An AI use policy for a small firm, section by section

It has to make six decisions in advance, in writing, so that nobody has to make them at speed on a Friday afternoon. A policy that does that in two pages is worth more than a twenty-page document nobody has read, because the test of a policy is whether a new hire can follow it in their first week.

The six sections below are the whole document. Each one closes with a sample clause written to be edited rather than admired.

Name the tools, by product name, and name the account type. A policy that says "approved AI tools" without a list has settled nothing, because every reader supplies their own list. The account clause is the half that gets skipped and the half that does the work: the exposure a small office should plan for is not an exotic tool but a good tool used on a personal free account, where the firm has no logs, no administrative control, and no terms it has reviewed.

Two provisions are worth adding while the section is open. First, a route for requesting a new tool, so the answer to "can I try this" is a process rather than a quiet yes. Second, a line covering AI features that arrive inside software the firm already runs, since those appear without anyone deciding to adopt them.

Sample clause. Approved AI tools are Claude Team on firm-provisioned accounts. Personal, free, or individually purchased AI accounts may not be used for any firm work, including drafting, summarizing, research, and file organization. AI features built into other firm software must be reviewed and approved in writing before use. Requests to add a tool go to the managing partner and are decided in writing.

Write the exclusion list as a list, and keep it somewhere a person can check in ten seconds. Categories beat adjectives: "sensitive information" is unenforceable, while a named list of matters, document types, and data categories is something a paralegal can actually apply at four in the afternoon.

• Matters under a protective order, or where the order’s scope has not been confirmed.
• Clients who have restricted AI use in an engagement letter or by written instruction.
• Categories the firm has decided to exclude outright, such as health information, financial account numbers, or government identifiers, unless the setup carries the agreements that work requires.
• Anything the responsible professional has not confirmed is covered by the firm’s vendor agreements.

Pair the list with a default. A policy that lists exclusions without stating what happens to an unlisted edge case leaves the edge case to a guess, and the guess will be made by whoever is busiest.

Sample clause. The excluded-matter list is maintained by the managing partner and reviewed quarterly. Material on that list may not be entered into any AI tool, in whole or in summary. Where a matter or document type is not addressed by this policy, it is treated as excluded until the managing partner confirms otherwise in writing.

This is the section that keeps a firm out of trouble, and it should be the most specific one in the document. Name the categories of work product, name who signs off on each, and state plainly that nothing goes out automatically. ABA Formal Opinion 512, issued in July 2024, applies the duties lawyers already carry to generative AI: competence, confidentiality, communication, supervision, candor, and reasonable fees. None of those are new obligations, and none of them are satisfied by a tool.

Citation checking deserves its own sentence rather than a clause buried in a paragraph. It is the failure mode with the shortest path from convenience to sanction, and a policy that names it explicitly gives a junior person permission to take the time.

Sample clause. No AI-assisted work product leaves the firm without review by the responsible attorney. This applies to client correspondence, filings, letters to counterparties, and anything submitted to a court or agency. Every legal authority produced with AI assistance is verified against the source by a person before it is cited. Automatic sending of AI-drafted material is prohibited without exception.

Record the configuration, with the date it was checked, so the answer to a client’s question is a document rather than a recollection. As of August 2026, Anthropic does not train models on Team and Enterprise content by default, and Enterprise adds custom data retention, audit logs, SCIM, and a HIPAA-ready configuration option. Those are facts about a plan, and plans change, which is why the date belongs in the policy next to the fact.

One configuration detail is easy to miss and worth writing down. Anthropic documents that history from a Claude Cowork session running locally is stored on the user’s own computer, outside standard retention, where an administrator cannot centrally manage or export it. A firm with retention obligations should decide deliberately whether local sessions are permitted at all. What each control actually does, in plain terms, is set out in Is Claude safe for client data?.

Sample clause. The firm uses Claude Team, on which Anthropic does not train models on firm content by default, confirmed against vendor documentation on the date recorded in Appendix A. Data retention, access, and connector permissions are configured by the administrator named in Appendix A and reviewed at least annually. The configuration and the date of each review are recorded in that appendix.

Decide the rule once, in advance, and name the person who decides the exceptions. Formal Opinion 512 discusses when the duty of communication with a client is triggered by AI use, and the analysis turns on the specific use, the engagement, and the jurisdiction. A guide cannot settle that for a particular firm. What a policy can do is stop the question being answered differently by four people under deadline.

Three disclosure surfaces are worth naming separately, because they have different answers: clients, courts and agencies with their own standing orders, and counterparties. One rule written for all three will be the wrong rule for at least one.

Sample clause. The firm’s default is to describe its use of AI tools in the engagement letter. Any disclosure beyond that default, and any court or agency requirement regarding AI-assisted filings, is decided by the responsible attorney, who confirms current requirements in the relevant jurisdiction before filing. Standing orders on AI use are checked at the start of each matter.

Access ends the day the job does, and the policy should say so with the same flatness as the rest of the document. This section is short, and it is the easiest one to leave out, because a small office that adds seats one at a time has no natural moment to write down how a seat is removed.

Cover the mirror image too. A policy that only handles departures leaves onboarding undefined, and an unsigned new hire with a provisioned seat is the same exposure viewed from the other end.

Sample clause. AI tool access is provisioned only after this policy is signed, and only to the matters and connected systems the role requires. On separation, AI accounts are deactivated on the final working day, together with all other firm access, and connected tool permissions are removed. Deactivation is recorded on the offboarding checklist.

Everyone with an account signs it, before the account exists. That includes partners, and it includes contract and temporary staff, whose document-heavy work is exactly what an AI tool gets bought for.

1. 01Everyone with an AI account signs before the account is provisioned, and new hires sign during onboarding.
2. 02One named person owns the policy. A document owned by a committee is a document nobody updates.
3. 03Review annually at minimum, and immediately whenever a plan tier changes, a new tool is approved, or a vendor changes its data terms.
4. 04Record the review date on the document itself, so an outdated policy is visible rather than assumed current.

Training is the other half of adoption. A signed policy that was never explained produces compliance on paper, and someone who has never had the excluded-matter list explained to them cannot apply it.

A policy is a management document. Whether it satisfies a particular bar rule, a regulator, a client contract, or a court’s standing order is a question for the professional responsible in that jurisdiction, and the answer varies by all four. A firm in a regulated practice should have counsel review the finished document before it circulates.

The other limit is scope. This document governs how people use approved tools; it does not decide which workflows are worth automating or how access to a document store should be scoped in the first place. Those decisions come earlier, and the beachhead version of them is in Claude for law firms.

Justo writes and delivers this document as part of a flat-fee engagement, alongside secure setup and training on your real work, from $3,500. Justo is attorney-founded; one of our three owners is a practicing attorney in Miami. Scope is on the pricing section of our main page.