Law firm AI governance: bar rules and the malpractice carrier questionnaire.

There is a lot of confident writing on this topic and not much survives a check. The most widely circulated claim, that more than 60 percent of legal malpractice carriers now ask about AI on their applications, traces to a vendor FAQ with no source attached, so we are not repeating it. Same for the claim that CNA added a supplemental AI questionnaire at renewal: primary-source review shows an ethics bulletin, which is a different thing.

What is documented is narrower and more useful. AmTrust's lawyers professional liability application, form LPLPRO-APP-01 0523, asks whether the firm allows the use of artificial intelligence software to draft documents and requests a description if the box is checked. In primary-source review of US LPL forms, that is the one verbatim AI question located.

The nearest documented read is from an adjacent line of business: Aon's accountants risk control lead, speaking in April 2026 about CPA firm underwriting, described underwriters asking whether firms use AI, police it, and have protocols in place, and expected more detailed questions within a year or two.

Underwriters ask about a risk once it starts producing claims, and that threshold was crossed in 2026. EPIC Insurance Brokers released its 16th Annual Lawyers' Professional Liability Claims Survey on May 21, 2026, drawing on senior claims executives at 13 LPL insurers that collectively cover more than 80 percent of Am Law 200 firms. Seven of the 13 reported an increase in AI-related claims over the past year, alongside the first rise in overall claim frequency in five years. EPIC's Eileen Garczynski put accountability in one line: the duty of competence cannot be outsourced to an algorithm.

As of a primary-source review dated May 2026, no major US lawyers professional liability writer, including CNA, Travelers, Chubb, Hartford, Markel, ALPS, and the state mutuals, had publicly filed an explicit AI exclusion on LPL paper. Filed AI exclusions exist on other lines: W.R. Berkley's form PC 51380 00 (06-24) carries an absolute artificial intelligence exclusion on management liability, not legal malpractice.

Read that as a window rather than an all-clear. The cyber market is the obvious analogue, and whether LPL follows it from silent coverage to questionnaires to exclusions is not yet on paper.

The reason to care about a checkbox is not the checkbox. Answers on an insurance application are representations, and in many states a material misrepresentation can rescind a policy retroactively without any intent to deceive. Applying Kentucky law, the Sixth Circuit in 2012 upheld rescission of an LPL policy in Continental Casualty Co. v. Law Offices of Melbourne Mills, Jr., PLLC, where the firm answered no on potential claims while a bar investigation was pending; the statute required the misrepresentation to be material and not intentional.

A firm that certifies it does not permit AI drafting, then discovers an associate ran a brief through a consumer chatbot, has a representation problem sitting underneath its next claim. The defense is not a stricter policy. It is knowing the answer, from a system, on the day the form is signed.

The confidentiality question has a physical answer and most firms cannot give it. Four things need to be contractually verifiable rather than believed after a demo: where matter content sits while a model processes it, whether that is inside your tenancy or a vendor's, whether any provider retains inputs or trains on them, and which agreement makes each enforceable by name and clause.

The consequence is a single controlled egress point. Every AI call leaves through one service the firm operates, which holds the credentials, applies the retention terms, strips what should never leave, and refuses anything not on the approved list. Pin the provider and model version there, so a vendor's quiet substitution shows up as a configuration commit.

Model Rule 1.0(k) defines screening as the isolation of a lawyer from a matter through timely procedures adequate to protect the information, and Rule 1.10 lets a proper screen avoid imputation. Your document management system enforces that through access control lists on matter workspaces; a retrieval layer on the same repository can flatten it in one indexing run.

Index-time filtering, where documents are excluded when the index is built, breaks the moment a wall goes up after ingestion, which is exactly when walls go up. The correct pattern is query-time authorization: every retrieval evaluates the asking user's live entitlements in iManage, NetDocuments, Clio, or Litify on Salesforce and returns nothing they could not open by hand. Log the exclusions too, because an answer that silently dropped three screened matters looks identical to one where nothing was screened.

This is the artifact everything else resolves to. Vendor consoles record usage per seat because seats are what vendors bill on; bar rules, outside counsel guidelines, and carriers all care about matters.

The record needs enough fields for a stranger to reconstruct one event a year later without interviewing anyone: matter number, user, tool and model version, task category, the documents or fields sent, the destination environment, whether that client's guidelines permitted it, the reviewing lawyer, the outcome, and timestamps. Written once at the point of the call, those ten fields turn an audit request into an afternoon rather than a two-week investigation that produces an estimate. Store it where your matter records live and make it exportable; a professional liability tail outlives any tool you are running today.

Every firm will tell you a lawyer reviews the output. Johnson v. Dunn is what that assurance is worth without a mechanism, and the firm there had a policy. Review that lives in a document is a habit; review that lives in a workflow is a control.

Make it structural. AI-assisted work product carries a status until a licensed attorney clears it, the clearing attorney is captured by identity, not initials in a comment, and the attestation is one sentence they can stand behind. Handle citations as their own gate: every authority checked against the primary source by a person, logged, before signature. The gate is also where the fee record comes from, so Rule 1.5 becomes a query.

With that layer running, the renewal answer stops being a drafting exercise. The packet is the dated policy and its version history, the approved and prohibited tool lists, the vendor diligence file with retention terms named by clause, training completion figures, the screening confirmation, the matter-level activity report, the incident procedure, and a named accountable partner.

Attach what exists and mark what does not as a commitment with a date. A dated commitment reads as credible; a present-tense description of a capability you lack ends a relationship.