
The Case of the Disappearing Margin
AI now does the junior work that big accounting and consulting firms bill by the hour. Who keeps the saving, the firms or their clients?
Report 001 · Dr Dan Epstein · version 0.17 · 12 October 2026
A Strategy Soup Scenario What is an exercised scenario?
The question
Professional services means the firms that sell thinking: accountants, auditors, lawyers and management consultants. A big one is shaped like a pyramid. A few partners, the senior people who own the firm, sit at the top, and many junior analysts sit at the bottom. The profit comes from billing junior hours at several times what the juniors cost, and the trade calls that bill the billable hour. AI is now good at much of the work the juniors do. So who keeps the saving, the firms or their clients?
Inside a firm the question sounds like this. When AI makes analysis cheap to produce, do we protect the pyramid that pays us, or undercut our own prices before somebody else does?
Our answer: in the end, the clients. In our story the firms keep the saving at first, by hiring fewer graduates and leaving the price where it was. Each job has two parts: routine work, about seventy per cent of it in our story, and the part a named, licensed person must check and sign. The clients’ own staff learn what the routine part costs to make, do more of it themselves, and stop paying the old fee for it.
The signed part stays with the firms, at a price. Their insurer charges more to cover unsigned AI work, meaning AI work that no named, licensed person has checked and signed. So the firms keep paying people to sign. A licence is a professional body’s permission to practise, and the person who signs can lose it if the work is wrong. That person is the one throat to choke.
How the simulation works
The simulation has five roles: the big firm, the challenger, the buyer, the workforce, and the insurers and regulators as a single role. The story has one character for each role, listed below.
The story runs as seventeen dated steps from 2026 to 2030, told from 2030 looking back. Nothing was decided by chance and the story was played once, so there are no counts of how often anything happened.
The real-world record stops on 1 July 2026, and every date after that is the story. Some of its late-2026 dates have now passed, and we have not checked the story against them. Every fact from the public record is checked against a source, and the deep mode marks each sentence that rests on one. A number with no source is an assumption we wrote to fit the others. Four such scales run through the story, listed here and drawn in Fig. 2 so you know they are ours.
ROLES IN THE SIMULATION, ONE OF THEM FOR INSURERS AND REGULATORS TOGETHER
5
POSSIBLE ENDINGS IN THE DEEP MODE, RANKED BY OUR OWN JUDGEMENT
3
DATED STEPS IN THE STORY, 2026 TO 2030, TOLD FROM 2030 LOOKING BACK
17
PLACES WHERE THE CHECKED EVIDENCE ARGUES AGAINST THE STORY, ALL LISTED IN THE DEEP MODE
8
In our story the Buyer began taking the saving, and the insurer began charging more for AI work nobody signed, in 2026 and 2027. The written rules came only in 2029.
The Firm keeps the payroll saving: fewer graduates, the same price
The Firm opens with a loud move and an unannounced one. The loud move is an alliance with a leading AI lab and a cloud provider, under an “AI investment” headline in the billions. Most of that number is existing technology spending relabelled. The real firms did this: EY’s own 2023 release describes its US$1.4 billion as money already spent. The second move is the cut. In 2026 the Firm hires fifteen per cent fewer graduates than in 2025, fifty-eight per cent fewer by 2028, and does not replace the juniors who leave. We chose those figures. Among the real firms the steepest one-year cut was KPMG’s, about a third, in 2024.
Fewer juniors at the same price means the money saved on salaries goes to the partner draw, the money the owners take home. The Firm’s new AI platform cuts what the work costs, not what the client pays, so no invoice explains the gap. Then it moves its proposals to fixed fees and subscriptions, so the invoice stops showing how many hours the work took. In the story that hides a sixty per cent fall in delivery hours, a figure we chose. The real McKinsey already earns about a quarter of its fees from prices tied to results rather than hours.
The Firm also builds a cheap unit of its own, with a separate brand and a lower price. It walls the unit off so a failure cannot hurt the core. Starved of talent, pricing freedom and partner status, the unit stays a toy. It is not ready when the insurer changes its prices in 2027.
The Firm’s own staff keep the time they save, so the partners cannot see it
The Analysts become the most AI-fluent people in the building, and they hide it. In the story a slide deck that took twenty hours takes four, and they pace the work to look like twenty. Owning up means a smaller bonus pool or being the obvious candidate for the next cut. The sharpest build reputations of their own on the side.
So the time the Analysts save stays with them and never shows up in the Firm’s accounts. This is a different saving from the payroll cut above, which the partners do keep. The partners look at their own numbers and conclude, reasonably, that AI is not changing their economics yet. That stays true until a challenger with no hidden slack to protect underprices them in the open. The hiding has real support: two large 2024 surveys found about half of workers reluctant to admit they use AI (Fig. 4).
The insurer writes the first rule on AI work, by changing a price
Every professional firm carries professional-indemnity insurance, the cover that pays the client when the firm’s advice turns out to be wrong. We call it the insurance from here on. The insurer reprices it once a year, at the renewal, and can refuse to insure a whole class of work. No statute and no consultation is needed.
Through late 2026 and 2027, in the story, unsigned AI work draws new questions at renewal. An argument starts over “silent AI”, a policy that never says whether AI-caused losses are covered. That argument is real and already named in insurance-law commentary. The Underwriter keeps a firm’s insurance premium flat when the firm can name the licensed person who owns each output. It raises the premium on unsigned AI work by fifteen to forty per cent at the 2027 renewal, a band we chose, or refuses to cover that work. Real insurers today give discounts instead (Fig. 5).
The rise is a small sum, because this insurance costs between a quarter of one per cent and five per cent of a firm’s fee income. The refusal is what stops the Firm. If the insurer will not cover a piece of work, every loss on it falls on the Firm. So selling unsigned AI work becomes the expensive and personally risky path, and paying a human to check and sign the work becomes the cheap one. The person who slows the Firm’s AI plans is its chief financial officer, reading the renewal.
Then in 2027 a wrong AI-produced piece of work lands on one licensed person rather than on the Firm. Management consultants need no such licence, so the case lands on the accountant, auditor or lawyer who signed the work. That person is struck off, meaning the licence is taken away and the career ends. “The AI did it” is no defence. Every consequential piece of work now needs one named person who can be punished when it is wrong. That person is the throat to choke.
The Buyer learns what the work costs and takes the fee back
The Buyer commissions a normal piece of work from the Firm. Behind a closed door, two of her own people rebuild sixty to seventy per cent of it with off-the-shelf AI tools, a share we chose. She compares the two versions and tells almost nobody. What she gains is a reference price: what the routine part of the work costs her company to make itself. That price never appears on an invoice, and it is the number that later cuts the Firm’s fee.
She moves while her bargaining power is highest. She re-cuts her contracts away from hours and towards results, and cancels first the consultants hired by the day to sit inside her own team. She builds an internal team of four to eight people.
In 2027 she splits her supplier list into two tiers. Brand-name firms sit on top, kept so the board sees a famous name on the work, and challengers underneath do the routine work. She gives both tiers the same brief every quarter, which puts a dollar figure on what the famous name is worth.
The Challenger sells commercial due diligence, the research a buyer commissions before it buys a company. The Challenger sells it for small deals, under thirty million dollars, at a fixed thirty-five thousand dollars, with the money back if it misses its deadline. We chose that price and set it at about a third of a big firm’s fee for the same small-deal job, which implies a big-firm fee of about a hundred thousand dollars. We derived that fee and did not check it. The Challenger wins the routine work she does not do herself.
By 2029 the written rules arrive, and they mostly ratify what the insurer priced in 2027. They name an accountable human on the record, make firms tell clients when AI was involved, and recognise a kind of insurance for AI work. The Firm buys the challengers that won the fixed-price work, at a premium. The market settles into three tiers. A few expensive accountable judges plus AI sit at the top, fixed products in the middle, and the buyers’ own teams underneath.
A famous name keeps the clients. It does not keep the old price.
What it means
In our story the clients end up with most of the saving. On scales we chose, the fee the client pays per job halves by 2030, and the Buyer’s company does 42 per cent of its advisory work itself, up from 4. The routine seventy per cent of each job goes to the client’s own staff or to cheap fixed-price firms. The Firm keeps the other thirty per cent, the part a named person must sign, and its partners keep their income because far fewer people share the fee. For a real firm the question is whether that signed thirty per cent of the work is a viable business at its cost base.
What this does not prove
The story was played once and nothing in it was decided by chance, so nothing was counted. We cannot tell you how often any of it happens. The deep mode ends on three possible endings, ranked by our own judgement. In the first, which we think most likely, the big firms survive smaller and more senior, behind compliance costs the challengers cannot afford. In the second the buyers do the work themselves, and in the third nothing much changes through 2030.
The most flattering claim is also the least trustworthy. All five roles conclude that human judgement, accountability and the apprenticeship, meaning training juniors under senior supervision, are the one advantage AI cannot copy. We will tend to say those abilities belong to humans, because judgement is what we sell. If AI keeps getting better at judgement, the whole story falls over.
Every number without a source is one we chose. That includes the intake cut, fifteen per cent in 2026, fifty-eight by 2028 and two-thirds by 2030, the sixty per cent fall in delivery hours, the seventy-thirty split between routine and signed work, the insurance repricing, and the four scales above. The checked evidence pushes back in eight places. Two are already on this page as figures, the insurance discounts and the buyers paying for more AI help. The largest of the rest is the time saving.
Five more places push back, starting with the real challengers, which price lower and promise less than ours. One sells due diligence at about US$50,000, against US$500,000 to a million from the big firms. That is a tenth or less of the big firms’ large-deal price, where our Challenger charges about a third of the small-deal fee we assume. Nobody publishes our money-back guarantee. In consulting the blame still lands on firms. Deloitte’s Australian firm partly refunded a US$290,000 government report with AI-invented references, and nobody lost a licence.
The hiring collapse is uneven, since EY and PwC cut by single digits and BCG grew its workforce to 33,500 in 2025. An insurer already covers AI failures. Since April 2025 a Lloyd’s product from Armilla and Chaucer has covered hallucinations, meaning AI inventing false information that reads as fact. And the small firms in our story die of not being found. Trade coverage says independents now lose on proving their method, not on the quality of the work.
What to watch
Six real-world signals, and where each one shows up.
The renewal form changes before any law does, and it is the thing to read first.
The record
Each chart is marked as real figures or as scales we chose.
In the story the fee the client pays halves, while the share of the work she does in-house rises from 4 to 42 per cent.
The real cuts to graduate hiring are steep and uneven. Our Firm’s first cut, fifteen per cent, sits between Deloitte’s and EY’s.
In two 2024 surveys about half said they were reluctant to admit using AI. Our Analysts hide their speed-up the same way.
The story needs insurers to raise the price of unsigned AI work. The real market is still cutting prices.
Real buyers are paying consultants for more AI help each year.
We assumed AI cuts delivery hours by sixty per cent. The measured figures are 25 to 30.
Want the evidence - the dated record, the sources, and our own list of weak points?
Read deep mode