Preparing for a quality review, the compliance officer of a firm with six offices asks the firm's IT provider which AI tools are in use. The answer is an email with four names: Microsoft 365 Copilot, ChatGPT, the bookkeeping software and the software that scans and reads purchase invoices. That same afternoon she watches a colleague in the accounts preparation team ask an unfamiliar tool to explain a general ledger.
That is a firm's AI register in a nutshell: a short list of what IT set up, set against a working day in which AI turns up in more places than anyone has reported. What belongs in that register for any organisation is covered in our article on what an AI register is and what to record. This piece is about what is different in an accounting firm: which tools, which fields and where the second source comes from.
Which AI runs in an accounting firm?
More than the chatbot. The general assistants are there, of course: Copilot in the Microsoft environment, ChatGPT and Claude for a client email or a note to the accounts, Perplexity for a tax question with sources. But most of the AI in a firm sits in software it has used for years. The bookkeeping software suggests journal entries, the invoice scanning software reads invoices, the accounts preparation software drafts a first version of the notes to the accounts and the data analytics tool in the audit practice summarises exceptions.
That second group never came with a new domain, a new account or a new purchasing decision: the vendor added the feature in an update and the firm has been using it ever since. It is usually missing from the register because nobody recognises it as AI. Why that is a structural problem is explained in our article on why tool catalogues date the moment you print them.
Then there is a third group: the tools staff have found for themselves. A transcription tool for the meeting on the annual accounts, a PDF summariser for lease agreements in an audit file, a browser extension that rewrites emails. They appear nowhere in the records. That is not a breach but shadow AI: picked up by staff themselves, without a decision and genuinely in use.
Which fields matter more in a firm?
The basic fields are the same for any register: vendor, purpose, owner, risk assessment, data use, decision, date and review. In a firm, four fields do more work than elsewhere.
| Field | What you record | Why it matters in a firm |
|---|---|---|
| Client information allowed | Public only, anonymised draft or file contents too; and what never goes in | This is where the duty of confidentiality is applied, tool by tool |
| Data processing agreement and location | Whether there is an agreement, where processing happens, whether the vendor trains on input | Without an agreement the tool is unsuitable for clients' personal data |
| Practice area | Accounts preparation, audit, tax, payroll, advisory | The same tool carries a different risk in payroll than in advisory |
| Embedded or standalone | AI feature inside existing software, or a service with its own domain | Embedded AI never went through a purchasing decision and otherwise stays out of sight |
The first field is what separates a list of tools from a register of decisions. Confidentiality is one of the principles in the Dutch professional code for accountants (the VGBA), and the accountant is expected to make sure that people working under their responsibility observe it too. The register is where that is spelled out for each tool: the text of a statute may go into Perplexity, a client name may not; the client file may go into Copilot, within the permissions already on it; nothing about a client goes into the free ChatGPT account. Exactly how the professional rules apply varies from firm to firm and is for the firm and its compliance officer to judge.
The second field depends on the account type. A free or personal account has no data processing agreement, and input can be used for training unless the user switches that off. The difference is explained in our comparison of free and business AI accounts, and our article on data processing agreements for AI tools covers whether you need one in place. If a tool processes the personal data of clients or staff, the same application also belongs in the record of processing activities under the GDPR, with a cross-reference in both directions.
What does a survey show, and what does it miss?
A firm has allowed Copilot for client work and ChatGPT on a business account for writing tasks that involve no file contents. The register has eight rows, filled in through a survey of the service line leads. A first look at which AI services are opened in the browser shows that, besides those two, Perplexity and a free ChatGPT account are in use, plus three specialist tools: a transcription tool in the advisory practice, a PDF summariser in the audit practice and the AI that drafts notes in the accounts preparation software. Two of them appear nowhere in the register.
That is not an incident, because nothing had been decided about those two. It shows what a survey is: a snapshot, taken by asking people who know about a tool. Someone who pays for a tool personally does not report it. Someone who started using one last week wasn't around for the survey. And someone who doesn't know that the new notes feature in the accounts preparation software is AI cannot report it.
How does observed use feed the register?
By adding a second source alongside the register: which AI tools are actually opened in the firm, how often and whether there is a decision. Not who, not what was typed: only which tool, at firm level. The gap between that list and the register is the backlog, and every tool in it goes into the register with the most honest status there is: "No decision yet". The next steps are set out in our article on moving from shadow AI to a current AI register.
BeeSensible provides that second source. The extension reports only the domain of an opened AI tool, at most once a day per device, with no user ID and no page content, and checks it against a catalogue of over 1,000 assessed tools with the vendor, country, training policy and a risk score for each. In the Inbox, an admin sees which tools have newly appeared and still have no decision, and records the decision; BeeSensible never makes that choice itself (see the page on the AI register).
