Article 4 of the EU AI Act requires accountancy firms that use AI systems to take measures so their staff, and others working on their behalf, have sufficient AI literacy. For firms using agents to reconcile ledgers, draft tax computations or sample audit evidence, partners, managers and trainees must understand what the agent did and where it errs.
What does Article 4 ask?
Article 4 applies to every provider and deployer of an AI system, not only high-risk ones. In substance, they must take measures to ensure, to their best extent, a sufficient level of AI literacy among their staff and other persons dealing with the operation and use of AI systems on their behalf. What counts as sufficient depends on the people's technical knowledge, experience, education and training, the context in which the system is used, and the people it is used on.
AI literacy, as the Act uses the term, means the skills, knowledge and understanding needed to use AI systems in an informed way and to be aware of their opportunities, risks and possible harms. It is not a certificate. It is a level of understanding fitted to the role. For the general rule, see what AI literacy means for accountancy firms using AI agents.
Who does it cover in an accountancy firm?
Everyone who uses or operates agents for the firm:
- Partners who sign audit opinions, tax returns and accounts that agents helped prepare.
- Managers who review agent-prepared work and decide what goes to the partner.
- Trainees and associates who instruct agents and work alongside them every day.
- Practice and IT staff who connect agents to client systems and set their bounds.
- Contractors and outsourced teams working on the firm's behalf, such as offshore processing centres or freelance bookkeepers using the firm's agents.
Clients are not covered by the firm's Article 4 duty, but they are the people the agents work on, which the article says should shape what literacy means.
What does literacy mean when an accountancy agent acts?
An accountancy agent is not a search tool. A governed agent in a firm might read a client's general ledger at depth 1, draft journal adjustments and tax computations at depth 2 for a person to post or file, and match routine bank reconciliations within tolerance at depth 3. Filing a return or posting to a client's ledger is held for a named person.
Literacy for each role follows from that.
Trainees need to know what the agent was allowed to do, what it actually did, and how to check it. A reconciliation the agent matched at depth 3 can still be wrong if a source was stale. Trainees must read the provenance of each figure: sealed data from the client's own ledger, signed confirmations from a bank, declared figures from the client's finance team, or external documents such as an emailed invoice. An external invoice cannot on its own authorise a payment, and should not drive an adjustment unchecked.
Managers need to recognise when an agent is confidently wrong, and to resist approving well-formatted work because it looks finished. That is automation bias, and in accountancy it hides in tidy schedules.
Partners need enough understanding to stand behind the work. When a regulator or client asks how a figure was produced, the partner must be able to answer, and the agent's append-only record lets them.
Everyone needs to understand refusals and lapses. A refusal means the request fell outside the agent's bounds, often because it involved a client the agent is not scoped for. An agent's certificate lapses after 90 days, and a run that clears no depth drops it to depth 1 the same day, so close and filing calendars must not depend on depth 3 work without a manual fallback.
The partner and board view of this is covered in what AI agent training an accountancy firm board needs.
How do you meet Article 4 in an accountancy firm?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Sufficient literacy for staff | Role-based training: trainee, manager, partner, practice staff | Training records by role and date |
| Literacy for persons acting on the firm's behalf | Contractors and outsourced teams trained before agent access | Contractor training records linked to access grants |
| Literacy fitted to context | Training uses the firm's own engagements and agents | Course content mapped to service lines |
| Awareness of risks and harms | Modules on automation bias, provenance and confidentiality | Assessment results on these topics |
| Kept current as agents change | Briefing at each agent recertification or scope change | Briefing records linked to agent change log |
What will a regulator or auditor ask?
Audit regulators inspecting a firm's quality management, and market surveillance authorities checking Article 4, will ask similar questions. How do you know the people using agents understand them? Show the training for the engagement team on this file. Who reviewed the agent's work, and what did they check? How are contractors trained? How do partners understand what they sign? Internal quality reviewers will add a file test: pick an engagement and trace an agent-produced figure back through the record to its source.
What does Article 4 not cover?
It does not prescribe a course, hours or a credential. It does not set audit, tax or accounting standards, which professional and statutory rules govern. It does not make every accountancy agent high-risk; most are not, so Articles 14 and 26 rarely apply. It does not cover clients' own literacy. And it does not replace confidentiality duties to clients, which shape what an agent may read. The Act's broader treatment of agents that take actions is covered in the EU AI Act and agents that act, and other sectors in Article 4 in e-commerce and in healthcare.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Where the Academy fits
Pryme Intelligence Academy's Foundations track takes about four hours and suits partners and anyone who needs the concepts; the Operator track suits managers and seniors who review agent work daily. Module tests are set in the learner's own industry and role, so an audit senior is tested on audit work. Our guide to AI literacy and certification for people who work with agents explains the levels.
Questions people ask
Does Article 4 apply to a small practice?
Yes. It applies to any deployer of an AI system, whatever its size. What counts as sufficient literacy scales with the practice and how it uses agents.
Do partners need the same training as trainees?
No. Partners need enough understanding to stand behind the work and challenge it; trainees need hands-on skill in instructing and checking agents. Article 4 asks for literacy fitted to the role.
Should outsourced teams be trained by the firm?
If they use or operate agents on the firm's behalf, the firm must take measures for their literacy. That can mean training them directly or requiring and checking equivalent training.