonomy, an evaluation harness, a corrected pricing driver, a deterministic tool, a rejected architecture and the reason it failed, an acceptance pack that can be reused with local cases.

A single contribution line can only show one of those. The second return is real and invisible, so it gets claimed rather than recorded — and a claim cannot be capped, audited, or killed. That is how the oldest sentence in professional services survives contact with evidence indefinitely: “we're investing in learning” becomes, as the source transcript for this work puts it, the oldest consultancy excuse in the book for losing money.

Two symmetrical failures follow, and firms tend to specialise in whichever one their culture prefers. Finance kills legitimate capability building because it looks exactly like margin leakage. Or delivery romanticises uncontrolled loss until “strategic account” becomes a permanent subsidy with no asset, no owner and no end date.

The Learning Subsidy Ledger: two sides, never netted

The artefact has two balanced sides, drawn in a fixed order. Order matters more than layout: the client side is drawn first, and it is not available to fund the second.

The closure discipline underneath this is inherited rather than invented: an engagement is not closed until the customer received the promised, falsifiably accepted outcome and the provider disposed the resulting learning into exactly one disposition with a named owner and a date.¹ What the subsidy adds is the money. The learning half of that closure has, in most firms, no owner, no budget line and no receipt. The ledger gives it all three.

No asset named, no subsidy approved

The approval instrument is one page, and it is written before the engagement, which is the only property that matters.

The obvious objection is that you cannot know in advance what you will learn. True, and it misses the distinction the record turns on. You cannot predict what you will discover. You can absolutely name the classes of artefact the engagement must leave behind and the test each must pass. The eval harness is foreseeable; which cases go into it is not.

And there is a second-order effect that is worth more than the accounting. A team that has pre-committed to producing an exception taxonomy, an eval harness and a corrected band driver structures the work differently from a team that hopes to learn something. Pre-declaration is not hygiene applied to engagement one. It is a design instrument that operates on it.

That is not merely an ethical position — it is, remarkably, the accounting position too. IAS 38 forbids exactly this move: “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.”² And it names the specific costs that can never be part of an asset: “identified inefficiencies and initial operating losses incurred before the asset achieves planned performance.”² The standards wrote the book's central rule down first. An overrun is an overrun.

What your accounts will actually say

Be honest about this early, because a finance partner will get there in about ninety seconds. Economically, this is capital formation in the delivery system. Statutorily, it is almost certainly an expense — and the more genuinely novel the learning, the more certainly it is expensed.

IAS 38 expenses all research-phase spend outright.² It capitalises development only when six criteria are met simultaneously. And it contains a paragraph that decides the question for a generative first engagement before you start arguing: “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.”² A first engagement in which discovery and construction are interleaved is, by default, all research.

Worse for a services firm, the thing a subsidy looks like it is buying — better people, sharper judgment — fails the control test outright: “an entity usually has insufficient control over the expected future economic benefits arising from a team of skilled staff and from training for these items to meet the definition of an intangible asset.”² The standard then names the condition under which knowledge is controlled — when it is protected by legal rights or a duty of confidentiality.² Which is the same rights condition the capability ledger already demands. “No asset named, no subsidy approved” is the management-side expression of an accounting control test.

The US position is moving, and not in the firm's favour for novel work. FASB's ASU 2025-06 replaced the old project-stage model with a probable-to-complete threshold, and added a novelty gate: “If significant development uncertainty exists, the probable-to-complete recognition threshold … is not met until that significant development uncertainty has been resolved.”³ The engagement a learning subsidy exists to fund is precisely the engagement that cannot book an asset.

None of that is a reason to give up the control. It is the reason you need your own. Investors have been saying so for years: more than 70% of CFA Institute survey respondents agreed that for many companies the most valuable assets do not appear on the balance sheet at all.⁴ The two-ledger device is the management answer to a real accounting asymmetry, not a workaround for it. And note the counter-argument from the same source, which the ledger has to earn its way past: “Greater flexibility in capitalization could be abused to manage earnings.”⁴ Every discipline in this instrument — named, owned, rights-checked, engagement-two-tested — is the price of being taken seriously by the person who said that.

The membrane has a price, and the ledger has to show it

Some of the most valuable learning from a first engagement will fail the rights or transferability gate, and it will fail precisely because it is valuable — it is specific. A client-specific regulatory exception, worked out at real cost by a principal over three days, may be genuinely brilliant and still be client truth. It stays on their side of the line.

The gate itself is not ours to re-derive here; the service architecture publishes it as three territories and a five-condition gate, with five dispositions and one of them mandatory.¹ What the subsidy adds is a consequence the gate did not have to price: a refused promotion is a subsidy line that will not be repaid, and it belongs on the ledger as one.

Recording refusals as a cost, rather than as an absence, is the only thing that keeps the gate honest in a bad quarter. A firm whose capability ledger has never shown a refusal is not running a gate; it is writing a ledger to look good. And there is a legal edge here that most firms have not noticed: a residuals clause protects information retained in an employee's unaided memory and “generally excludes tangible materials.”⁵ So the artefact-based capability this instrument demands — a written taxonomy, a harness, a decision tree, a regression test — is exactly the class of learning residuals does not cover.

Engagement two pays the receipt

The receipt has one design constraint that decides everything else: engagement two must be delivered by ordinary capable staff. Transfer means ordinary staff lead materially more because shared infrastructure changed — not because the same heroes worked late again.⁶ If engagement two still needs the same people at the same density, you did not invest in learning. You subsidised the customer.

That is not a rhetorical flourish, and it is not only our claim. The cleanest empirical study of learning-by-doing ran exactly this test by accident. In an automobile assembly plant, each ten-fold increase in cumulative production halved average defect rates — and when a second shift with different workers began, it started below the first shift's defect rate.⁷ The researchers' conclusion is the engagement-two rule in economics form: “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.”⁷

Change the humans. See whether the improvement survives. That is the test.

The transfer measurement set — estimate error, exception classes, escalation rate by class, scarce-expert density and the rest — is already published and should be referenced rather than rebuilt.¹ The subsidy's receipt adds the two rows that measurement set deliberately left out, because they are money rather than mechanism: cost-to-serve and engagement-two contribution against target. A flywheel can be turning while the appropriation still has not been repaid, and a principal needs to be able to see both.

The row that makes firms quit

Expect measured disposition load to rise in engagement two before it falls, because instrumentation reveals the true denominator.¹ Engagement one's disposition load looked low; it was invisible. A firm that concludes the new way of working “created overhead” and quietly stops counting has chosen narrative over control, and having chosen it once, never gets a real number again.

Four cases, quickly

Do not oversell the curve

It is tempting to reach for learning-curve theory and let it do the arguing. Resist that, because the literature is more useful when read honestly than when quoted selectively.

Progress ratios vary enormously — from 55% to over 100% across a survey of more than a hundred studies, with a modal ratio around 81–82% — and the variation appears “not only across industries, products, and processes, but also for subsequent runs of the same product within the same plant.”⁸ Services learn more slowly: the pizza-franchise study found a progress ratio of 93%, roughly 7% improvement per doubling, a very long way from the seductive 20%.⁸ And the whole log-linear picture is confounded, because “constant exponential progress yields the same log-linear relationship predicted by a learning curve”; in the cleanest natural experiment available, experience explained only 40–67% of the observed cost reduction.⁹

Two consequences for the instrument. First, engagement two must be measured to be cheaper; it is not cheaper by theory. Second, the subsidy needs an expiry, not just a ceiling, because organisational knowledge depreciates — 17% per week in the franchise study, 25% per month for wartime Liberty ship construction.⁸ Competition with learning and forgetting is “akin to racing down an upward-moving escalator”: if sales slow, the firm slides back up its own learning curve.¹⁰ A subsidy whose engagement two lands eighteen months later may have bought an asset that evaporated before it was used.

There is a first-of-a-kind literature that says the same thing from the other end, and it produced a definition worth stealing. In nuclear construction, non-recurring costs can add roughly 30–35% to the cost of a first unit, and are meant to be amortised across later ones. But the OECD Nuclear Energy Agency had to add a caveat: some projects are 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.”¹¹

The incentive inversion, made countable

All of this collides with the way professional services actually pays people. As the source transcript puts it: “within a company, you still try to keep your own intellectual property to yourself.” If you are the only person who knows the trick, your utilisation rises, your indispensability rises, your promotion prospects rise. “Please document everything for the benefit of the organisation” has been asked politely for thirty years, and it loses to utilisation every time, because only one of the two is counted.

The fix is not exhortation. It is a second ledger on the person. Every escalation should reduce the probability of the next equivalent escalation⁶ — so make escalation classes retired an attributable, countable contribution, the way billable hours are countable. The senior expert becomes a canon author, a product improver, an exception resolver, rather than the human API every project must call.

The industry numbers make the case for the firm rather than against it. Billable utilisation across professional services fell to 66.4% in 2025, the lowest SPI Research has recorded, while EBITDA held at 9.9%.¹² A firm with a third of consultant time unbilled that still calls capability-building unaffordable does not have a capacity problem. It has a denomination problem. And read the counter-reading honestly too: at roughly 10% EBITDA, an uncapped subsidy is genuinely dangerous. That is why the ceiling is not optional.

Why firms keep funding a bet that has already failed

The kill condition needs a design, not a promise, because the failure mode is documented. Escalation of commitment is “one of the most robust and costly decision errors addressed in the organizational sciences”, and the meta-analysis finds that one of its most powerful drivers is ego threat, with time invested among the strongest antecedents.¹³ The same work finds that “the sharing of decision authority may lead to greater levels of escalation”¹³ — which argues against a committee and for a single named subsidy owner with a dated kill obligation, plus an assessor who did not approve it.

Stage-gate practice has been saying the structural half of this for decades: “once a project begins, there is very little chance that it will ever be killed”, which is exactly why gates with teeth — investment decision points — correlate with profitability.¹⁴ A learning subsidy is a gate output: resource commitment approved, ceiling set, deliverables and date for the next gate agreed. What is new is the second ledger the gate is judging.

This is not discounting, and here is the difference

A learning subsidy is not a loss leader, not penetration pricing, and not land-and-expand. The legal tests supply a clean separator even though no one here is accusing anyone of anything. Predatory pricing is defined by its target and its payoff mechanism: pricing below cost with, in the US formulation, “a reasonable prospect … of recouping its investment in below cost prices”¹⁵; in the Australian formulation, sustained pricing below one's own cost of supply that damages or deters competitors.¹⁶

A learning subsidy points the other way. Its target is the firm's own future cost curve, and its payoff mechanism is a named reusable asset that lowers the cost of engagement two. One is aimed outward at a rival; the other is aimed inward at your own production system. If a principal cannot say which one they are doing, they do not have an investment thesis — they have a discount.

What would falsify all of this

Three observations would tell you the instrument is ceremony rather than control, and they are cheap to check.

1. No refusals, ever. If across three first engagements the capability ledger has never produced a refused promotion, the rights gate is decorative.
2. The same people, twice. If engagement two is consistently staffed with the people who delivered engagement one, the firm has not accepted the rule, whatever its documents say.
3. Promised equals accepted, every time. If the ratio is always 1.0, the firm is promising only what it had already built, and the subsidy is buying nothing.

And the honest limit of the argument: there is no published dataset of paired first-versus-second engagement margins for professional services. This is a measurement design, not an established empirical regularity. One pair of engagements cannot establish a curve, and anyone who tells you otherwise is selling you a slide.

Monday

Take the last three engagements your firm called strategic. For each one, answer a single question: what asset was named before the discount was approved? Count how many have an answer.

Then, before the next below-target engagement is signed, write the one page — amount, ceiling, expiry, named asset list, kill conditions, approver, assessor. It takes an hour. It is the difference between an investment thesis with a built-in kill condition and tolerating an early loss while hoping scale will fix it.

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