Writing a law firm AI policy: an outline you can adapt
By the ChamberOne team 4 minute read
Somewhere in your firm, someone has already pasted a paragraph of a client letter into a public AI tool to tidy the wording. They meant no harm and probably told nobody. A written policy does not stop that on its own, but it gives people a clear line and gives you something to point to when you need to act. This outline is a starting point for managing partners and compliance leads. It is general information, not legal advice, and it should be checked against the rules your own regulator or bar publishes.
Key points
- Keep the policy to a few pages, with a named owner, a version number and a review date.
- List approved tools and their permitted uses, and treat everything else as not approved until reviewed.
- Say what data may go into which category of tool, and what checking is required before output is relied on.
- Plan for incidents and near misses, make reporting safe, and train everyone on real examples.
Keep it short enough to be read
A twenty-page policy that nobody opens protects nobody. Aim for a document of a few pages that a new trainee can read in one sitting, with a one-page summary that goes on the intranet. Use direct statements: staff may do this, staff may not do that, ask this person when unsure.
Give the policy an owner by name, a version number and a review date. AI tools change often, and a policy that is two years old is probably describing tools you no longer use.
Section one: scope and approved tools
Define what counts as an AI tool. Include chatbots, drafting assistants, transcription tools, summarisers, features built into your email or document software, and browser extensions. People often forget the last two.
Then keep a list of approved tools, each with its permitted uses. A tool might be approved for internal drafting of non-client text but not for client documents. Say plainly that anything not on the list is not approved until someone has reviewed it. Say who does that review and how a staff member asks.
Section two: data rules
This is the section that matters most. Spell out what may and may not be entered into each category of tool. A simple structure works: public information, firm-internal information, client confidential information, and especially sensitive material such as matters under a confidentiality wall.
For each approved tool, record what you know about where data goes, whether the provider keeps it, whether it trains on it and where it is processed. If you do not know, say so in the register and treat the tool as restricted until you find out. Your own client agreements and any duties of confidentiality should shape what you allow.
Section three: review standards
State that the lawyer who relies on AI output is responsible for it. Then say what checking looks like in practice. For research, every authority is found and read in a trusted source before it goes into anything. For drafting, the lawyer reads every line as if a stranger wrote it. For summaries, the key points are checked against the document.
- No citation goes into advice, a letter or a court document unless the person has opened the original.
- Output is treated as a first draft, never as a finished work product.
- Anything filed with a court follows that court's rules on AI, which the lawyer checks each time.
- Supervisors may ask how AI was used on any piece of work, and staff answer honestly.
Section four: telling clients
Decide the firm's default position on telling clients about AI use, and when specific consent is wise. Cover engagement letter wording, client questions and any client who has asked you not to use AI on their matter. Make sure the practice matches what the engagement letter says. A short, honest sentence is better than a long one nobody understands.
Section five: incidents and near misses
Plan for the day it goes wrong. Someone pastes the wrong document. A fabricated citation reaches a draft. A tool changes its terms. Say who staff tell, how quickly, and what happens next, including who decides whether a client, a court or a regulator needs to be told. Check your regulator's guidance for duties of that kind.
Make reporting safe. If people fear punishment for admitting a slip, they will hide it, and you will find out later and at greater cost. Record near misses as well as incidents, because they show where the rules are unclear.
Section six: training and review
Tie the policy to a short training session for everyone, including support staff and the partners who think it does not apply to them. Use real examples from your own practice area. Show what a fabricated citation looks like and what an over-confident summary looks like.
Review the policy on a set date. Add a standing agenda item for a partners' meeting every few months: which tools are in use, what has gone wrong, what rule needs changing.
- 1Train everyone on real examples
- 2Meet on a set date or standing agenda item
- 3Ask which tools are in use
- 4Note what has gone wrong
- 5Change the rule that needs changing
- Then back to step 1.
A worked example of one rule
Suppose a five-lawyer firm writes this rule: client names, addresses and account details are never entered into a public AI tool; a lawyer may paste a clause with the names replaced by letters into an approved tool to ask for plainer wording. The rule is specific, easy to remember and easy to test. A partner can ask any lawyer to show how they would apply it to a real document.
Compare that with a rule such as 'use AI responsibly'. Nobody can fail to follow it, so it changes nothing. When you draft each section, ask whether a busy associate at six in the evening would know exactly what to do.
Where a rule needs judgement, say who is to be asked. A named person with a short turnaround beats a committee that meets quarterly, and staff will use the route if it is quick.
- Names, addresses and account details never go into a public AI tool
- Names replaced by letters before a clause goes into an approved tool
- Easy to remember and easy to test
- Nobody can fail to follow it, so it changes nothing
- A busy associate at six in the evening would not know what to do
What a policy cannot do
A policy does not make a tool safe, and it does not make staff competent with it. It also cannot be enforced if you cannot see what people use. Pair it with technical measures where you can: blocking unapproved tools on firm devices, switching off AI features you have not approved in your own software, and keeping a record of usage you can review.
This article is general information about the law, legal practice and technology, current when it was published. It is not legal advice. Laws and guidance change: check the sources linked above and take advice on your own circumstances from a lawyer qualified where you are.
Read next
Australia's new statutory tort for serious invasions of privacy
The statutory tort for serious invasions of privacy commenced on 10 June 2025 under Schedule 2 of the Privacy Act 1988 (Cth). This article explains the five elements, defences, remedies, time limits and what the reform means for practitioners.
6 minute readClient confidentiality and AI tools: a checklist before you paste anything in
The questions to answer about retention, training use, access, hosting and sub-processors before client information goes into any AI tool.
5 minute readShadow AI in law firms: staff are already using it, so what next?
Why bans fail, how to find out what is being used, and how to offer an approved alternative with clear rules.
5 minute read