Product names, plans and features change often. Check the vendor’s current documentation before relying on any detail here, and apply your own jurisdiction’s professional-conduct rules.
We have spent four posts on making AI genuinely useful inside a firm. All four move client data somewhere it was not before. That is not a reason to stop, it is a reason to write down the rules, once, in a form people actually apply.
Most firm AI policies do not survive contact with a busy Tuesday. They are too long, too abstract, and written to satisfy a hypothetical regulator rather than to be followed by a paralegal deciding right now whether to paste something.
Why most policies fail
Three failure modes, and nearly every unusable policy has at least two.
It is about categories, not tools. "Generative AI tools must be used in accordance with professional obligations" tells nobody anything. The person deciding needs to know whether they may put this document into that specific product on that specific plan.
It is long. Past about two pages, a policy stops being a reference and becomes a document that was circulated. If it cannot be held in mind while working, it will not be applied while working.
It lives in a folder rather than in the setup. A rule that exists only as a document depends entirely on memory and goodwill. A rule reflected in how the tools are configured is followed by default.
The best AI policy is mostly enforced by the configuration, and the document just explains why.
Tier the data first
Before anything else, sort your information into tiers and say explicitly which tools each tier may reach. This single table prevents more incidents than the rest of the policy combined, because almost every real slip is someone making this judgment alone, in a hurry, with nothing to apply.
A structure that works for most firms:
- Tier 1, public or non-client. Published material, research questions, your own marketing, general legal questions with no client facts. Any approved tool, no restriction.
- Tier 2, client work, de-identified. Substantive drafting where names, identifiers and distinguishing facts have been removed. Approved business-plan tools only.
- Tier 3, identified client material. Real matter documents with client identity attached. Only tools on the approved list, on a business or enterprise plan with training on firm data switched off, and only where the client engagement permits it.
- Tier 4, restricted. Matters that do not go to any external service: sealed material, certain criminal and regulatory work, anything under a protective order or a client instruction that forbids it. Locally hosted models only, or no AI at all.
Two things make this work. Name the actual products in each tier, not "approved tools". And make sure Tier 4 has somewhere to go, which in practice means a local model, because a tier with no permitted route is a tier people quietly ignore.
The seven sections
What we put in every policy we write. This is the whole structure, and it fits on two pages.
1. Scope and who it binds. Everyone: partners, associates, paralegals, contractors, temporary staff. Contractors are the usual omission and the usual problem.
2. Approved tools. A named list, with the plan each is approved on, because the plan is what determines whether your data trains a model. "ChatGPT" is not an entry. "ChatGPT, Business plan, firm SSO" is.
3. The data tiering. The table above, as the centre of the document.
4. What requires approval, and from whom. One named person. A route for requesting a new tool, and a commitment to answer within a few days. Slow answers are what drive people to unapproved tools, so the response time is a security control.
5. Verification. Covered below, and it is the section that does the most work.
6. Client-facing obligations. Whether your engagement letters address AI use, whether any client has instructed otherwise, and who checks before a new matter opens. Some clients now impose their own terms, and those override your defaults.
7. What happens when it goes wrong. A named person to tell, a commitment that reporting promptly is treated as the right thing to do, and a short account of what the firm will do. A policy with no route for reporting a mistake teaches people to conceal them.
The verification rule
The one section to get exactly right.
A general model will produce a confident, fluent, entirely wrong answer. Not occasionally, and not only about obscure things. Citations that do not exist remain the standard example because they are the one that reaches the press, but the ordinary case is subtler: a summary that misses the operative clause, a chronology with a transposed date, a recitation of a rule that was amended last year.
The rule we write, near enough verbatim:
No AI output goes to a client, a court, or any third party until a
person with the relevant competence has verified every factual and
legal assertion in it against a primary source.
Verification means reading the source. It does not mean the output
seemed plausible, and it does not mean asking the model whether it
is confident.
The person who sends it owns it. "The AI produced it" is not an
explanation available to anyone in this firm.That last line is the one that changes behaviour. Everything else is process.
Attach it to the workflow rather than the document. A drafting skill that outputs its sources alongside its text makes verification a fifteen second job. A skill that outputs polished prose with no sources makes it a twenty minute one, and twenty minute checks get skipped under deadline. Design the tooling so the safe thing is also the fast thing.
What to leave out
Policies get ignored because they are bloated, so be disciplined about what does not go in.
- Explanations of how AI works. That is training material. A policy is rules.
- Long prohibition lists. A short, justified list of what is off limits gets followed. A long one gets dismissed wholesale, including the parts that mattered.
- Aspirational language. "Staff should exercise appropriate judgment" is not a rule, it is a way of writing a document without making a decision. Make the decision.
- Anything you will not enforce. A rule nobody checks teaches people which rules are decorative, and they will generalise.
Keeping it alive
Three habits, none expensive.
A named owner. One person who owns the approved list, the request route and the review date. Not a committee.
A quarterly review. Twenty minutes. What did people start using that is not on the list, what did we approve, what broke, what changed in the tools. Add a line, remove a line, re-date it.
A record of the decisions. When you approve a tool, write one line about why and on what plan. In a year, when a client asks how you assessed it, that line is the answer, and reconstructing it from memory is not.
The policy is not the interesting part of AI in a law firm. It is the part that lets the interesting parts happen without an incident, and it is genuinely a day's work to get right, not a quarter's.
We write these as part of every AI enablement engagement, drafted around the actual tools we set up and the practice areas they touch, rather than adapted from a template. That is the work.