AI policy for the legal department: what it should contain

An AI policy for lawyers is a short document in nine points. It answers which tools may be used, what may be entered into them, how the output is reviewed, who is responsible and what happens when something goes wrong. Here is the outline I use, with what the Swedish Bar Association's guidance requires of Swedish law firms and what the EU AI Act requires of everyone.

Updated by Nils-Erik Jansson

What the policy is for

The policy should contain nine points, as the outline below shows. It refers to the risk assessment, the contract and the training, and is therefore short.

The policy is needed as soon as anyone in the department uses AI, whether under an Enterprise agreement or on a personal account. It does three things. It sets out what is permitted, so that those who want to do the right thing can do it. It draws the line for what may never be entered. And it places responsibility where it belongs, with the lawyer who uses the output.

The risk assessment, the contract with the vendor and the training are separate documents. The policy refers to them. That is why it can be short.

The outline

  1. Scope. Who is covered (everyone in the department, consultants, trainees) and which tools the policy applies to: both those the company has bought and those anyone can open in a browser.

  2. Permitted tools. A list with names and plans. "Claude Enterprise under the company's agreement" is a permitted tool. "Claude" is not, because a personal account is governed by different terms and is not covered by the undertaking not to train on the content [4]. The list is owned by a named person and updated when agreements are signed or terminated.

  3. What may be entered. Start with what is allowed: public information, your own templates, anonymised contracts, information about counterparties to the extent the matter requires it. Then set out what may never be entered, and link it to the list of matters kept out of the tool. For law firms, client information without a data processing agreement and a transfer assessment is such a limit [1].

  4. Matters kept out. The list itself, by type of matter. It is drawn up in the risk assessment and set out in the policy so that everyone can see it. Possible entries: matters with information subject to special access controls, disputes where a request from a US authority cannot be accepted, and matters where the client has said no.

  5. Review of output. Responsibility for advice is the same whether or not the advice is based on output from a tool [1]. The policy says how this applies: case law, statutory references and factual statements are checked against the source before they are used, and whoever passes a document on is responsible for its content. Nobody sends an AI generated draft to a counterparty without having read it.

  6. Transparency. When does the department tell the client or the business that AI has been used? The guidance names transparency as one of the questions the guidelines should answer [1]. My advice is a simple rule: always when asked directly, in the terms of engagement as standard, and never as an excuse for a mistake.

  7. Contracts and data protection. A reference to where the data processing agreement, the transfer assessment and any data protection impact assessment are kept, and who owns them. The policy does not repeat them.

  8. Incidents. What a member of staff does after entering something that should not have gone in: whom to tell, how quickly, and what the department then does with the vendor. This requires the vendor's deletion procedure, so it is included.

  9. Training and follow-up. Who is trained and when, and that the training is documented as one of the measures that Article 4 of the EU AI Act requires of the organisation [3]. The date of the next review of the policy. The guidance says that guidelines and training should be kept up to date [1].

Mistakes to avoid

  • The policy prohibits everything. The tools are then used anyway, privately and without a trace. A policy that only says no has controlled nothing.
  • The policy is written for IT. It covers encryption and access rights but says nothing about what a lawyer may write in a prompt about a dismissal. The lawyers do not read it because it is not about their work.
  • The policy names a tool the company has no agreement for. A permitted tool without a data processing agreement is a contradiction. Check that every entry in the list has an agreement behind it.
  • Nobody owns the lists. Permitted tools and matters kept out change. Without an owner, the policy goes out of date.

How to keep it up to date

One review a year is the minimum, with another when an agreement is signed or terminated, when the vendor changes its terms, or when an incident has shown that a rule did not work. The training is repeated for new joiners and when tools change. The document itself changes little. The lists change often.

The guidelines are one of five parts of the data protection package, together with the transfer assessment, the data protection impact assessment, the draft supplementary agreement and the training. For organisations introducing Claude, they are part of the setup. Which matters are kept out is decided by the transfer assessment, see the guide on transfer impact assessments for AI tools.

Read also the AI policy template.

Download

Free. No sign-up and no email address.

AI policy template for the legal department

AI-policymall för juristavdelningen

Frequently asked questions

Must a Swedish law firm have an AI policy?
The Swedish Bar Association's guidance on generative AI says that the firm should adopt and implement guidelines on which tools may be used, how they may be used, which legislation may apply, fact checking, transparency and ethics, and that staff should be trained [1]. It is a recommendation, but it is linked to the principal's responsibility for staff under the Bar's Code of Professional Conduct. Without guidelines, each member of staff's own judgement becomes the firm's policy.
How long should the policy be?
I suggest two to four pages. Longer policies are harder to read, and what fills them out belongs in the risk assessment or in the contract with the vendor. The policy should take ten minutes to read and answer the question of what I may do today.
Is the company's general IT policy enough?
Not if it only covers tools and access rights. The legal department also needs rules on what a lawyer may enter about a counterparty, an employee or a client, and on how an AI generated draft contract is reviewed before it leaves the department. The legal department's policy can be an annex to the IT policy, but it needs to exist.

Sources

  1. [1]Vägledning om användning av generativ AI i advokatverksamhet, Sveriges advokatsamfund, 12 juni 2025, in Swedish; Swedish Bar Association guidance on generative AI, sections 3 and 5
  2. [2]Uppdaterad vägledning om användningen av externa IT-tjänster i advokatverksamhet, Sveriges advokatsamfund, 15 april 2026, in Swedish; Swedish Bar Association guidance on external IT services
  3. [3]Regulation (EU) 2024/1689 (the AI Act), Article 4, consolidated text of 27 July 2026, EUR-Lex, read 2026-11-17
  4. [4]Commercial Terms of Service, Anthropic

Back to AI for legal

More under AI for legal