Three things happened this quarter, and together they are the reason your managing partner has a new item on the agenda. One, two of the country’s largest legal malpractice carriers added “written AI use policy” to their renewal questionnaires. Two, the general counsel of a major financial-services client sent a three-page data-handling addendum to every outside firm, asking for the firm’s AI policy as an attachment. Three, the state bar in your jurisdiction either released a formal opinion on generative AI or is about to. The through-line is that an AI use policy is no longer a forward-looking governance document. It is a prerequisite for doing business.
The good news is that a working policy is not the seventy-five-page epic the compliance industry would like to sell you. It is ten well-structured sections, a short list of named tools, and a review cadence the firm can actually keep. The policy most partners need in 2026 is short, specific, and written for the lawyers who have to follow it — not for the regulator who might read it someday.
What follows is the framework we walk partners through before writing a single line of policy language. It is not legal advice; every jurisdiction has its own wrinkles and your ethics counsel is the right person to sign off on final terms. But this is the skeleton that stands up under scrutiny, gets adopted by the partnership inside a month, and survives contact with the associates who will actually test its limits.
Why every firm is being asked for a policy right now
The pressure is coming from four directions at once, and each one asks for a different kind of document. A useful policy satisfies all four without pretending any of them does not exist.
Malpractice carriers want to know that the firm has thought about AI as a source of professional-liability risk. Their question is narrow: do you have a written policy, is it current, and does it address client-data handling and supervision? They are not grading the policy’s prose. They are checking a box that used to say “cyber training” and now also says “AI.”
Sophisticated clients — particularly in-house teams at financial institutions, healthcare operators, and regulated industries — are adding AI language to their outside-counsel guidelines. They want to see the firm’s policy on approved tools, client-data segregation, and disclosure before they hand over privileged material. Some already require a signed addendum. By Q3 of 2026, most will.
State bars and ethics committees are the slowest and the loudest. Their interest is primarily in Model Rules 1.1 (competence), 1.6 (confidentiality), 5.1 and 5.3 (supervision). A formal opinion in your jurisdiction may not require a policy, but it will almost certainly describe the elements of one. Aligning your policy with those elements is the cheapest form of forward compliance.
The lawyers inside the firm are the fourth and most consequential audience. Without a policy, every associate invents their own. That is how a summary-generation tool ends up running unreviewed discovery material through a consumer-tier account over the weekend. A policy is, among other things, the firm’s permission structure: what you may do, what you must do, and what is off-limits entirely.
What a working policy actually does
There is a version of this document that is compliance theater — aspirational language, broad prohibitions, no named tools, no owners, no dates. It exists to be produced on request and ignored the rest of the time. We are not writing that one.
A working policy does three things. It tells the lawyers what to do today, with specific tools and specific workflows. It draws a defensible perimeter that a regulator or carrier can understand without calling for a meeting. And it has an owner — a named partner or committee — with a review cadence and a path for updating the document as the firm’s tooling matures. If any of those three is missing, the policy does not survive the first associate who ignores it.
The ten sections every working policy needs
Ten sections is a target, not a ceiling. A firm with one AI tool in production might keep each section to a paragraph; a firm with a bespoke in-house stack may need an appendix for the technical architecture. But all ten belong in the document, in roughly this order.
- Scope and applicability. Who does this policy bind — partners, associates, paralegals, contract attorneys, outside co-counsel? What activities count as “AI use” for purposes of the policy? A one-paragraph definition avoids a dozen future arguments.
- Approved tools and deployment models. A short list of named systems, each with a deployment-model tag: firm-hosted, private tenant, or multi-tenant SaaS. Tools that are not on the list are not approved. “Contact the governance committee” is the only legitimate path to add one.
- Prohibited uses and data categories. The narrow list of things nobody may do with an AI tool, regardless of approval status. The most common entries: feeding privileged material into a consumer-tier account; generating legal analysis the attorney does not personally read and revise; using AI output as a substitute for judgment on a matter the lawyer has not studied.
- Supervision and review requirements. The operational language that mirrors Model Rules 5.1 and 5.3. Who reviews AI output before it reaches a client, under what standard, and what documentation the reviewer must keep.
- Client disclosure and consent. When does the firm tell the client it is using AI on their matter, in what form, and who decides. Most firms will settle on a default of disclosure by engagement letter for any material workflow, and affirmative client consent for anything that touches a privileged communication end-to-end.
- Data retention, logging, and audit. What the AI tool retains, for how long, who can retrieve it, and how the firm logs prompts and outputs. The right answer depends heavily on deployment model; this is the section where that choice pays off.
- Incident response. The steps a lawyer takes when something goes wrong — bad output that reached a client, a prompt that contained more than it should, a vendor breach. A simple decision tree beats a ten-page IR plan every time.
- Training and competence. The minimum training every lawyer using AI tools completes annually, and the recertification requirement for anyone onboarding a new tool. Competence is a Rule 1.1 obligation; in 2026, AI literacy is part of it.
- Third-party vendors and privilege protection. The due-diligence standard for any vendor whose tooling touches client data. This is the section that points to your separate vendor-due-diligence checklist — eight questions, no fewer — and names the partner who has to countersign before a contract is executed.
- Review cadence and amendment process. When the policy is reviewed, by whom, and how amendments are proposed and adopted. Twice a year is a reasonable baseline. A policy without a review date becomes inert within eighteen months.
Section-by-section: what to write, what to leave out
The section that most firms get wrong is the second one: approved tools. The temptation is to write it in the abstract — “AI tools that meet the firm’s security standards are permitted.” That is a non-policy. What the lawyers need is a named list: tool A for drafting, tool B for discovery summarization, tool C for intake triage. Next to each, the deployment model and the workflows it is cleared for.
The section most firms over-write is prohibited uses. Broad prohibitions read tough and produce nothing except selective enforcement. A short, specific list — written around the behaviors the partners have actually seen go sideways — produces compliance. “Do not paste opposing counsel’s production into a consumer-tier account” is worth more than a paragraph of abstract principle.
Supervision language should mirror the text of Rules 5.1 and 5.3 almost verbatim, because that is the language a regulator will read it against. The operational detail — who reviews, under what standard, with what documentation — goes in a short implementation guide beneath the policy. That separation keeps the policy stable while the operational detail evolves with tooling.
Client disclosure is where partners get nervous, and with some reason. The right default in 2026 is a short, plainspoken clause in the engagement letter: the firm uses AI tools to improve accuracy and efficiency, lawyer judgment governs every deliverable, and the firm maintains confidentiality and privilege standards consistent with its professional obligations. Anything more specific in the engagement letter becomes a negotiation item. Anything less, in front of a sophisticated client, reads as evasion.
Vendor due diligence is the section that ties the AI policy to the rest of the firm’s compliance infrastructure. The checklist is short — deployment model, data path, retention, sub-processor list, termination, audit rights, model training, incident history. If the firm cannot answer those eight questions about a vendor, the vendor is not ready to be approved, no matter how good the demo.
The thirty-day rollout that actually lands
A policy that takes six months to adopt will be obsolete by the time it is signed. The firms we see adopt quickly run a tight four-week cycle that forces decisions rather than inviting them to drift.
Week one: inventory. A single partner, with help from the IT lead, catalogs every AI tool currently in use across the firm. Paid accounts, shadow accounts, browser extensions, the lot. The output is a spreadsheet with tool name, owner, workflow, deployment model, and privilege exposure. Most firms are surprised at the length of the list.
Week two: triage. The partner and a small working group categorize each tool: approved as-is, approved with guardrails, pending vendor review, or discontinued. The goal is not perfection. The goal is a defensible baseline the policy can name on its first day in force.
Week three: draft and circulate. Ten sections, ten paragraphs, two pages of cover material. The draft goes to the partnership with a deadline and a short list of comment categories. Line edits are welcome; architectural rewrites are deferred to the next review cycle.
Week four: adopt and train. The partnership votes, the policy takes effect, and every lawyer completes a thirty-minute orientation keyed to the approved-tools table. Training is the step firms are most likely to skip and most likely to regret. The policy binds only the people who have read it.
Three traps that turn a policy into shelf-ware
The first trap is the aspirational tool list. The drafting committee includes products the firm is considering but has not yet vetted, because the list looks thin without them. Six months later, nobody can tell the difference between approved and aspirational. The fix is to keep the list honest and reconvene the committee when something new is ready to add.
The second trap is the unowned policy. Nobody’s name sits at the bottom. No review date. No process for amendment. The document reads like a memo to the file. The fix is to name a partner — title, not person, so the policy survives turnover — and set a review date no further than six months out.
The third trap is the compliance-only policy, written to satisfy carriers and regulators and nobody else. The associates read it once, notice that it says nothing about the tools they actually use every day, and ignore it thereafter. The fix is to write the document for the lawyers in the firm, with the regulator looking over their shoulder — not the other way around.
Where this intersects with the rest of your AI stack
A policy is a necessary piece of the firm’s AI governance, but it is not the whole picture. The choices that most determine whether the policy holds up are the ones the firm makes before the policy is drafted: which deployment model to adopt for work that touches privileged material, which vendors to take seriously, which workflows to build in-house and which to rent from the market. The policy documents those choices. It does not make them.
If the firm is about to publish a policy and has not yet made those decisions, the right sequence is to defer the policy by a few weeks and run the deployment-model analysis first. A policy that bakes in the wrong answer on deployment is harder to fix than a delayed policy that gets the answer right.
If you want an outside set of eyes on a draft — or on the deployment-model analysis that sits behind it — that is exactly the kind of thing a thirty-minute bottleneck audit can close out in a single call. Bring the draft, bring the approved-tool list, and we will walk it with you section by section.
