The cut came before
the capability.

I spent a day this week refiling our own accounts. Not a project — housekeeping. By the end of it I had changed my mind about how much time this profession has.

What 500 CFOs actually said

The survey is worth reading slowly, because most coverage picked the wrong headline.

Eight per cent of these CFOs have AI tools or autonomous agents deployed at scale in finance. Eight. Meanwhile 64 per cent say the function is shifting away from junior roles, and 91 per cent expect finance headcount to be flat or lower — 61 per cent flat, 30 per cent down.

Read those together. The technology is not yet doing the work. The seats are being removed anyway. The cut is not following the capability. It is anticipating it.

At the other end of the same org chart, the senior people are leaving on their own. Sixty per cent of CFO departures in the first half of 2026 were retirements or moves to board seats, against an eight-year average of 41 per cent. At the largest listed companies it was 69 per cent.

So: the most experienced are walking out the front door, the entry seats that would eventually replace them are being closed, and the machine that is supposed to cover the gap is running at scale in fewer than one company in ten. All in the same budget year.

Six months ago it could not read a receipt

Here is my own evidence, and it is the reason I now think the 64 per cent are early rather than wrong.

Six months ago I tested AI on our own paperwork and it could not reliably read a petrol receipt. Not a hard task. It failed it.

This week I put 369 documents through it — bank statements, utilities, digital subscriptions, withholding tax, receipts, agreements, a vehicle file. Read, named to one formula, filed into a numbered structure, receipts tied to the bank transfers that paid them, statement lines matched back to the documents behind them. One day. A handful of small errors, all of them caught by looking.

The part that actually changed my mind was not the accuracy. It was this: once every document had a standard name, the rest stopped being intelligence and became arithmetic. Matching a statement line to a receipt was a rule, not a judgment. No clever model required. The hard part had been the naming, and the naming is now automatic.

That is the whole profession's exposure in one sentence. The middle of accounting was never difficult. It was just expensive, because a human had to look at every document.

What actually disappears

Three things survive, and they are not the three people assume.

Capture survives — the receipt in someone's pocket, and the one line that says what the payment was actually for. No model can generate a fact that was never recorded. But capture is not a career; it is a task that attaches to whoever already touched the transaction. The salesperson, the driver, the owner with a phone.

Audit survives, because someone independent has to test the record. Though it survives smaller: when looking is cheap, sampling becomes census, and one reviewer covers what a team of associates used to.

The signature survives, and this is the sturdiest floor of the three. A tax agent, a licensed auditor, a licensed company secretary — these are people who can be struck off. AI can prepare everything and still cannot sign it. That is a legal floor, not a technical one, which is why it does not move when the next model ships.

Everything between capture and signature — transcription, coding, filing, matching, chasing, reconciling — is the part that just became arithmetic. It is also, by headcount, most of the profession.

Where the people go

This is the question I keep coming back to, and the honest answer is not the comfortable one. The three survivors cannot absorb the displaced: capture creates almost no seats, audit is capped by licensing rather than demand, and the signature was always one person per company.

So, roughly in order of volume:

1. The seat is simply never refilled. Someone retires, resigns, or finishes their contract, and the replacement is quietly not hired. This is where most of the adjustment happens, and it is invisible, because nobody is retrenched.

2. Into the exception queue. The five or ten per cent the machine cannot match — the messy, the disputed, the undocumented. Real work, but roughly one seat for every ten removed.

3. Into checking the machine. Controls, evidence, explaining why something was booked the way it was. Genuinely new, genuinely small.

4. Sideways, at a lower price. Credit control, procurement, operations, admin at a smaller company. The knowledge keeps its value; the wage it commands falls.

5. Out of the profession. The typing pool was not retrained into computing. It became administration and retail at flat wages, and the typing dispersed into everybody else's job. That is the closest historical match to bookkeeping, and it is not a happy one.

Two shortages, one cause

Now the part almost nobody is pricing in.

The junior seat was never only a job. It was the training system. Judgment is not taught in a lecture hall; it is accumulated by doing reviewed work, badly, for several years, under someone who corrects you. Remove the bottom of the pyramid and you have not just removed jobs — you have removed the only machine that manufactures senior people.

A founder whose post crossed my feed this week put it in one line: judgment is the finance asset with no vendor. He is right, and the implication is brutal. In ten years you get two shortages at once: a surplus of displaced mid-career staff, and a shortage of people qualified to sign. Same cause, opposite ends of the same building.

And notice who pays. The people already in seats mostly keep them — they absorb more work, not less. The entire cost lands on the ones who were never hired: this year's graduates, the ones who picked the sensible degree because it always led to a job. The safest degree fell first, and for the same reason: work that is well defined and teachable is work that is well defined and automatable. You will not find these people in any retrenchment statistic, because you cannot count a job that was never advertised.

Malaysia is not behind on this

The instinct here is always “that is a Western problem, we are years behind.” On this one, we are not — and the reason is our own regulation.

MyInvois has spent two years converting Malaysian commerce into structured data. Phase 4 brought businesses with RM1 million to RM5 million of turnover in from 1 January 2026, with the relaxation period running to the end of this year. The exemption line was raised from RM500,000 to RM1 million, and the planned final phase was cancelled — so the smallest operators are out, and everyone above the line is in.

The point is not the compliance burden. It is what it leaves behind. An economy whose invoices are already machine-readable has done the expensive half of the automation for free, and handed it to the tax authority at the same time — which is the argument I made about enforcement becoming a search rather than a lottery. The clerk whose job was to type an invoice into a system is competing with a file that was born structured.

To be fair

Four honest qualifications, because I am making a strong claim.

Eight per cent is a real number. Most finance functions genuinely cannot do this today. Legacy systems, group policies, auditors who want their own formats, data sitting in four places. My one clean day on our own books is not a listed company's month-end close, and I am not going to pretend otherwise.

Anticipation can be wrong. A 64 per cent expectation is a plan, not an outcome. Plans get reversed — several large companies have quietly rehired after cutting on an AI assumption that did not hold.

The errors are still there. Small ones, in my own run, this week. In a domain where a wrong number is not a typo but a misstatement, “mostly right” is a category error. Somebody still has to look.

And this cuts my way, which you should discount for. I sell AI capability. When I tell you the work is collapsing faster than people think, notice that I benefit from you believing it. Check my reasoning, not my conclusion.

The bottom line

Why three months, and not three years? Because the decision is not waiting for the technology — the survey already proved that. It is waiting for the budget cycle. Headcount for January is being decided in October and November, by people who have read the same survey I have, in a year when the seniors are retiring at an eight-year high and the relaxation period on structured invoicing runs out on 31 December.

The capability crossed my own threshold in six months. The intake decisions for next year are being made right now. Those two clocks are about to meet.

So if you are in this profession: stop defending the middle, because the middle is arithmetic now. Move toward the two ends — the capture, where the facts are born, and the assurance, where somebody carries the consequence. Write down the judgment calls of the people about to retire while they are still in the building; that file will outlast the org chart. And if there is a licence in reach, go and get it, because a licence is the one thing in this industry that a model cannot hold.

If you employ them: the cheapest thing you will ever cut is the junior seat, and the most expensive thing you will ever buy back is the senior it would have become.

I have been wrong about timing before, and I would rather be early than comfortable. But I do not think this one is far away, and I would not plan the next quarter as though it were.

Prepare for the earthquake in accounting and finance in the next three months.