On a Wednesday afternoon a trainee lawyer asks her supervising partner whether she may summarise a sixty-page judgment in ChatGPT. The judgment is public, but she wants to paste in the notes from the client meeting so the summary speaks to the matter. The firm's AI protocol says "carefully" and "no confidential information". The partner reads it, hesitates, and says: better not. The trainee does it that evening at home, on her own account.
That is the moment an AI policy works or fails. A usable policy for a law firm answers a small number of questions concretely: which tools, with which account, which client information never, what to do in doubt, who assesses, and how the lawyer sees the decision at the moment of typing. The eight questions every AI policy has to answer are in how to write an AI policy. This piece is about what changes when the reader is a lawyer.
Why is "no confidential information" not enough?
Because at a law firm almost everything is confidential. The duty of confidentiality covers everything a client has entrusted in that capacity: the facts, the documents, the strategy, and often the mere fact that someone is a client. A rule that keeps all of it out of the tool keeps the tool out of the work. Then what the trainee did happens: the work moves to an account the firm cannot see.
A policy that works says what can be done and how. Summarising a public judgment is fine. Pasting in the meeting notes may be fine too, but without the names of the client and the other side, and in the firm's business account. That is an instruction someone can follow on a Wednesday afternoon, and it keeps the work inside the environment the firm has agreements about.
Which questions does the policy answer for a firm?
| Question | What a usable answer looks like at a law firm |
|---|---|
| Which tools are allowed? | A list of names with the account next to each: the firm's business ChatGPT licence is a different tool from ChatGPT on a private account |
| For which work? | Per tool, the tasks that fit (public case law, a clause without parties) and those that do not (a complete file, a client meeting) |
| Which information never? | Who the client is where that is itself confidential, the strategy and settlement range, the other side's documents received under confidentiality, special category data from criminal and family matters |
| What about a tool with no decision? | Public information only, and this is where you request the assessment |
| Who assesses, and how fast? | The partner responsible for quality or compliance, with IT and the DPO where present, within two weeks |
| Who is covered? | Everyone working under a lawyer's responsibility, so the secretariat, paralegals and trainees too |
| How do people see a change? | Not in an email, but at the moment the tool is opened |
| How does the firm know the policy still holds? | By setting the list in the policy next to the tools that are actually opened, in numbers, without names |
The last three rows are missing from most protocols. They are exactly the ones that decide whether the policy still says anything, six months on, about the tools that exist by then.
Why is the account type the first agreement?
Because the same tool on screen can be two different services. With a free or personal ChatGPT account, input can be used by default to improve the model unless the user switches that off; with the business variants that is excluded by default, as OpenAI describes it. With a free account there is no data processing agreement, and nobody at the firm can manage members, settings or shared links. The difference per vendor is in free versus business AI accounts. A policy that says "ChatGPT is allowed" without naming the account is allowing the free account. So say it precisely: the business licence is allowed, the private account is not.
One nuance belongs here. The business account settles the GDPR side: training, administration, and the data processing agreement that Article 28 of the GDPR requires once a third party processes personal data for you. It does not settle the duty of confidentiality. The information still goes to a party outside the lawyer's circle, and whether that fits in a specific matter remains the lawyer's judgement. That is why the table has a separate row for what never goes in, even with a business account.
Which client information never goes in?
The general list is in what not to share with AI. For a law firm three categories come on top: the identity of the client where the fact of the relationship is itself confidential, the strategy and the settlement range, and the other side's documents obtained under confidentiality. What can go in is usually more than people think: a public judgment, a statutory provision, your own clause with [claimant] and [defendant] in place of the names. The rule of thumb the firm writes down: if someone could recognise the matter from the input, it comes out before it goes in.
What happens with a tool nobody has decided about?
A firm of twenty-five lawyers has a policy with three allowed tools. For some months the international practice has been using a translation tool for correspondence with a foreign client. The policy says nothing about it, so nobody asked, and the correspondence names both parties.
The policy then gains a fourth flavour: no decision yet. Every tool not on the list falls into it automatically, with an instruction everyone can follow: public information only, and report the tool to the quality partner. The translation tool is assessed within two weeks and becomes limited use, with a data processing agreement and without party names. Record the vendor's answers with a date, because AI services change their terms faster than a firm revises its protocol. Why "no decision yet" is the most honest of the four statuses is set out separately.
How does the policy reach the lawyer at the moment of typing?
The protocol sits in the quality handbook, the chatbot sits in the browser, and the trainee at the start needed an answer while the text box was open. A policy that exists only in the handbook arrives too late.
The decision works when it appears where the tool is opened. For an allowed tool nothing happens. For a tool without a decision a calm notice appears: public information only, and a button to continue. For a tool that is not allowed, the approved alternative is the first button, with continuing after a deliberate press-and-hold. For a limited-use tool, the condition the firm wrote itself appears, in one sentence.
BeeSensible does this in the browser: the firm records a decision per tool, and that decision appears when someone opens the tool. The extension blocks nothing and decides nothing; continuing is always possible, and what happens next is counted without a user id. When someone pastes case information into a supported AI tool, party names, matter numbers, amounts and company registration numbers are marked as they type. How that works is on the page about AI tools. This is what the overview looks like in which the decisions sit next to the observed use:
