What changed
Two things happened at once, and they are usually discussed separately. Generative AI moved from pilot to production across professional services and inside the organisations those firms advise. And the entry-level intake that has always fed the profession contracted. The connection is widely assumed and hard to prove — firms cut for several reasons at once, and cost is rarely only one of them. The causation is not what matters here. What matters is that the work junior people used to do is now substantially done by a model, in firms and in-house alike, and that almost nobody has thought carefully about what else that work was doing.
Why it matters
Every apprenticeship profession ran the same machine. Accounting, law, consulting, banking, engineering: juniors produced, seniors reviewed, principals signed. The machine delivered the work, and — incidentally — it manufactured expertise. The only way to learn which evidence to discard, or to see that a question is wrong before anyone starts answering it, was to produce a hundred pieces of work and be corrected on them. Training was a by-product of production. Nobody designed it. It fell out of the model.
AI removes the production. It does not replace the by-product.
The response so far has largely been to substitute one input and keep the process: the junior now produces in AI, the senior still reviews. That looks like continuity. It is the most consequential mistake being made across advisory right now, and it is being made in-house as often as in firms.
A model amplifies whoever is holding it. Given a well-framed question it produces work of real quality — faster and cheaper than the pyramid ever managed, and often better. Given a poorly framed one it produces something equally polished, equally confident, and wrong in ways that do not show in the output. It has no view on whether the question it was handed was the one that needed answering, whether the data it was given excludes the cases that would have changed the conclusion, whether the authority it is relying on still stands, or whether the assumption the whole thing rests on describes the business as it actually operates. It answers what it was asked, well.
Nor is a good answer one instruction. It is a few hundred small decisions — push back here, concede there, discard that line, chase this anomaly, re-ask the question because the first framing was subtly wrong. Knowing when to push and when to concede is the hardest judgement in the exchange, and the one that tracks experience most closely. An inexperienced operator concedes early, because the answer is fluent and they have no basis on which to doubt it. That is not a training gap that closes in a year. It is the thing the hundred corrected files used to buy.
Then there is what review can no longer catch. Review worked because the reviewer could see the working. Someone’s draft showed what was consulted and what was set aside; a wrong turn three steps back was findable. AI-assisted work arrives finished, and the two hundred decisions that produced it are gone. The reviewer sees a coherent document and checks it for internal consistency, arithmetic and presentation — all of which pass. What is not visible is the framing error at the start, the exception that was never surfaced, or the question nobody thought to ask. Review has become a check on the artefact rather than on the reasoning, and the errors that matter live in the reasoning.
So the answer is not more review. It is that the person holding the judgement has to be the one in the exchange, making the calls as they arise rather than reading the transcript afterwards. That is a different shape of organisation from the one professional services spent a century building — and not one a leveraged partnership adopts easily, because its economics require the leverage that has just become unnecessary. The constraint is structural, not cultural. It will not be fixed by a strategy document.
There is a pricing model buried in all of this, and it does not survive either. The pyramid ran on the billable hour — value proxied by time, time priced by grade. It always rewarded the wrong thing: the slower the work, the larger the fee, and efficiency was a cost the firm absorbed rather than a service it sold. The arrangement held only because hours were a passable proxy for effort, and effort for value. AI cuts both links at once. When a week’s work arrives in an afternoon, hours measure nothing a client should pay for; and hours-by-grade — the entire arithmetic of the pyramid — charges for a structure that is no longer doing the work. The profession has been slow to say this out loud, because saying it unpicks how it charges.
We should be plain about our own position, because the argument implicates it. ajiho exists in its current form because of AI. A practice of this shape could not have been built ten years ago — the work would have required a team, a team requires a firm, and a firm requires the pyramid described above. That is no longer true, and the consequence for clients is not a discount on the same product. It is a different one: the senior person doing the work rather than reviewing it, in a fraction of the time the old model needed, priced on the answer rather than the hours behind it. Not cheaper hours — no hours. The trade-off between quality, speed and cost dissolved at the point the expert took the keyboard, because the thing that created it — time billed by grade — had stopped describing the work.
The short-term cost is real, and it is carried by people at the start of their careers who did nothing wrong. It is worth saying that the same collapse in the cost of production that removed those roles also removed the barrier to building something alone. What that generation makes will look very little like what it was going to be hired into, and is likely to be considerably more interesting.
What this means for your business
- Ask who did the work, not which firm did it. The relevant question is no longer how many people were on it. It is how much experience was present where the analysis was actually made.
- Apply the same test internally. Wherever your own teams have put AI at the bottom of a process and kept sign-off at the top, you have the problem described here — in finance, legal, procurement and treasury as much as anywhere else.
- Test the reasoning, not the document. Ask why that was excluded, why that assumption, why that conclusion. Work that cannot answer a question will not answer a challenge either.
- Treat a clean-looking output as neutral information. It used to be weak evidence of care. It is now evidence of nothing at all.
- Stop paying for time. If your advisers still bill by the hour and by grade, you are funding a structure AI has hollowed out and rewarding the slowest route to an answer. Ask what you are buying — the hours, or the judgement.
Sources
ajiho analysis. No external data is relied on in this piece. No client or confidential information is used in any Signal publication.