Most accountancy firms train the wrong people at the wrong time. They brief partners and managers, then launch agents in busy season, leaving juniors to review bookkeeping, reconciliations and payroll drafts they have never prepared themselves. Qualifications are taken as proof of agent skill. The fix is to train reviewers first, on real client files, before the agent's depth widens.
Why is the junior reviewer the weakest point?
Accountants have always learned by doing the preparation. A trainee posts the journals, reconciles the bank, builds the working papers, and learns from the manager's review notes what good looks like. Agents take over exactly that preparation. A bookkeeping agent categorises transactions and drafts journals at depth 2; a reconciliation agent matches routine items at depth 3 and holds the exceptions.
So the person who now reviews the agent's output is often a junior who never did the work by hand. They can see whether a reconciliation balances. They cannot always see that a recurring payment was categorised as a repair when it is plainly an improvement, because they have not built the instinct. Firms that train only partners and managers miss this entirely. The people who most need to understand verdicts, sources and exceptions are the ones closest to the agent.
The answer is not to stop juniors reviewing. It is to give them a structured route: decide the agent's drafts alongside an experienced colleague, compare the decisions, and learn from the gap.
Is a professional qualification enough?
No, and assuming it is causes quiet errors. A qualified accountant knows the standards, the tax rules and the client. That does not tell them what it means that an action was held because a source was external, or that the agent's certificate lapsed last week and its routine postings are now waiting for approval.
Agent literacy is a separate, smaller body of knowledge: depth 1 to 5, the four verdicts, the five bounds in the decision chain and which one decided, provenance classes, and lapse. A senior manager can learn it in an afternoon. Without it, they will misread the queue in ways their qualification cannot catch.
Why does launch timing matter so much?
Because accountancy firms run on deadlines. The temptation is to switch the agent on before the year-end or filing peak, to get the help when it is most needed, and to train people after. That puts untrained reviewers in front of the largest volume of held actions the agent will ever produce, under the most time pressure they will ever face.
The pattern that results is rubber-stamping: drafts approved because they look complete and the deadline is close. The record then shows a high approval rate that says nothing about quality, and when certification comes round the evidence cannot support more depth. Launch at the quiet end of the cycle, let reviewers practise while volumes are low, and arrive at the peak with a trained team.
What goes wrong with generic training material?
The client file is where the risk sits, and generic examples skip it. Real client records arrive with mixed provenance. A bank feed from the client's bank is signed by a named party. A spreadsheet of cash expenses the client prepared is declared. A supplier invoice the client forwarded by email is external and cannot authorise a payment on its own, which matters for any firm running payroll or bill payments for clients.
Training on tidy, invented ledgers teaches none of that. Train on anonymised real files from your own client base, including the awkward ones: the client who changes bank details by email, the director's loan account that keeps moving, the payroll run with a late starter. Those cases are what reviewers will meet.
Why is file review not the same as agent review?
Firms already have a review culture, so they assume it transfers. It partly does. But traditional file review happens after the work, against a checklist, often days later. Agent review happens at the point of action, case by case, and the reviewer's decision is filed in an append-only record that explains why the action was approved, edited or refused.
That changes two habits. Reviewers must write reasons a colleague could follow, because the record will be read by the agent's owner and possibly by a client or regulator. And they must act on patterns: if the same correction appears every week, it goes to whoever owns the agent's brief, rather than being fixed silently each time.
What should an accountancy firm do instead?
Five changes, in this order:
- Train reviewers first, starting with juniors. Pair each with an experienced reviewer for the agent's shadow period.
- Give everyone the agent vocabulary, including partners, in a short session set in accountancy work.
- Launch at a quiet point and widen depth only after reviewers have been assessed on real decisions.
- Use your own anonymised client files as the training material.
- Make approval depend on competence. A firm can adopt the practice of granting approval rights for client payments only to people with a current certificate.
The full sequence for any organisation is in the guide to training teams for AI agents. For the detail of content, see what an AI agent training programme for accountancy firms should include; for timing and owners, the AI agent training plan template for accountancy firms; and for planning the spend, how much AI agent training costs for accountancy firms.
Other professional firms make related mistakes. Law firms get some of the same things wrong, particularly around supervision of junior staff.
Where the Academy fits
Pryme Intelligence Academy suits this order. Partners and managers can take Foundations, about four hours and free. Juniors and seniors who review held actions take the Operator track, which ends in a written exam and a two-week practicum of real review decisions signed off by a Pryme assessor. Module tests are set in the learner's own industry and role, so the cases look like accountancy work.
Questions people ask
Do client-facing partners need hands-on agent training?
They need enough to answer a client who asks how the work was done and who checked it. That means the vocabulary and an understanding of what the record shows. Hands-on reviewer training is for those who approve held actions.
How do we train juniors who never did the work by hand?
Give them the manual version on a sample of cases first, then have them review the agent's drafts of the same cases. Comparing their work with the agent's teaches what to look for faster than either alone.
Should clients be told that agents prepared their work?
Tell clients plainly how agents are used and who reviews their output, in engagement terms and in conversation. Clients care less that an agent drafted the work than that a competent person checked it.