The Throat to Choke

The box above this section says how the exercise was run. The subject is professional services, the firms that sell thinking: accountants, auditors, lawyers, management consultants. An AI played five roles in turn, the incumbent firm, the challenger, the buyer, the workforce and the regulator, and I refereed every move across a story that runs from 2026 to 2030.

A note on "you". From here I write as if you run one of these firms, because that is who the exercise was built for. If you do not, read "you" as "a managing partner".

The question on the table is the one every managing partner is privately turning over. When AI makes analysis cheap to produce, do we protect the pyramid that pays us, or undercut our own prices before somebody else does? The pyramid is the shape of a big professional firm: a few partners at the top, many junior analysts at the bottom, and profit that comes from billing junior hours at several times what the juniors cost.

The question is the wrong one, and showing you why is the point of the exercise. The real fork, the point where this story could still go more than one way, is where you let the price cut land, and who is holding the blame when it does. Two parties you never sit across a table from set those terms while you fight the war you expected. One is the insurer who covers your firm when its advice turns out to be wrong. The other is the buyer's own internal team. 58, 22, 19

One reading convention before we start. A marker like the three above points at a checked fact in the Sources table at the end, with its source, a quote from it, and the date the quote was checked. A marker with a small amber star beside it points at a fact that holds with a caveat, and the Sources table states the caveat. Where a marker sits on a sentence that mixes fact with story, it backs the pattern only. The invented company and the exact figure are mine. A number with no marker is one I wrote. The full statement of that rule is in the seams.

Five players carry the story I wrote out of those five roles. The Firm is a composite of a Big Four accounting firm (Deloitte, EY, KPMG, PwC) and a top-three strategy firm (McKinsey, Bain, BCG). From the Big Four it takes the audit anchor, the scale of the partnership and the professional-indemnity exposure that comes with signing audits, valuations and tax positions. From the strategy firms it takes project pricing and the leverage model, the pyramid above. I built it from the steepest real cutters, so it copies no one company. Throughout, "the Firm" means only this invented company, and "the big firms" means the real Deloitte, EY, KPMG, PwC and the large strategy firms. The Challenger is a new firm of two to six senior people with no juniors under them. The Buyer is the chief strategy officer of a large company, who decides what work is bought outside and what is done in-house. The Analysts are the Firm's own junior staff. The Underwriter is the market that carries the Firm's professional-indemnity risk, the cover that pays the client when the Firm's advice is wrong.

Two phrases recur, so here they are once. AI-heavy work means work where AI produced most of the output. Unsigned AI work means AI-heavy work that no named professional has checked and put their name to. The whole story turns on the difference between them.

What follows is written from 2030, looking back.

The Firm opened with a loud announcement that changed almost nothing, and a quiet cut that changed a great deal. Remember what the Firm is: a composite, Big Four in audit anchor, partnership scale and indemnity exposure, strategy firm in project pricing and leverage.

The loud move was an alliance with a leading AI lab and a cloud provider, under a ten-figure "AI investment" headline. 1, 6, 5 Most of that number was existing technology spending, relabelled. EY's own launch release describes its US$1.4bn in the past tense, as money already spent. 2, 3 The Firm built an internal delivery platform and put an "AI capability" line on every proposal. It was inoculation: the Firm turned the threat into a service line it could sell, so AI arrived on the books as revenue, and the price stayed where it was.

In the same two quarters it cut its graduate intake, the number of new graduates it hires each year, by fifteen to forty per cent, and stopped replacing the juniors who left. 7, 26 Internally that was "raising the bar". On the books it was the one lever that lifted margin at once and in silence: fewer juniors, the same price to the client, and the saved cost going to the partner draw, the money the owners take home. The platform cut what delivery cost the Firm 40✱ and did not cut what the client paid.

The lift arrives before any change to pricing. On hourly billing the Firm pays a junior for every hour of the year and bills the client only for the hours that land on a job. Cut the juniors and the whole salary goes, while the platform and the remaining staff absorb the billable hours. Paid hours fall faster than the fee. So the first phase of this story looks better than it is: a cost cut booked against an unchanged price, which lasts only until the client learns what the hours now cost to make.

2026 headline (composite, an invented line): "The Firm commits $3bn to AI, says graduate hiring will be 'more selective.'" The two halves of that sentence are the whole strategy.

The Buyer, the Firm's largest client, meanwhile commissioned a normal piece of work from her usual strategy firm. Behind a closed door she had two of her own people rebuild sixty to seventy per cent of the same piece of work with off-the-shelf AI tools, then compared the two finished versions. She told almost nobody, because she was gathering intelligence and no purchase hung on it.

She came away with two things. One was a document to put in front of her chief financial officer later. The other was a new reference price in her own head: what the routine seventy per cent of the work costs her company to make itself. Once the platform is built and paid for, each extra piece of that work costs her close to nothing. That reference price is the most important number in this scenario, and it never appeared on an invoice.

The Challenger was two to six senior operators, no pyramid, no brand. 12 It turned one painfully specific job into a fixed product: commercial due diligence for deals under thirty million dollars, ten business days, thirty-five thousand dollars fixed, money back if it missed. Commercial due diligence is the research a buyer commissions before it buys a company. Deals under thirty million dollars sit at the floor of that market, so the price comparison this report builds on this product is a floor case. Small-deal pricing shows the lowest price the work can carry and says nothing about the fees on large deals.

The guarantee came first, because a firm nobody has heard of cannot otherwise get past a buyer's fear of being blamed. Then the price, at twenty-five to forty per cent of what the big firms charge. A tenth would have branded the category cheap and left nothing to pay for building a reputation. A third sat in the one band the Firm could not follow, because the Firm still has to cover its juniors, its offices and its partner draw. I chose both numbers.

The first body to set rules for AI-heavy consulting was not a regulator. It was the insurer, and it did it by changing a price. 58

Every professional firm carries professional-indemnity cover, the insurance that pays the client when the firm's advice turns out to be wrong. A small firm buys it from an insurer and gets a renewal quote every twelve months. A firm the size of the Firm funds it through a captive: an insurer the Firm owns, which holds capital against the first slice of every claim, called the retention, and buys reinsurance for losses above that slice. This report assumes the Firm runs a captive. For such a firm the price of risk arrives as capital the captive has to hold and as the price and wording of the reinsurance above the retention, never as a quote. Reinsurers reprice that layer every twelve months, and they can carve a whole category of work out of it on their own. No statute, no scandal, no consultation.

Through late 2026 and into 2027, AI-heavy work started drawing new questions from the reinsurers and from the actuaries who set the captive's capital. A quiet argument began about whether "silent AI", a policy that never says whether AI-caused losses are covered, already excluded this work. 22, 23 In the game the Underwriter held the price flat where the Firm could name a licensed person who owned each output, and loaded unsigned AI work by fifteen to forty per cent or refused to cover it. For the Firm that landed as more capital locked in the captive and a dearer, narrower reinsurance layer: a cost on the balance sheet that never appears on an invoice.

So the Firm's own risk management became a tax on its AI strategy, and the brake was the chief financial officer, who owns the captive's capital. Selling unsigned AI work is now the expensive, personally risky path, and paying a human to check and sign the work is the cheap one. Reinsurers' loss models, meaning their own estimates of what future claims will cost them, set the working policy on AI in consulting for 2026 through 2028. The people who built those models did not think they were regulating anything.

The Firm stopped showing the client how many hours the work took. Proposals moved to fixed fees, outcome-based or "value-based" pricing, and managed-service subscriptions, which charge a standing fee for work delivered continuously 9, 45✱. All of it was sold as some version of "aligning our incentives with your outcomes." The billable hour, which means charging the client for time, had become the Firm's enemy. 57✱ Keep selling time and eventually the client asks why the task the AI just collapsed still costs two hundred hours. Cutting the link between price and input let the Firm keep the difference when AI cut delivery hours by sixty per cent. The honest name for value-based pricing here is hiding the cost collapse from the invoice.

The Firm set the trap for itself: once price no longer follows input, the client is free to ask what the result is worth to her.

The Buyer re-cut her outside contracts away from hours and towards results and access. She stood up a small internal team of four to eight people, light on seniority and heavy on tooling. She killed staff-augmentation spending first, before touching anything the board sees. Staff augmentation means renting consultants by the head to sit inside her own team. To the partners she was blunt: "I'm not paying analyst rates for analyst work anymore." She moved when her bargaining power was at its peak, and she cut the safest thing first.

The window was short, because the firms were frightened and had not yet folded their AI savings back into the old price. Her internal team kept the saved margin, and the knowledge that departing consultants used to carry out of the building in their heads.

The partners were reading bad numbers about their own business, because their own staff were hiding the real ones.

The Analysts became the most AI-fluent people in the building, and they hid it. 28, 29 The slide presentation that used to take twenty hours took four, and they paced the work to look like the old speed. Telling the truth meant either a smaller bonus pool or being the obvious candidate for the next cut. 50✱ The sharpest built reputations of their own on the side: industry teardowns, public templates, a personal brand that did not need the Firm's logo.

If they declare the speed-up, the gain belongs to the Firm. If they hide it, the gain is theirs. So the workforce absorbed the saving before it reached the Firm's accounts, which is why the measured return on AI lagged the demonstrations. 49✱, 37 The partners looked at their own numbers and concluded, reasonably, that AI was not changing their economics yet. That held until a challenger with no hidden slack to protect underpriced them in the open.

The Firm did build a unit of its own built around AI. It built it so that a failure could not hurt the core.

Separate brand, lower price, its own accounts, thinly staffed, and aimed at the work the Firm was about to lose anyway: benchmarking, diligence support, market sizing, routine modelling. 11, 43✱ The Firm walled it off and kept it out of the pricing conversation, so it could not set a reference price the partnership would have to live with. Taking your own work at a forty-to-fifty per cent margin beats a challenger taking it at thirty and climbing.

Starved of talent, of freedom to set prices, and of partner status, the unit stayed a toy, and it was not ready when the repricing arrived.

Then a case landed on a person, and it changed how every partner thought about risk. An AI-produced output turned out to be wrong: an audit, a valuation, a tax position, a filing, and nobody had actually checked it before it went to the client. The penalty fell on a licensed individual, and the AI, the firm's tool and the firm itself carried none of it. 46✱ A licence here is the professional body's permission to practise, the thing an accountant, an auditor or a lawyer needs in order to work in the field at all. Losing it, which the professions call being struck off, ends the career. That person was struck off, or sued, or both, and the insurance claim was contested because nobody could show a human had checked the work.

Professional bodies needed one case like it to turn the soft 2026 guidance, that the human remains responsible, into something felt in the gut. 25 "The AI did it" was no defence. Every consequential piece of work needs one named person who can be punished when it is wrong, and that person is the throat to choke this report is named for.

The Buyer made her suppliers compete directly against each other for the same work, every quarter.

A thin top layer of brand-name firms, kept for board-facing cover and for genuinely new thinking that no template can reproduce. Challengers built around AI underneath for everything operational. Both layers got the same brief, and she compared what came back.

The famous logo still protects the board from criticism, and that use is the last to fall. Bidding the identical brief put a dollar figure on that protection, so she could watch whether it was shrinking fast enough to justify moving a firm down a tier.

"AI gives you the analysis. We give you judgement, the board-defensible decision, regulatory cover, indemnity, and someone accountable when it's wrong." That was the Firm's new pitch, and this time it was honest about what it was defending. It leaned into audit, risk, regulation, restructuring and deals, the work where no model can carry the blame. 25, 48✱

That defended the top of the pyramid and did nothing for the middle. Salaried managers had risen by running teams of juniors, and half their value had been coordinating people who no longer existed. Those who could direct AI-heavy output survived. Those who had passed work between humans washed out. 13, 33, 54✱ The Firm was slow to react for an ugly reason: those people had no vote in the partnership.

Graduate hiring fell far enough to see from the street. 7, 26 The MBA bargain came under open question. An MBA is a two-year business degree, usually paid for with debt, and its promise was an entry-level job at a salary that services a hundred-and-fifty-to-two-hundred-thousand-dollar loan. Both ends of that promise were shrinking. 31, 32 The sharpest twenty-two-year-olds went around the qualification instead, towards AI skills, a specific industry, or building something of their own.

The path was always a financial bet: take the debt, and let the pyramid salary pay it off. When the salary and the exit job both shrink, the bet stops paying. The Firm had, in effect, chosen a 2035 partner shortage to fund 2027 margin, and partners are paid on this year's profit.

By 2028 the margin arithmetic had changed for good, and pricing could no longer hide it. So the Firm rebuilt the career model itself. 33, 54✱

Fewer but bigger engagements. One senior person overseeing AI-heavy output, where a team of juniors used to sit. 12 A thinner manager tier. A quietly reset up-or-out clock, the rule that says get promoted on time or leave, and a quietly lowered rate of admission to the partnership. The people pushed out were the partners' own former protégés.

Removing the salaried middle, whose billable hours no longer covered their cost, protected the partner draw. It happened late, because each cut admitted that the business model had changed.

Years of quiet hiring cuts turned into a sudden political problem, and nobody had planned for what came out of it. Job losses across the white-collar and graduate professions became a political headline, and politicians reached for blunt instruments: talk of hiring mandates, laws forcing firms to tell clients when AI did the work, training levies. Professional bodies scrambled to sell the qualification as a guarantee of trained human judgement.

The professional bodies also defended the junior analyst pyramid harder than anyone expected. Partly for their members. Partly for a reason they would never say aloud: fewer juniors trained now means fewer licensed humans later, and their fee income depends on a steady supply of them.

The trained judgement that walked out of the big firms re-formed into small firms.

Ex-pyramid talent built networked boutiques, meaning small firms that specialise in one thing, and part-time collectives. Pick one industry, turn the work into a fixed product, undercut the Firm on price, match it on quality, on a cost base of three people plus AI, against the Firm's thirty plus AI. 12, 13 The best moved up-market on purpose and added a thin premium line for a named accountable human.

The binding constraint had moved. It used to be capital, brand and an army of analysts. Once enough trained judgement had left and the tooling had become cheap, the constraint became trust and distribution. Most of these firms still failed the way small firms always fail: they could do the work and they could not be found.

The binding rules finally arrived in writing, and they surprised nobody who had been watching the insurers. 58 A named accountable human on the record. 25 A duty to tell clients when AI was involved. Audit-trail requirements. A recognised kind of professional-indemnity policy for AI-heavy work, with its own pricing tables. And the rules quietly redefined the qualification itself, from the person who does the work to the person accountable for the work, who certifies the AI did it correctly.

By 2029 the rule-writers had case law, and from the insurance market they had reserves, the money set aside against AI-heavy work, and prices. Settled claims they did not have. A professional-indemnity claim runs years from the wrong advice to the paid loss, so a 2029 loss is invisible in 2029 and the 2027 losses were mostly still open. The rules were written on reserving and pricing, and mostly ratified what the insurers had priced in 2027. The rules permit AI, the firms keep the productivity, and a licensed, insured, disciplinable human stands at the signing point of every consequential output. 46✱, 48✱ The pyramid survived by turning upside down, behind a wall of compliance cost the incumbents could afford and the challengers struggled to clear.

The Firm acquired the challengers that were winning the fixed-product work, at a price it justified in public and resented in private. The purchase removed a competitor that was setting the price, and bought the capability the Firm had strangled inside its own wall.

The surviving challengers had to choose a lane. Some stayed capped boutiques of fifteen to thirty seniors, took the cash out, and never raised money. Others turned their tooling into software other people could run, a bigger prize and a brutal fight against funded competitors and the AI labs. The ones who refused to choose died between the two.

The buyers landed on a standing rule. Do the recurring, embedded work yourself, as the largest companies already do through in-house delivery centres. 19 Save outside spending for the genuinely new, the politically dangerous, and the work that must come from an independent party. The market settled into three tiers: a few expensive accountable judges plus AI at the top, fixed products in the middle 11, 43✱, and the buyers' own internal teams underneath. The old mass-apprenticeship analyst job, the one that made all those partners, is largely gone.

You have read one path through. It was not inevitable. A handful of hinge points decide which world you wake up in, and they collapse to three endings. Pick the one your own beliefs support, then notice what you had to believe to pick it.

The Firm runs the harvest playbook end to end: alliance theatre, quiet intake cuts, fixed fees that keep the cost collapse off the invoice, and a walled-off unit it never lets win. The starting facts are real, the alliance announcements 1, 2, the intake cuts 7 and the pricing shift 9. The motives, and the walled-off unit, are my invention. The insurer and the 2027 liability case make protecting the human-review tier the cheap path, and the Challenger hits a ceiling at the regulated tier. The Firm buys the capability it starved and survives as a smaller, more senior accountability business, earning a higher margin per head. 33, 54✱

What it looks like: half the graduate intake of 2024, fewer partners, and the average partner draw defended or higher. Revenue lower, gross margin per head higher. A bought-in fixed-product unit runs the routine seventy per cent of the work. Buyers do the routine analysis themselves and hire the Firm only for the thirty per cent that needs a signature. The Firm defended whether clients hire it, and lost how much the work costs.

The belief it stands on: the AI labs never accept liability for professional-grade output, so a licensed human must. 48✱ If a lab or an insurer ever credibly covers model output, the signature stops being scarce, the work becomes cheap and interchangeable, and this ending collapses.

A handful of flagship companies very publicly run strategy in-house, with AI and a couple of ex-pyramid hires, and other buyers copy them. Buyers bring in-house anything they use more than twice. 19 The consulting line halves because demand went away, and the challengers won little of it. The Challenger's hardest competitor turns out to be the Buyer's own newly capable team, plus a flood of identical tiny boutiques with no distribution.

What it looks like: company strategy teams double in capability, outside spending falls forty to sixty per cent, and the long tail of staff-augmentation and research vendors disappears. The big firms keep the work that is genuinely new, too politically hot to hold in-house, or required to come from outside.

The belief it stands on: buyers build internal capability that lasts, and the hidden fixed costs stay manageable. Platform, data governance, security and the scarce senior people who run it all cost money. Near-zero cost for each extra piece of work has to survive company overhead landing on it. 20

Everyone overestimated how rational buyers are. Procurement inertia, long relationships with the big firms, the preference for a safe pair of hands, and the comfort of a big-firm name on a board presentation keep the money flowing long past the point where the Challenger's economics are obviously better. 17, 35, 55✱ A confidently wrong AI deliverable from a no-name boutique becomes a public disaster and pushes buyer trust backwards. AI stalls on exactly the judgement and taste tasks the pessimists assumed it would take.

What it looks like: the big firms adopt AI to fatten margin while holding price, and largely get away with it through 2030. Challengers win clients and scrape by, capped by how much senior time they have. The analyst who stayed and served the apprenticeship inherits a thinner but intact firm at higher pay.

The belief it stands on: trust, risk transfer and board psychology are stickier than the economics. Buyers keep paying the brand premium because they were always buying defensibility, someone to blame and political cover. AI made the analysis cheap and left those three alone. If the buyer's reference price resets the moment "AI did the analyst work" becomes consensus, this flips to A or B fast.

Everything above is written as memory, and none of it has happened. Two dates matter here. The real-world snapshot behind this report was taken on 1 July 2026, and that is the "now" the story looks forward from. The text was last revised on 8 September 2026, which is why the byline carries the later date. Next year's graduate intake decision is live on somebody's desk right now. The questions reinsurers ask above a captive's retention have not been redrafted. The canonical liability case has not been filed.

The future in this document is a fork, and every branch is still open. Choices that look administrative are deciding, this quarter, which ending you are walking towards. A hiring number. A change to a pricing model. Whether the AI unit reports to somebody with the power to make it win.

I built this by having an AI play the five roles and by refereeing the result myself. Here is where I do not trust it, including where I do not trust it because it is an AI.

The whole thing flatters human judgement, and an AI wrote the first pass. If AI keeps getting better at judgement, the one thing every role in this report assumes AI cannot do stops being true, and two of the three endings above fall over. All five roles land on the same conclusion, that human judgement, accountability and training juniors under senior supervision are the one advantage a competitor cannot copy. That is suspiciously comfortable, and it is the least trustworthy load-bearing claim in here. The three abilities at stake are framing a problem well, judging what counts as good work, and catching your own confidently wrong output. An AI trained to sound authoritative will tend to say those belong to humans, and so will I, because judgement is what I sell. I have left the claim in because I think it is more right than wrong over this window. How much judgement is genuinely irreducible I could not settle, and nor could the machine.

I chose every date after 2027 for the telling. The 2027 liability case, the 2028 political shock, the 2029 written rules: that sequence is clean because I made it clean to tell. Real case law and legislation are lumpy and differ by country. I am fairly confident in the order, insurer first and regulator second. The timing could compress into eighteen months or stretch past 2030, and the claims lag pulls it later, because settled losses arrive years after the reserves and prices the 2029 rule-writers see.

The captive is an assumption. This report assumes the Firm funds its professional-indemnity cover through a captive it owns, with reinsurance above a retention, as large firms commonly do. The pack holds no claim on it. If the Firm bought cover on the open market instead, the same repricing would arrive as a renewal quote, faster and in plainer sight.

The Buyer is the figure most likely to be modelled wrong, in the same direction, by everyone. An AI reasoning about buyers is biased towards assuming they are rational, price-sensitive and capability-building. Each of the five AI-played roles, asked where it was wrong, said it had assumed buyers were more rational than they are. Naming the bias did not remove it. Procurement inertia, board psychology and the value of an outside firm as political cover are the three most likely to be under-weighted across all five roles at once. 17, 20 A room of real executives already knows how its own board behaves, better than I or the machine can guess.

The challenger wave might be a thousand identical boutiques that mostly fail. The roles assume a challenger can build reputation and distribution fast enough to matter, then confess that no distribution is how small firms die. I cannot settle whether the challenger surge is a structural shift or a churn of look-alikes. 34 Do not treat "challengers reset the reference price" as inevitable.

Every number without a marker is an archetype: a figure I invented to be internally consistent, with no measurement behind it. They include the ratios of juniors to partners, the sixty-to-seventy per cent the Buyer's team reproduced, the seventy-thirty split between routine work and signature work, the fifteen-to-forty per cent insurance repricing, the margins on the walled-off unit, the slide presentation that took twenty hours and now takes four, the capped boutiques of fifteen to thirty seniors, the halved graduate intake in Ending A, and the forty-to-sixty per cent fall in outside spending in Ending B. Those numbers are internally consistent because I wrote them to be. None is benchmarked to your firm. Replace every invented figure with your own real numbers before you act on any of this.

In eight places the checked evidence pushes back.

The insurance repricing is a direction of travel. Today's pricing runs the other way. The 2025-26 professional-indemnity market is soft, which means insurers are competing on price. An Australian specialist broker reports discounts of five to fifteen per cent for firms with clean claims histories 21. Premiums benchmark at roughly a quarter of one per cent up to five per cent of fee income 39. My fifteen-to-forty per cent repricing is therefore a multiple applied inside a small band of fee income, and the real money is smaller than the headline reads. The mechanism is loading: by May 2026 three carriers were seeking approval to exclude AI-driven losses from their wordings 58, and "silent AI" is a named debate in insurance-law commentary 22, 23. If the soft cycle persists, the tax lands later and lighter than my timeline needs.

The real challengers price lower and promise less than mine. My Challenger charges a third of what the big firms charge, and matches quality. The best-documented live one sells commercial due diligence at about fifty thousand US dollars against a half-million-to-a-million-dollar incumbent job 11, and positions itself as "80% of the insight at 5% of the cost" 42✱. That is two and a half to eight times cheaper than my band, depending on which ends you compare, and it does not claim to match quality. Nobody in the pack publishes my money-back guarantee 43✱, and my story leans hard on it. The closest real analogue is the incumbents' own outcome-linked pricing 45✱. The most capitalised real challenger can draw up to three hundred million US dollars of investment from Warburg Pincus, and it aims at large mid-sized companies, well above the cheap end of the market 41✱, 10. My own comparison starts from a deal under thirty million dollars, small-deal pricing at the floor of the market, and says nothing about large deals. Capital has not stopped being a constraint. The pricing move is the weakest-evidenced one in this scenario.

The boutiques may not fail where I say they fail. I have the challenger wave dying of no distribution. Trade coverage of productisation records a different failure mode: independents now lose on whether their delivery is demonstrable, structured and traceable, even when the raw quality is there 44✱. The binding constraint may be a provable method plus visibility.

In consulting the blame still lands on the firm. Courts have sanctioned individual licensed attorneys for AI-fabricated citations, including disqualification from a case 46✱. But the highest-profile consulting incident to date landed on the firm as a commercial remedy. Deloitte Australia partially refunded a government report containing AI-fabricated content, and nobody was struck off 24. My canonical case assumes the penalty migrates from the firm's invoice to the individual's licence, and in consulting that has not happened yet.

Buyers are running the other way. The share of clients paying outside consultants for AI support rose from 81 per cent to 88 per cent into 2026 17. Strategy, which is board-facing work, was one of only two service lines to shrink in 2023, the opposite of my staff-augmentation-first cut order 18. And Gartner found that 80 per cent of large companies using autonomous AI tools had cut staff without getting the returns they expected 20. The cheap internal engine my Buyer builds is, on today's evidence, neither cheap nor quick.

The measured time saving is half the figure I assumed, and uneven. My timeline assumes delivery hours fall sixty per cent. The best-documented figures are McKinsey's own claim of up to 30 per cent of time saved by its Lilli platform, on searching and synthesising 40✱, and a field experiment on 758 consultants, published in 2023 and run with GPT-4, that measured 25.1 per cent faster completion 14. In that experiment the consultants using AI were 19 per cent less likely to be correct on tasks outside what the model is actually good at, which the study calls its competence frontier, and 23 per cent worse on the judgement-heavy task 15, 16. Real, concealable compression exists, and the same experiment rated the AI-assisted work 40 per cent higher in quality 30. My figure still runs ahead of the evidence, and the gains are jagged: AI helps a lot on some tasks and hurts on others. That 2023 experiment is the best measurement in the pack and the oldest, and the pack holds no later measurement of the same design.

The hiring collapse is uneven across the profession. KPMG cut UK graduate intake by about a third 7, 26 and Deloitte by about a fifth 27, but EY and PwC trimmed by single digits to low teens, below my own fifteen-to-forty band 8. BCG grew, with 7 per cent revenue growth in 2025, its twenty-second consecutive year, and headcount up to 33,500 55✱, 56✱. My Firm is a composite of the steepest cutters. If the political clock fires, it fires on an uneven, contested labour signal.

An affirmative AI indemnity already exists. Ending A stands on nobody credibly indemnifying model output. Since April 2025 a Lloyd's-market product co-developed by Armilla and the underwriter Chaucer has affirmatively covered hallucinations, meaning AI inventing false information that reads as fact, along with model drift, meaning a model's answers quietly getting less reliable over time, and related AI failures 47✱. Microsoft's Copilot Copyright Commitment already indemnifies commercial customers for copyright claims, and says nothing about whether the output is correct 48✱. The limits are small today, but the seed of the thing that collapses Ending A is being sold.

This last section is about the assumptions your own team is betting the firm on without having said them aloud.

You are treating clients value trust and accountability and clients will keep paying the old price as one claim. They are two claims. A reputation for judgement protects whether clients hire you, not what the work costs. 11, 42✱ So ask, out loud: if buyers confine us to the thirty per cent that truly needs a signature, is that thirty per cent a viable business at our cost base?

You cannot starve your new capability to protect today's revenue and also have it ready when you need it. Ask who inside is actually allowed to make the new unit win, and whose pay that win comes out of.

Your binding constraint over the next two years may be the capital your own captive has to hold against AI-heavy work, and the price of the reinsurance above it, long before any regulation arrives. 58, 22 Has anyone worked out your insurance and your compliance cost as a bill that rises with how aggressively you use AI? Is your chief financial officer already the real brake on your strategy, and do you know it?

The fastest erosion may come from your own customers building the capability in-house, ahead of any competitor. 19 How much of your current revenue is work a client could do itself the day it has the tooling and one good hire?

And the quiet existential one. Cutting your intake of juniors gets booked as an efficiency move. It may also be a decision to trade a 2035 shortage of senior people for 2027 margin, on the assumption that senior judgement can be bought later. 7, 27 Where do your 2035 leaders come from if you saw off the ladder below them now?

You answer those by putting four to eight of your own people in a room and forcing the disagreement into the open. Give somebody in that room the job of attacking the comfortable answer.

That is what the Workshop is. Half a day. You bring the decision that is keeping you up. I bring the scenario and the red team. We run your assumptions until the ones that cannot survive contact fall over, in a room on a Tuesday, before the market runs the same test over three years.

And if a workshop is not the right tool, I will tell you. Sometimes the honest answer is that you have an execution problem, or a data problem, or a nerve problem, and half a day of structured disagreement would just be theatre. The exercise on this page is the free version. The real value is running it on your own decision, before the fork closes.

- Dr Dan Epstein, The Long Game Project

v0.14 - 8 September 2026. Argument pass after a Big Four partner cold-read the essay: the composite Firm split by trait, the insurer beat rebuilt around captive funding (confessed in the seams), the intake-cut margin mechanism added to the first beat, the sub-thirty-million-dollar deal marked a floor case, the 2029 beat qualified for claims lag, the 2023 GPT-4 experiment dated at the point of use, seven undated Sources rows marked so, and contrast-shaped sentences cut from about thirty to two. No number, marker or heading changed.

v0.13 - 6 September 2026. Plain-language rewrite of the whole document against the house clarity standard, so a reader from another field can follow it without knowing how professional-services firms make money. Every technical term is now defined in the sentence that first uses it, the contract opens by saying plainly what this document is and how it was made, and metaphor, epigram and closing riddles are cut throughout. No number, date, name, claim or marker changed, the Sources table is untouched, and every section heading keeps its wording, while ten beat titles are made plainer with every date prefix unchanged. A fresh-reader pass then added the missing definitions of a professional licence, being struck off, fork, archetype and soft market, stated in plain words what the title means, and separated the five machine-generated actor-plays (incumbent, challenger, buyer, labour, regulator) from the five figures who carry the story.

v0.12 - 5 September 2026. Editorial pass over the whole document before republication, against the house prose standard. No claim changed and no Sources row, number, date or beat heading was touched. Four things beyond prose did change. Two markers moved off the two-tier panel sentence, which is an archetype: the claims they pointed at record the live challenger as aimed at the upper mid-market, the opposite of the tier the sentence puts it in, and the collision on real challengers already carries them. A two-sentence reading convention now sits in The contract, so the reader meets it before the first unmarked number rather than in the seams. The eight collisions sit under their own heading, Where the evidence already cuts against me, the fixed section the house standard requires. And the fork no longer promises six hinge points it never lists. The beat opener that had become a metronome ("X happened, and here is the turn") is broken up in five beats, two beats now end flat on a fact instead of a closing line, and single-word italic stress drops from twenty-two instances to eight. Passives with a known actor, filler adverbs and two self-announcing asides are cut. Three beats are re-blocked so their shape is no longer uniform. The "you cannot throttle it and also have it ready" line was stated twice in nearly the same words and is now stated once, in the So what section. Editorial pass by the editing assistant, awaiting author review.

v0.11 - 26 August 2026. The master text re-synced to the v0.10 interactive edition, so the site now renders the current prose from the text and the evidence pack. No claim changed and no citation moved. Figures are re-set as standard charts of verified pack data (the intake cuts, the hiding surveys, the McKinsey people-and-agents configuration, the MBA on-ramp cost, the buyers-going-the-other-way slope), with the timeline of the five deciding moves placed in the contract. Extraction and re-set by the editing assistant, awaiting author review.

v0.10 - 7 August 2026. Figure density restored the conventional way: four figures added from unplotted evidence-pack data - the measured AI-hiding surveys, the McKinsey 40,000-people-25,000-agents configuration as a typeset figure, the MBA on-ramp cost, and the buyers-going-the-other-way slope placed inside the collisions section. Nine figures total. No invented icons anywhere.

v0.9 - 7 August 2026. Symbol-stripping pass: the four remaining v0.6 pictorial figures deleted, the side-rail standings and watch-signal blocks removed, the timeline strip stripped of its lightning-bolt icons, and the abstract price-position chart cut. The invented game-token icons are retired everywhere. The page now carries prose, ruled typography, and two conventional charts of verified data.

v0.8 - 7 August 2026. Figures rebuilt as charts first: three charts of verified evidence-pack data (the intake slopegraph, the price positions, the jagged frontier) in the house style, three structural figures re-set as typeset plates, and the freeform pictorial drawings retired.

v0.7 - 7 August 2026. Prose re-synced to the v0.5 text after a restraint pass, and five static figures placed inline.

v0.6 - 31 July 2026. Readability pass across both the interactive edition and its markdown source, so the two stay in sync. No claim changed and no citation moved. Every semicolon in the narrative prose is gone (20 of them), replaced by a full stop or restructured, including three in the beat headings. The worst chained sentences were split, including four changelog entries that had become harder to read than the report. One marketing adjective went ("bespoke pyramids" is custom-built pyramids). The 58-claim evidence pack, the source tiers and the amber shading for authored numbers are untouched.

v0.5 - 20 July 2026. Design revision of the interactive edition, same day. The first interactive skin failed its reader test: boxed citation chips, badge rows and a card-stack dashboard rail read as machine furniture rather than editorial design. This pass rebuilds the presentation as a print-first system. Citations are set as superscript markers, the side rail as ruled margin notes, and figures are redrawn in ink with teal reserved for verified fact and amber for authored uncertainty. Reading-convention wording updated from "brackets" to "markers" to match. Prose otherwise unchanged from v0.3.

v0.4 - 20 July 2026. Interactive edition. Presentation remake under the house report standard: a side rail that moves with the scroll (the timeline, key numbers per beat, standings, signals to watch), three pictorial figures including the liability-flow centrepiece, and a fork selector with linkable endings. The prose is unchanged from v0.3. Every number on a figure is either cited or marked as an archetype.

v0.3 - 19 July 2026. Production hardening pass, off the back of an adversarial pre-publication review. The reading convention now says exactly what a bracket verifies on a composite sentence: the pattern, not the fictional actor or the exact figure. Three citations were re-pinned to their seed facts, so fictional motive language can never read as a verified claim about a named firm. The reclassified-spend line is now explicitly the EY-documented pattern, the "aligning our incentives" pitch line is marked as a composite, and Ending A's playbook cites are split per move. The one real firm previously named inside the fiction (Ending B) is now the composite Firm. An eighth evidence collision was added - the boutiques may fail on provable method, not visibility alone 44✱. The audit-rigour wording in the v0.2 entry and the Sources preamble was corrected to what the pack actually records: dated live re-checks on a ten per cent sample. The pack records no human audit, and claiming one was exactly the kind of overreach this document exists to kill. A copy-edit pass ran throughout.

v0.2 - 6 July 2026. Grounding pass. The scenario's real-world signposts are now bracket-cited to an adversarially verified 58-claim evidence pack. Every claim carries a verbatim source quote verified at retrieval, with dated live re-checks on a ten per cent sample recorded in the pack. The Sources appendix at the end lists every cited claim. Each citation survived an adversarial check that it supports the specific sentence it sits on - about a third of the proposed citations did not survive and were cut. The seams section gains the collision list: seven places where the evidence already cuts against the scenario. One correction. The codification beat previously name-checked Chubb as the 2027 price-setter. No evidence of a Chubb AI-PI product was found, and the documented pioneers are Chaucer/Armilla and Munich Re, so the line now reads "the carriers". Signpost watch from v0.1: the first insurer renewal language is arriving - three carriers seeking approval to exclude AI-driven losses by May 2026 58. The first consulting liability event exists at firm level only - Deloitte Australia's partial refund 24. The publicly run in-house strategy function has not yet been observed.

v0.1 - 1 July 2026. First public draft. Single-run exercise: five AI actor-plays, one human referee. Endings ranked by my prior, not by evidence. Known soft spots flagged in the seams above - chiefly the self-flattering "human judgement is the moat" claim, all post-2027 dates, and buyer-rationality assumptions. This document will be updated as reality moves: the first insurer renewal language, the first canonical liability case, and the first publicly run in-house enterprise strategy function are the three signposts I will watch to decide which ending we are actually walking towards. If you are reading a later version, the dated beats that have since resolved are marked.

Every marker in the text resolves to a verified claim below - 54 of the pack's 58 claims are cited. A claim marked * holds with a recorded caveat. Full provenance, verbatim quotes and retrieval dates live in the evidence pack (throat-to-choke-v1).