The compliance officer of a forty-lawyer firm is asked by the partnership for "the AI list". She opens the file made last year: two rows, the legal research environment and Microsoft 365 Copilot. The same week an associate in corporate mentions he is trying a contract analysis tool someone recommended at a conference, and the family law team's dictation software turns out to have started producing summaries since an update.
An AI register for a law firm records, per application, which tool is used, for what, by which practice group, with what case information, and what the firm has decided about it. The fields are largely the same as for any organisation. Three of them weigh more heavily at a firm, and the list of tools that belong in it is longer than the two rows most firms start with.
Which AI tools run at a firm?
More than the research environment the firm licenses. Five kinds of AI show up at firms, and most registers name one or two.
The general chatbots: ChatGPT, Claude, Copilot, Perplexity, often through a free or personal account. The AI features in the legal database, which for a few years now do not only search but also summarise and draft. Dictation and transcription software that, after an update, suddenly summarises conversations. Specialised tools for contract analysis and due diligence, often connected to the document management system. And browser extensions that rewrite or translate pages and read along on every page that is open, including the case management system; why those belong in the register is in browser extensions are AI tools too.
Part of this was never procured: a free chatbot appears on no invoice, an AI feature arrived with an update, an extension is installed in three clicks. A register built from the purchasing records misses exactly those rows.
Which fields count extra for a firm?
The base fields are in an AI register: what it is and what you record in it, with a completed example. For a law firm, three fields make the difference, and two are missing from most registers.
| Field | What you record | Why it weighs more at a firm |
|---|---|---|
| Confidentiality | Which case information the tool may see and which never: nothing traceable to a client or matter, only anonymised documents, or full files | The duty of confidentiality and legal privilege sit with each lawyer and do not transfer to the vendor |
| Processing location and jurisdiction | Where processing takes place, where the vendor is established, and which government can demand data | Confidential documents in a jurisdiction where privilege does not apply is a different risk from processing in the EU |
| Agreement | A data processing agreement under the GDPR, plus the contractual terms on confidentiality, retention and sub-processors | A free account has neither; a business account has the first, and the second has to be read |
| Connections | Whether the tool reads in the document management system, the mailbox or the database, and with whose permissions | A tool that reads case files itself has a different reach from a tool that only sees what someone pastes |
| Client terms | Whether clients have set anything about AI in their terms or the engagement letter | Some clients rule out AI use on their matter; the register has to be able to show that |
Confidentiality is the most important field, and it is the one organisations outside the legal profession do not have. A data processing agreement governs what a vendor does with personal data under Article 28 GDPR. It does not govern the lawyer's duty to keep silent about what a client entrusted to them. That duty stays with the lawyer even when the tool belongs to someone else, and it is that duty the supervising dean looks at. So "processing agreement signed" is not a decision about case information. The decision is the sentence in the confidentiality field.
The Netherlands Bar has published recommendations on AI in the legal profession, organised along the profession's core values. Under confidentiality the substance is that a lawyer knows and controls the data flows, and puts no confidential data into free tools. The register is where that knowledge of data flows becomes concrete: per tool, per application, with a date.
What does a row look like?
Take the family law team's dictation software. Vendor: European, processing in Germany. Purpose: transcribing dictation and, since the update, summarising. Owner: the family law partner. Confidentiality: full dictation, so case information, including special category data about children and health. Processing location: EU. Agreement: processing agreement signed, audio retained thirty days, sub-processors named. Connections: none. Decision: allowed for dictation, limited use for summarising until the new feature has been assessed. Date and review date attached.
Last year the same software was one row with the status allowed. The update turned it into two applications, and the second needs its own decision. That is why every row has a review date.
What is not in it, and why does it lag?
After an afternoon's work the forty-lawyer firm has seven rows. A first look at which AI services are opened from the firm's browsers shows that, next to the research environment and Copilot, ChatGPT, Claude and Perplexity are in use, plus the contract analysis tool from the conference and a transcription tool someone uses for client meetings. Three of those appear nowhere.
That is not an incident: nothing had been decided about those three. It shows what a survey is: a snapshot of one moment, taken by asking people who know about their tool. Whoever does not know their dictation software is AI, or pays for a tool privately, does not report it. Why surveying more often does not fix that is in why an AI tool catalog ages the moment you print it.
The register needs a second source: which AI tools are actually opened from the firm. Not who, not what is typed in, only which tool and how often, at firm level. Every tool that appears there and not in the register enters it on the most honest status there is: no decision yet. How those steps run is in from shadow AI to a current AI register.
BeeSensible supplies that second source: the extension reports only the domain of an opened AI tool, at most once per day per device, with no user id and no page content, and sets it against a catalog of 865 assessed tools, each with vendor, country, processing agreement and training policy. An administrator sees which tools still have no decision and records the firm's decision; BeeSensible never takes that decision itself. More on how that works is on the AI tools page.
