An accountancy firm proves it with four pieces of evidence that link to each other: a map of who approves which agent work, a certificate for each approver that can be checked at source, a practicum log showing they practised on real client files, and the filed record of each approval showing the approver was certified on that date.
Who will ask, and what do they want to see?
Accountancy firms face questions from more directions than most. Audit regulators and professional bodies inspect the quality of engagements. Clients, especially larger ones, ask how their books and returns are prepared. Professional indemnity insurers ask about controls. And if the firm deploys agents in the EU, Article 4 of the EU AI Act expects sufficient AI literacy among staff.
They all want the same thing underneath: confidence that when an agent prepared a reconciliation, a set of adjusting entries or a tax computation, a competent person checked it before it counted. "We trained everyone" does not answer that. "This reviewer, certified until this date, approved this entry" does.
Which evidence answers the question?
| Evidence | What it shows | Where it comes from |
|---|---|---|
| Role map | Which roles approve which agent outputs, at which depth | The firm's own policy |
| Certificate per approver | The person passed a test and supervised practice | The issuer, checked at source |
| Practicum log | Real decisions made under supervision, with results | The firm's training records |
| Approval record | Who approved each agent action, when, and what they changed | The workspace's filed record |
| Renewal history | Certificates were kept current | The certificate register |
The approval record is the piece firms most often lack. Without it, the others show that people were trained, not that trained people did the work. The AI agent training plan template for accountancy firms shows how to plan the first four.
What should accountancy staff be trained to catch?
The evidence only means something if the training addresses the errors agents make in accounting work. Three matter most:
- Source confusion. A bank feed is sealed by the bank. A client's spreadsheet of accruals is declared. A supplier statement emailed in is external. An agent may post adjustments from all three with equal confidence.
- Plausible but wrong classification. An agent can code a capital purchase as an expense because the invoice description looks routine. Reviewers need to check the coding of material items, not just the totals.
- Held versus refused. Posting a journal above the client's materiality limit may be held for a reviewer; filing a return may be refused for the agent entirely. Staff must know which, and why.
A firm whose evidence pack shows training on these, tested on its own client files, has a far stronger answer than one with a generic course certificate. The piece on what accountancy firms get wrong about AI agent training covers the usual gaps.
A worked example
Take a firm of 60 people providing bookkeeping and year-end accounts to small businesses. It deploys an agent to reconcile bank accounts and propose adjusting entries at depth 2. Only seniors and managers may approve entries.
When a professional body reviewer inspects a sample of files, the firm shows: the role map stating that adjusting entries need a certified senior; the certificate numbers for the four seniors who approved entries in the sample, each checked at source on the day of review; their practicum logs; and the filed record for each sampled entry, showing the agent's proposal, the senior's edit and the approval time. One sampled entry was approved by a senior whose certificate had lapsed a week earlier. The firm can show that it found this itself, reassigned the senior's queue and had the entries re-reviewed.
Which controls keep the evidence trustworthy?
- Approval rights depend on a current certificate, so a lapsed approver cannot approve. A firm can adopt this depth by depth.
- Certificates are checked at the issuer, not accepted as PDFs.
- The register is reconciled monthly against the people actually approving.
- Records cannot be edited after the fact. An append-only record is what makes an approval history credible.
For EU deployment, the Article 4 note for accountancy explains the literacy duty. Similar evidence sets for neighbouring professions are in how law firms prove staff are trained and how property management firms do it. The guide to AI literacy and certification covers the overall approach.
Where the Academy fits
Pryme Intelligence Academy gives firms the certificate and register pieces of the evidence. Certificates are issued through VerifyDoc.ai and can be checked at source by number or QR code, and a lapsed or revoked certificate shows as such. For teams, an admin view shows progress, results, certificates and expiry dates in one place.
Questions people ask
Does CPD training count as evidence?
It shows learning time, which helps. It does not show competence at approving agent work unless it includes a test and supervised practice on the work itself.
Should partners be certified?
Partners who approve agent work directly should be. Partners who sign off engagements need enough literacy to judge whether the review beneath them was sound.
How long should the evidence be kept?
As long as the firm keeps the engagement files it relates to, so any approval can be matched to the approver's status at the time.