An accounting practice can put AI agents to work across hundreds of client ledgers, but it cannot scale them without trained reviewers. Every client has its own chart of accounts, its own quirks and its own limits, and the review that catches errors does not scale by itself. Practices that train reviewers first grow; practices that skip it grow their error rate.
Why do accounting practices hit the review wall so quickly?
Because practice work multiplies by client, not by transaction. A finance team inside one company reviews one ledger. A practice reviews dozens or hundreds, each with different suppliers, different VAT treatment, different owner habits. An agent can categorise bank transactions, match receipts and draft returns at depth 2 for all of them at once. The output arrives faster than the old team could produce it.
The review does not arrive faster. A manager who used to review twenty sets of books a month now faces the same twenty plus the agent's drafts for fifty more. If reviewers are not trained to review agent work efficiently and well, the practice has two choices: slow down, or approve without looking. The second is how practices end up filing wrong returns at scale.
What makes agent work in a practice hard to review?
Three things that do not apply in-house. First, provenance varies by client. Under the Pryme Context Protocol, a bank feed is sealed, a client's note saying "that was a personal expense" is declared, and a receipt photographed and emailed is external. Some clients send everything properly; others send a shoebox. The reviewer needs to know which client is which and weigh the agent's work accordingly.
Second, the agent's confidence is uniform; the client's reliability is not. An agent will categorise a vague bank description for a disorganised client with the same certainty as for a tidy one. Third, the cost of error falls on the client and the practice's professional standing, not on the agent. The partner who signs a return answers for it.
Which reviews must stay with a trained person?
A practical split for a typical practice:
| Work | The agent | Depth | Stays with a trained reviewer |
|---|---|---|---|
| Bank categorisation | Codes routine transactions | 2, then 3 once certified | Unusual items, owner transactions, anything new |
| Receipt matching | Matches receipts to payments | 2 | Missing or doubtful receipts |
| VAT or sales tax returns | Drafts the return | 2 | Approval and submission |
| Payroll for clients | Prepares the run | 2 | Approval, and every bank detail change |
| Year-end accounts | Drafts adjustments | 2 | Judgements, estimates and sign-off |
| Client queries | Drafts replies | 2 | Anything that could be advice |
Submitting a return to a tax authority is hard to undo cleanly, and approving payroll moves money. Both sit at depth 4 or stay with a person entirely.
What does a trained practice reviewer do differently?
They review by exception and by client risk. Instead of reading every categorised line, they check the items the agent held, the new suppliers, the large or round amounts, and anything built on declared or external facts. They know each client's patterns. They read verdicts and understand which bound held an item: perhaps the practice's own limit for that client, perhaps the agent's certificate.
They also know what the agent cannot know. That the client mentioned a new van on the phone. That a director's loan account has been creeping up. That the previous accountant treated a recurring item differently. Context like this lives with people, and good reviewers write it back into the client's sources so the agent learns it too. Every verdict and every human decision is filed in an append-only record, so a practice can show which reviewer approved a return and on what basis.
Which roles in a practice need training first?
The people who approve. Managers and seniors who sign off client work need operator-level skill before the agent runs at volume. Bookkeepers who review categorisation day to day need the same core skills applied to their queue. Partners need enough to understand depth, certification and the record, because their name is on the output. The guide to which roles in accounting practices need AI agent training sets this out role by role.
Market context matters. Credit control for practice clients in Nigeria and financial control in a German practice face different rules and client expectations; see the briefs on credit controllers at accounting practices in Nigeria and financial controllers at accounting practices in Germany.
What should a practice do before scaling?
Train reviewers before raising volume. Pick one service line, such as monthly bookkeeping, and run the agent for a small set of clients with trained reviewers. Measure correctness of review decisions, not speed. Expand clients only when reviewer capacity and accuracy keep up. Remember that certificates lapse after 90 days, and a run that clears no depth drops the agent to depth 1 the same day; a practice running many clients needs cover for that moment, especially near filing deadlines.
The same pattern holds in other sectors; see how e-commerce and consumer lending meet the review wall. The full approach to training teams is in training teams for AI agents.
Where the Academy fits
Pryme Intelligence Academy offers a free readiness assessment for teams of up to 25 people: the team takes the Foundations test set in its industry, and the report shows scores by concept, the gaps and the first agent to start with. For a practice, that shows which reviewers are ready before client volume rises.
Questions people ask
Can a small practice use agents without dedicated reviewers?
Yes, if the people who already review client work are trained to review agent output. The reviewer role is a skill, not necessarily a new hire.
Should the agent submit returns to the tax authority?
Most practices should keep submission with a named person. A return submitted in error is hard to correct and affects the client's standing with the authority.
How does a practice handle clients who send poor records?
Treat their data as weaker and review more of it. The agent's provenance labels make this visible, so reviewers can spend their time where the records are worst.