A compliance officer at a firm with an assurance practice and an advisory practice gets a question from a partner on Tuesday: may the tax practice use Claude for advisory memos, because two people already do. The firm's AI policy was adopted last year: five pages, with the words careful and confidential in it, and Copilot as the preferred tool. About Claude, about an advisory memo with a valuation in it, and about the account those two people work on, it says nothing.
That is the moment an AI policy works or does not. A policy for a firm does not have to be long, but it has to answer that question: is this tool allowed, with this file, on this account, and if not, what instead. The general structure of such a policy is in writing an AI policy, and a one-page example in AI policy example. This piece is about what is different for an accounting firm.
Which questions must a firm's policy answer?
The same questions as anywhere else, with answers that fit the work of a firm.
| Question | What a firm writes down |
|---|---|
| Which tools are allowed, on which account? | A list of names with the account type next to each; Copilot inside the firm's own environment is not the same chatbot on a personal account, and the AI feature in the compilation package is on the list too |
| Which data never goes in? | Client names combined with figures, citizen service numbers and payroll data, IBANs, figures before adoption or publication, valuations, anything from an audit file |
| What differs for assurance compared with advisory work? | Which tools are allowed in which practice, and that AI output in an audit file is not audit evidence without the accountant's own work |
| Who assesses a new tool, and how fast? | An owner in the practice answers the questions, the compliance officer tests, and there is a deadline |
| What do you do with a tool that has no decision? | Public information only, and this is where you request the assessment |
| How does the floor hear about a change? | Where the tool is opened, not in an email from three weeks ago |
| How do we see whether the policy is followed? | The list from the policy next to the tools actually opened, in counts, without names |
The second row is different in a firm from most organisations. Most of what must never go in is not personal data: a revenue figure, a valuation, a figure the client has not yet adopted. The GDPR says nothing about that; the duty of confidentiality in the VGBA, the Dutch professional rules for accountants, does. So write that row not as a privacy rule but as a confidentiality rule, with examples from the firm's own work. What a junior accountant does with it in practice is in ChatGPT and client files.
What differs for the assurance practice compared with advisory work?
More than one rule, and the policy has to say which.
In advisory work it is about confidentiality and about the agreements with the client: some engagement terms say something about who may process the information, and the vendor of a chatbot is a third party. In the assurance practice the audit file comes on top. What a chatbot says about a balance is an idea, not audit evidence; the accountant does the work and records what was done. A summary from an AI tool in the file without that work next to it is the first question asked in a quality review.
That does not mean the assurance practice uses no AI. It means the policy says per practice which tools are allowed and for what: a tool for rewriting a client email can be allowed in both practices, a tool for summarising an audit file needs a decision of its own. How much weight that carries differs per firm and per engagement; the compliance officer and the firm's quality policy draw the line, not this article.
What is the compliance officer's role?
A different one from the role they usually get. In many firms the AI policy lands with the compliance officer, because it is about rules. That does not work, for the same reason "the DPO will handle it" does not work elsewhere: whoever tests should not decide and own at the same time. A compliance officer who approves the tools personally can no longer say independently, later, that something is wrong.
This division works better. The compliance officer tests a new tool against the professional rules and the quality system, and together with the DPO or privacy officer against the GDPR: data processing agreement, where processing takes place, whether a DPIA is needed. The decision itself belongs to the board or the partners. And each tool has an owner in the practice: the head of the compilation practice for the AI feature in the package, the tax partner for Claude. That owner answers the questions, knows the work the tool is used in, and drives the reassessment when the vendor changes something. How those roles sit in general is in AI governance roles: who does what.
How does the policy reach the compilation practice?
Not through the document. A five-page policy is read once, at the training or on joining, and the tool is opened hundreds of times after that. What does stick is one rule per situation, at the moment the situation arises. How to write that without drowning people in rules is in how do you communicate AI policy.
Back to the partner and his question. The two tax advisers open Claude and see a short notice: no decision has been made about this tool yet, enter public information only, and this is where you request the assessment. They write the memo in Copilot that afternoon, without the family's name and without the valuation, and request the assessment. Three weeks later the firm has decided: limited use, only in the firm's business environment and without client data, with Copilot as the route for anything a client appears in. From that moment, everyone who opens the tool sees that one sentence.
That is not a complicated setup. It is the list from the first row of the table, with one of four statuses per tool, visible where the work happens.
BeeSensible does this in the browser: the firm records a decision per tool in the AI Tools module, and that decision appears when someone opens the tool, with the approved alternative as the first button. The extension blocks nothing and decides nothing; continuing is always possible, and what happens afterwards is counted without a user id.
