The Terminal Value Doctrine: Professional Services

When the customer acquires your production function, the labour-priced unit compresses from both sides — and no competitor ever shows up in the loss report.

Scott Farrell · LeverageAI · August 2026 · A doctrine article for managing partners, boards and CFOs of firms whose product is cognition

The client opens a folder and puts something on the table before you have finished sitting down.

It is their own baseline analysis. Market sizing, competitor profiles, a first-pass model with the assumptions listed down the side. It is not as good as yours. It is good enough that the conversation now starts at a point you used to bill for arriving at — and they are pleased with it, because it took a weekend and cost nothing.

Nobody in that room experiences this as a competitive event. No competitor was mentioned. No deal was lost. That is precisely why it matters. In the loss-reason column of your CRM there is no code for what just happened, because nothing was lost yet. The engagement that would have started with six weeks of discovery will start with two. The four-person team will be quoted as two. And eighteen months from now, when the revenue line finally reports what began in that meeting, every explanation on offer — pricing pressure, a soft market, a bad sales quarter — will be wrong in the same direction.

This article is the short form of a doctrine we have now written at book length: the professional-services continuation of the Terminal Value Doctrine, which asked boards to judge AI capital by its effect on future enterprise value rather than near-term workflow ROI. The original doctrine worked a mid-tier consulting firm as one of three illustrative variants and said, explicitly, that the variant was a worked example, not an industry report. This is the industry report — or rather, the industry doctrine: what happens to a firm whose entire product is cognition when the customer acquires part of the production function, and what a firm has to prove before it deserves to call itself the successor.

1. The revenue nobody won

Start with the observable fact pattern, because it is already public and dated.

Gartner's consulting segment — advisory work helping firms execute strategy — fell about 13% in Q4 2025, and Reuters attributed part of the softness to companies using in-house AI tools to do planning and performance work internally. The share price dropped 22% on the news.⁴ Klarna announced it had cut external agency spend by 25% — translation, production, CRM, social — by generating the work in-house, while increasing the number of campaigns it ran.⁵ In legal services, the most instrumented professional vertical, 67% of corporate legal departments expect AI to change how hours are billed, 71% of buyers already prefer a flat fee for an entire matter, and the researchers' own summary sentence is the one that matters: clients have moved faster than the industry has.^1,2 Fifty-two per cent of corporate counsel plan to handle more work in-house within five years.³

Read the direction of that gap out loud: the demand side repriced before the supply side decided to. Not a startup. Not a platform. The people already paying you. Practitioners inside the legal market state the mechanism plainly: generative AI "will disintermediate lawyers from many legal solutions... AI won't take away all the hours. But it'll take away enough to do some serious damage."²⁴ And public markets have started pricing the same mechanism one industry over: software indices fell 20–23% in early 2026 while the broader market stayed flat, with the analysts' summary line — "risk is up, and terminal value assumptions are down" — including their own hedge that investors may be selling first and asking questions later.⁶ The argument here does not need the re-rating to be correctly sized; it needs only the direction, and the direction has a price on it.

Your customer does not need to switch supplier to disintermediate you. They can simply buy less of you.

The sequence is quiet. "We used to need four people; now we need two." Then: "We can build most of the dashboards ourselves — could you just help with the difficult bits?" Then: "We've got the requirements pretty well worked out. Can you quote this very narrow piece?" And eventually: "We think we'll do this internally." No competitor won the missing revenue. The requirement for an external supplier contracted.

This is why the standard loss telemetry cannot see the threat. "Lost — no decision" and "client proceeding internally" are economically different events wearing the same CRM code. The first is a deferral; the second is your market disappearing one work package at a time. A firm that wants to see this before revenue reports it needs a different ledger — we'll come to it — because market-share loss and market disappearance produce identical quarterly numbers and require opposite responses.

2. The externalisation share

To reason about this without hand-waving, change the unit of analysis. Not the firm. Not the industry. The revenue unit — the thing a client actually buys: a report built, a system configured, a strategy reviewed, a migration delivered.

A professional-services firm exists because a client chooses to externalise some combination of problem framing, research and analysis, specialised judgment, construction and configuration, verification, implementation capacity, and authority or consequence. For each outcome class, define the Externalisation Share:

E  =  work purchased externally ÷ total work required

External revenue pool:      R  =  Q × E × P
Inherited labour demand:    L  =  Q × E × h

Q  = total demand for the outcome
E  = share externalised
P  = price of the external unit
h  = supplier labour per external unit

AI moves every variable at once, and not in the same direction. Q may rise — much more analysis, software and assurance becomes economically possible when cognition is cheap. E falls — clients self-supply more of the work, starting with reproducible cognition, because that is the part of your invoice they most resented paying for. P bifurcates — it falls for generic cognitive output and can rise for verified commitments, independent authority and transferred risk. And h falls, because your own delivery is getting more productive too.

That last pair is the trap. Supplier productivity (h↓) does not offset demand contraction (E↓). It multiplies the capacity surplus. Which brings us to the arithmetic every managing partner should run on their own numbers before their next board meeting.

3. The two-sided squeeze, worked exactly

A firm historically sells 90 of every 100 available consultant-days. Ten days of slack. Now let AI and client self-service reduce demand by only 10%: the firm sells 81 days. Unsold capacity has gone from 10 days to 19 — a 10% demand fall nearly doubles the bench, because slack absorbs the shock first. Nothing about that requires a dramatic competitor, a lost logo, or a headline. It requires several small compressions arriving at once, which is exactly what the data now shows.

Then add your own productivity programme. Let d be the proportion of external demand retained and p the delivery productivity multiplier:

Old-unit capacity requirement  ≈  d / p

Clients retain 80% of former demand:      d = 0.80
AI makes delivery 25% more productive:    p = 1.25

0.80 / 1.25  =  0.64

This is why "our people are 25% more productive" can be an alarming sentence rather than a reassuring one. The productivity is real: the largest field experiment run inside a strategy firm found consultants using GPT-4 completed 12.2% more tasks, 25.1% faster, at higher quality.⁸ The demand-side movement is real too, and the benchmark data captures both blades closing at once: billable utilisation across 509 professional-services organisations fell to 66.4% in 2025 — the lowest point in the survey's history — while industry growth ran at 5.2%, roughly half its historical rate, and generative AI was already inside 27.1% of delivered projects.⁷ Record productivity adoption; record-low utilisation. That is not a paradox. That is the multiplication, showing up in a benchmark.

And the substitution mechanism is visible at the task level. When enterprises wire AI into their own operations through the API — that is, when your client's team builds rather than chats — 77% of usage shows automation patterns, predominantly full task delegation, versus just 12% augmentation.⁹ Delegated whole tasks are precisely the work that used to be externalised to junior consultants. The client is not collaborating with the model the way your analysts do. The client is replacing the purchase.

The firm's instinctive answer — "we'll use AI too, and get faster" — is not wrong. It is incomplete at the level that matters. The firm is asking: how can we produce our existing unit more efficiently? The customer is asking: why am I buying that unit at all? Those are not the same altitude, and no amount of excellence at the first question answers the second. If the water is rising, making everyone a better swimmer is useful at a local level. It is not a strategy for owning dry land.

4. The exception sink crosses the commercial boundary

Here is the mechanism almost everyone misses, and it is nastier than simple demand loss: partial substitution is worse than replacement.

A traditional engagement bundled ordinary and exceptional work. Routine research, repeatable analysis, drafting, configuration — the normal distribution — plus a smaller amount of senior judgment, political exceptions and liability-bearing decisions: the tail. The normal work did three economic jobs at once. It produced billable volume. It subsidised the difficult tail. And it trained less-experienced people through repeated exposure to normal cases.

Now let the client's AI take the first pass. The client researches the market, drafts the requirements, generates candidate architectures, challenges your estimates — and brings you only the disputed, novel or consequential questions. The consultancy loses junior-leveraged volume and retains senior judgment, political complexity and professional exposure. Revenue falls much faster than cost-to-serve.

The customer self-insures the ordinary cognitive work and cedes only the tail to the consultancy. The consultancy becomes a reinsurer of expertise — often without charging a reinsurance premium.

In insurance language this is adverse selection, and it reprices everything. Historical blended day rates were calibrated to the old case mix — a lot of normal, a little tail. They no longer price the new distribution. Senior effort rises as purchased volume falls. Project variance gets harder to underwrite. The firm can appear busy — flattered, even, by the seniority of what walks in the door — while its unit economics quietly break.

We first documented this mechanism inside a workflow, not across one: in our Tesla service case study, automated triage absorbed the routine tickets that had built the service advisers' judgment, then handed the humans only the rare, ambiguous, emotionally loaded failures. The human owned the broken residue of a system they didn't control — de-triage, not triage. The industrial literature predicted it in 1983: automate the routine and the operator's skills decay from disuse, so that "a formerly experienced operator who has been monitoring an automated process may now be an inexperienced one" — needed most exactly when something has already gone wrong.¹⁴ Fresh survey data shows the same case-mix shift arriving at scale: AI reduces routine volume and shifts human work toward the complex residue.¹⁵

What is new in professional services is that the sink now crosses the commercial boundary. It is not your automation de-skilling your people. It is your client's automation, de-triaging your economics — from the outside, rationally, without hostile intent, one retained work package at a time.

5. The pyramid was also an apprenticeship machine

The pyramid was never just a margin structure. Junior analysis, report construction, data cleaning, application configuration — that work manufactured the firm's future judgment. Whoever did it saw hundreds of ordinary cases, recurring client misunderstandings, subtle data problems, the same requirement expressed ten different ways. The boring exposure was where a consultant quietly built a map of the real world.

That base is now being dismantled, visibly, in public data. Job-postings analysis of the leading consulting firms shows overall hiring about 20% below peak, consultant hiring down around 40%, senior hiring up 55% since 2020 — and by 2025, AI roles outnumbered entry-level consultant postings.¹⁰ The UK Big Four cut graduate intakes by between 6% and 29% in a single cycle, with AI's absorption of entry-level tasks named as a cause.¹¹ Economy-wide payroll data puts a number on the on-ramp itself: employment for 22-to-25-year-olds in the most AI-exposed occupations is shrinking at 3.8% per year. The researchers' phrase is exact: the technology "isn't eliminating work across the board. It's eliminating the on-ramp."¹²

Harvard Business Review's verdict on the same data is calibrated: consulting "isn't disappearing; it's being fundamentally reshaped" — AI automating exactly the research, modelling and analysis that junior consultants existed to do.²⁷ Uber's co-founder put the unreshaped version less gently: "If you're a traditional consultant and you're just doing the thing, you're executing the thing, you're probably in some big trouble... Push a button. Get a consultant."²⁵ And industry commentary has started saying the second half out loud: AI "is altering the structure of the professional services firm itself by compressing traditional leverage models and eroding the apprenticeship layer firms have relied on for decades to train, assess, and produce future partners."¹³ The early repetitions that anchored the learning curve are disappearing. Training programmes alone cannot close the gap, because the pyramid's teaching was a by-product of paid production — and the paid production is what the client just took back.

Run the tape forward. AI performs the normal analysis. Junior people stop seeing normal cases. Only difficult exceptions reach humans. Current seniors resolve them from experience accumulated under the old model — and the firm mistakes amplified senior judgment for a renewable capability. Ten years later it has a small group of ageing experts and no mechanism for creating their replacements. The firm that "right-sizes" by deleting its junior layer has consumed a compounding asset and reported it as efficiency.

The successor firm therefore needs an explicit apprenticeship system, designed rather than inherited: ordinary consultants inspecting sampled normal cases, not only escalations; AI decisions replayed and adjudicated; de-identified cases turned into simulations and held-out evaluations; authority increasing only with demonstrated calibration; seniors recording why alternatives were rejected. The test is stark: can someone who did not build the original expertise become a reliable judge through the firm's operating system — or is the knowledge base merely embalming the principals?

6. The race to compile

Everything so far is diagnosis. The strategic contest underneath it can be stated in one sentence:

The professional-services firm is racing its customers and its future competitors to compile its own expertise.

Three parties are trying to turn the same fuzzy work into machinery. The client is compiling depth of local context — its own history, data, requirements and judgment into internal AI capability. The AI-native entrant is compiling new production economics with no legacy cost base to protect. The firm must compile the one thing the other two cannot: breadth of transferable variation — twenty years of engagements, exceptions, corrections and cross-client pattern recognition. The provider wins only where cross-client variation, independence or accountable operation outweighs the client's context advantage and the entrant's cost advantage. If the firm compiles first, its accumulated discrimination becomes machinery it can sell repeatedly. If the client compiles first, a growing fraction of the firm is unnecessary. If the attacker compiles first, the firm becomes the legacy provider.

Which is why the historical IP sitting in proposals, project folders, emails and partners' heads suddenly matters so much. It is not a knowledge-management programme. It is raw material in a race to determine who owns the next unit of production.

But capture is not compilation, and here the firm's own operating instruments betray it. Our old firm — IC Consulting, in the early 2000s — ran a WIP system that measured available hours, worked hours, productive hours, billable percentage, billable amount, average charge rate. That system did not merely administer the business; it told management what value was. If the ledger can see billable hours but cannot see reusable learning, the rational manager maximises billable hours and treats capability construction as leakage. The old system measured effort accurately. The successor must account for something the timesheet never captured — and that requires a second operating ledger.

1. Client observation        something happened in one engagement
2. Candidate invariant       a hypothesis about what recurs beyond that client
3. Rights-cleared abstraction fingerprints removed; contract boundaries satisfied
4. Callable capability       encoded as a test, rule, adapter, pricing boundary
5. Independent reuse         invoked by another team, without the originator
6. Realised contribution     demonstrably improved cost, quality, risk or value

Three ledgers, then, replace the one the pyramid ran on: a Demand-Side Disintermediation Ledger (what is the client no longer buying externally — split into competitor displacement, scope compression, internal substitution and stack exit, so the revenue nobody won finally gets a row of its own); a Capability WIP Ledger (what has the firm converted into demonstrably reusable production capital); and a Successor Proof Ledger (what new unit has produced paid demand, independent acceptance, transfer and attractive economics). The old firm counted hours. The successor counts how much demand remains external, how much experience has become callable machinery, and whether the new unit can win twice without the hero.

7. The last Domino consultant

Now the correction that keeps this doctrine from collapsing into an obituary — because zero is the boundary case, not the forecast, and the difference is worth real money.

I lived a demand-side crash once before. I spent years in the Lotus Notes/Domino market. Domino carried email and custom applications as a joint product; when Microsoft Exchange removed the anchoring email use, the remaining application estates became hard to justify as a standalone platform — and more than 95% of the suppliers left the market. Here is the part the doom narrative misses: the few who stayed inherited the remaining work of every client still running the platform, and made good money right up until the last client left. Supplier capability exited faster than demand. The survivors enjoyed scarcity rents on a declining category.

The public record confirms the shape at every scale. IBM's mainframe business posted its highest revenue in twenty years in 2025 — six decades into the category.²¹ The average COBOL developer earns $115,475 a year, comfortably above the median software developer, in a language routinely declared dead.²² And the Notes/Domino franchise itself — twenty years after losing the category war — was still worth $1.8 billion to HCL in 2018.²³

Declining terminal value and strong harvest returns can coexist.

So separate three objects that the phrase "terminal value" keeps blurring. The current unit of sale — the labour-priced hour, the bespoke project — is the thing whose value trends toward zero as the customer self-supplies its underlying scarcity. The book of business has harvest value: runoff cash, tail scarcity, migration tolls. The enterprise is worth whatever migrates. As a valuation:

value today = runoff cash
            + tail scarcity and migration income
            + convertible assets
            + successor options
            − restructuring liabilities

The board must not mistake tail profits for a recovered future — that is the Domino lesson. But it should not destroy them either, because harvest cash is precisely what funds the migration. The failure mode is not harvesting. It is harvesting while calling it a strategy.

8. What will the client still buy?

The strongest case against the doom thesis is not that traditional consulting carries on. It is that the market bifurcates rather than collapses — and the honest version of this doctrine holds that counter-case inside itself, with a falsifier attached.

Cheap cognition makes the customer a better producer and a better buyer: better able to specify a bounded need, compare specialists, supervise delivery, verify outputs. A customer might stop buying a six-person, twelve-week analytics programme and instead buy an independent semantic review, a security test, a complex connector, a regulatory opinion, a bounded implementation commitment and continuous verification. External hours collapse while the number of specialist transactions rises. AI can also create demand for services nobody previously offered economically — exhaustive reconciliation, continuous assurance, every-account analysis — services suppressed for decades because human breadth and frequency costs made them irrational to sell.

The aggregate numbers say this recomposition is real revenue, not consolation. Accenture tripled generative-AI revenue to $2.7 billion in FY2025 and grew 7% overall.¹⁹ EY's AI-related revenue rose 30%.²⁰ The Big Four all grew in FY2025 — while PwC's headcount fell by 5,600 and its growth slowed for the third consecutive year.¹⁸ Aggregate demand is growing while the labour-priced externalised unit compresses. Both facts are true at once; the doctrine's variables (Q rising, E and h falling) are exactly what lets you hold them together without hand-waving.

And the industry's own reference firm has now described the migration on the record. McKinsey's global managing partner counts the firm as "40,000 humans and 20,000 agents," describes moving "away from pure advisory work... and a fee-for-service model" toward underwriting outcomes, and says about a third of revenues already sit on that model — with the ambition of a majority.¹⁶ When the temple of the leverage pyramid tells you a third of its revenue no longer prices labour, the question is no longer whether the unit is migrating. It is whether your firm's migration will be proved before your runway runs out.

But here is the discipline most strategy documents skip: run the boundary case against your own successor. Assume your best clients have frontier AI, their entire internal knowledge estate compiled into a usable corporate wiki, excellent build agents and good evaluation machinery. What would they still rationally buy from you? The original Terminal Value Doctrine's consulting variant proposed kernel publishing and subscription advisory as the successor. Attack that answer with the same weapon: an AI-native client can ingest the same kernel and apply it without the adviser. Subscription advice is not automatically terminal value — it may be another temporary migration form. What survives the test is narrower and harder: offers backed by an advantage the client cannot cheaply reproduce — lawful cross-client variation, independent authority, superior verification machinery, accountable operation, transferred risk. Four things clients externalise: cognition, capacity, authority, consequence. AI collapses the first two. The durable successor sells the second two, wrapped around whatever cognition remains genuinely scarce.

Two honest cautions belong beside the counter-case. First, production capability is not yet substitution everywhere: 94% of B2B buyers now use LLMs, and are running two-thirds of the buying journey before ever engaging a seller — yet the same survey finds this has not yet reduced their reliance on vendors and third-party experts.²⁶ The first pass has moved in-house; the purchase, so far, has not always followed. That is exactly why the leading indicator has to be the externalisation share, not the revenue line — the behaviour moves before the budget does. Second: why won't clients simply do everything themselves? The best current evidence: most of them can't — yet. About 95% of enterprise generative-AI pilots are failing to reach P&L impact, not because the models are weak but because of the learning gap between tools and organisations.¹⁷ That failure rate is the successor market: bounded, accountable, fixed-price conversion of AI capability into working systems, sold to clients who have discovered that capability is not the same as outcome. Treat the number as decaying evidence — it describes 2025-vintage organisational maturity, not a law of nature — which is precisely why the window is a window.

9. The sign flip: what "AI-native" actually means

Most firms use "AI-native" as decoration. The doctrine gives it a definition with a derivative in it.

For the legacy professional-services firm, every model release hurts: the customer can do more, juniors differentiate less, billable hours compress, implementation gets easier, generic advice gets cheaper. Better AI → enterprise value down. For a firm that has reorganised — expertise compiled into callable machinery, evidence systems, evaluation harnesses, bounded offers — every model release makes its assets more productive: the kernel answers better, the compiled project history becomes more exploitable, smaller teams carry more work. Better AI → enterprise value up.

An AI-native professional-services firm is one that has reorganised its assets and commercial model so that advances in AI increase its enterprise value rather than decrease it.

That is the sign flip, and it is the cleanest test in the doctrine because you cannot theatre your way through a derivative. AI, on this view, is not a tool the firm adopts. It is industry pressure — pressure that doesn't affect all assets equally. Stranded assets (first-pass analysis, deck factories, pyramid economics) get worse under pressure. Convertible assets (decades of engagements, methods, judgment) become economically feasible to compile. Compounding assets (the kernel, the evaluation machinery, the evidence systems) are made more valuable by the very force destroying the old model. Pressure doesn't automatically make diamonds — pressure also crushes. The material decides. AI pressure widens the spread between good and bad capital allocation; the same force liquidating one firm compounds another.

Management's job, then, fits on two clocks. The runway clock: how long can harvest cash support the transition under measured demand erosion? The proof clock: how long until a successor unit produces paid customer value, engagement-two transfer and credible unit economics? The survival inequality is one line: time-to-successor-proof < runway-under-erosion. Everything in this doctrine — the ledgers, the compilation race, the harvest discipline — exists to move one clock or honestly read the other.

And the proof standard is three migrations, or it didn't happen. Production migration: paid bounded units rise relative to scarce-expert dispositions — otherwise you have productised the expert, not the service. Budget migration: customers move spend from consultant-days into the successor commitment — without it, the offer is an efficient sidecar. Capability renewal: the system creates more people capable of future judgment than it consumes — otherwise the smaller, sharper firm is quietly exhausting the seniors the old pyramid made and replacing them with nothing.

10. The author is inside the blast radius

One more thing, because the reader deserves to know where this doctrine comes from and what it would take to kill it.

I run a professional-services firm. This doctrine, applied honestly, threatens my business model too — and it is not lost on me that I am a services organisation telling services organisations the end of the old unit is near, while selling migration help. That is a principal–agent conflict, and disclosure alone doesn't dissolve it. Structure does: in our own offer architecture, the diagnosis is a separately priced decision product whose valid outcomes include stand pat, harvest only and build nothing; the construction is separately commissioned; the decision pack is portable to another builder. The book nominates the equation. The engagement is allowed to discover that its coefficients do not apply to your firm.

The same honesty applies to our evidence. Our own specimens — the compiled IP wiki, a production deployment that removed a CMS translation layer at brochure depth, a governed BI delivery specimen built on synthetic engagement evidence, a proposal-compilation system — are working organs, not yet a proven organism. Each sits on a dated evidence ladder (argued → designed → implemented → internally used → externally sold → repeated → transferred → economically scaled) at its honest rung, and the commercial figures attached to our successor offers are designed hypotheses, not validated market prices. The doctrine also names its own falsifiers: if external spend per resolved consequential decision holds or rises among AI-mature clients, if externalisation share stabilises, if bifurcation demand exceeds displacement — the strong form of this argument is wrong, and the book that develops it says exactly what evidence would show that.

Because the rain isn't coming. It's raining — and AI is not just the rain falling on your business; it is also giving your customers boats. The response that deserves capital is not "swim 20% faster." It is: make the current vessel seaworthy; harvest while the old routes remain navigable; convert everything valuable onboard into portable capital; build successor vessels faster than the water rises; and keep searching for the new coastline — urgency without a landfall forecast, because under AI Fog nobody honestly has one.

Conclusion

The industry conversation is stuck one level too shallow. "How do we use AI?" produces productivity. "How do we sell AI?" produces trinkets bolted to the old catalogue. The consequential question is what AI is doing to the economic conditions that made the firm valuable in the first place — and the answer arrives quietly, as revenue nobody won, while your own productivity programme enlarges supply into the contraction. Measure the externalisation share before revenue reports it. Price the tail you're inheriting. Protect the apprenticeship machine you're about to delete. Compile before your client does. Harvest without calling it strategy. And hold your successor to the three migrations, because until production, budget and capability have all moved, it didn't happen.