When To Hire AI Product Counsel, And How To Recognize One

What Does AI Product Counsel Actually Do On A Launch?

A pricing feature is two weeks from release. It personalizes an offer using a model trained partly on scraped reviews, it explains its reasoning in a sentence the user sees, and it runs in the EU. Your general counsel has three questions back and no time this week. AI product counsel answers that on Tuesday, in the design review, before the code that creates the problem is written.

The work is unglamorous and specific. Where did the training data come from and can that be shown a year from now. What does the model output claim, and what happens when it is wrong in the direction that costs a user money. Does the feature need a disclosure, and does the disclosure need to be in the interface rather than the terms. Is this a high-risk system under the EU AI Act's employment or credit categories, or is it plainly outside them, and which sentence of the file records that decision. Under that Act, core obligations for high-risk systems (risk management, data governance, documentation, human oversight, registration, post-market monitoring) apply from 2 August 2026, with penalties reaching 35 million euro or 7 percent of global annual turnover for banned uses ³. Those dates and thresholds are jurisdiction-specific and move; treat any date in an article as a prompt to check with counsel, not as advice.

The deliverable is rarely a memo. It is a decision recorded in the ticket, a rewritten UI string, a documented reason a feature launched in three markets and not five, and a model card with a section nobody on the engineering team wanted to write. The role reports into legal and lives in product's calendar.

When Is Your GC Not Enough For AI Features?

Three signals, and you probably have two of them. Shipping decisions wait on legal answers more than twice a month. A single feature raises data provenance, output liability, and a notice requirement in the same review, which the pricing release above manages on its own. And the general counsel has started answering AI questions by scheduling outside counsel, the honest move and also a recurring five-figure line item that scales with your roadmap.

Company size is a bad trigger. A twenty-person company shipping a model that touches hiring, lending, health, or education needs this person before a four-hundred-person company selling internal analytics does. Legal hiring reporting for 2026 describes sustained interim demand for product counsel and compliance work alongside a shortage of candidates who combine legal expertise with AI literacy ¹. That shortage is the reason to start looking a quarter early rather than the week the launch slips.

The distinction from privacy counsel matters at the point of hire. Privacy counsel owns a body of law and applies it across the company. Product counsel owns a surface, the thing you are shipping, and pulls in whatever law it touches, which on an AI feature usually includes privacy, consumer protection, IP, and sector rules at once. If you hire privacy counsel expecting product coverage you get excellent DPIAs and a launch calendar that still stalls.

Good AI Product Counsel Writes In Ship-Or-Don't Language

The separating trait is that this lawyer writes in ship-or-don't language. Ask for the last hard call they made and a strong candidate names the feature, the two options they gave the product lead, the one they recommended, and what they were wrong about later. A weaker candidate describes a framework and a risk matrix, which is a real skill and a different job.

The behaviors that hold up under a reference check all share one property: they attach to a shipping decision rather than to a body of law. A candidate worth hiring reads a model card and says which section is load-bearing and which is decoration. They separate training-data risk from output risk without being prompted, because the mitigations are unrelated, provenance and licensing on one side, guardrails, evaluation and user recourse on the other. They argue about interface copy down to word choice, since the disclosure that satisfies a rule is usually a sentence in a product surface rather than a clause in the terms. Within twenty minutes they ask what happens when the model is confidently wrong about the pricing feature specifically, not about AI features in general. And before the conversation ends they name a feature of yours they think is fine, which is what tells you they will not say no to everything and call it caution.

Performed fluency has its own signature. Vocabulary is current and unattached to a decision: RAG, agents, alignment, all correct, none of it connected to a launch that shipped or did not. The candidate cites regulation by name but not by scope, so the EU AI Act is invoked for a feature nowhere near the risk tiers it defines. And the writing sample is long. Product teams do not read long. The best sample you will receive is often a page and a half with a recommendation in the first paragraph, closely related to what a good AI-fluent paralegal produces when the work is genuinely reviewed rather than generated and forwarded.

The Best Candidates Spent Three Years Inside A Product Organization

Three backgrounds produce this person reliably, and the third one surprises people. Product counsel from a consumer platform who spent four years on features rather than contracts. Privacy counsel who kept getting handed the AI questions and taught themselves the technical half, a drift that shows up in career guidance as the AI governance and risk track ². And engineers or technical PMs who went to law school in their thirties, undervalued in resume screens because their legal tenure looks short.

The practice behind the skill is unromantic. The candidates who are good at this used the systems. They ran the model against their own drafting, watched it fabricate a case citation or an obligation, and built the habit of asking for the source before the sentence rather than after. They have opinions about where an assistant helps (first-pass issue spotting on a spec, comparing two versions of a policy, drafting the plain-language version of a notice) and where they refuse to delegate (the scope call, the interpretation, anything that gets recorded as a decision). That refusal is not conservatism. It is the thing you are hiring, and it is close to what a good AI output verification counsel does full-time on a larger team.

Where to look: AI startups two rounds ahead of you that just hired their second lawyer, in-house teams at consumer platforms with mature product-counsel functions, the IAPP's AI governance community, Tech GC and In-House Connect circles, and law-and-technology programs at Stanford, Berkeley, and Georgetown, which produce candidates who read papers without flinching. Recruiters will send you regulatory specialists; the productive redirect is to ask for people whose last three years were spent inside a product organization, whatever the title on the badge.

What Does AI Product Counsel Pay, And Where Does The Work Happen?

No published salary series exists for the title, so treat any precise number for AI product counsel as inference rather than data. The nearest defensible anchor is the in-house legal band, because the scarcity sits in the AI half rather than the seniority. One salary aggregator, as of 1 September 2026, puts United States legal roles at a $246,000 median total compensation, with the 25th percentile at $179,000 and the 75th at $325,000 ⁴.

Product counsel and privacy counsel sit toward the upper half of that spread on the same page. AI product counsel at a funded startup tends to land near the 75th percentile once equity is counted, because the candidates are scarce and usually being recruited out of a job they like.

Two practical notes. Equity does more work here than in most legal hires, because the person is being asked to take a title that is not yet a recognized rung on a career ladder, and upside is how that risk gets paid. And the band is regional in the ordinary way: the figures above are national, and a Bay Area or New York offer should be read against local comparables rather than the median.

On location, the work is remote-friendly with a hard exception. Design review is where the value is created, and design review is a live conversation. Fully distributed teams do this well when reviews are scheduled and recorded; hybrid teams do it badly when the product organization is in one room and counsel is on a screen nobody looks at. If your engineering team is on-premise three days a week, ask for two of those days. Classified, regulated, or on-premise-only deployments (defense, some health systems) impose their own answer, and candidates screen themselves out early when you say it plainly in the posting.

How Do You Close AI Product Counsel, And What Kills The Offer?

What closes this candidate is scope, stated concretely. They want to know they will be in the room before the design is set, that they can say no and have it hold, and that vendor contracts will not eat the week. Say which meetings they attend, name the product leads they will sit with, and describe the last decision the company made against short-term revenue on a legal judgment.

Three things kill the offer, all of them recoverable if caught early. Reporting the role into a general counsel who treats product review as approval theater, which the candidate detects in the third interview by asking who overrules whom. A compensation structure that pays them like a contracts attorney because that is the band the company already has. And a hiring process that never lets them touch the product, which reads as a company that will not let them touch it later either. Give them a real feature and an afternoon; the conversation that follows is the interview.

On assessment, resist the writing sample as the whole screen. An issue-spotting memo is exactly the artifact a model produces well, and you cannot tell which one you received by reading it, so the artifact tells you very little either way. What separates candidates is watching the work happen: hand them the pricing spec from the top of this piece, an assistant, and forty minutes, and see where they check a confident claim against something outside the conversation. The same instinct is what makes a good AI agent manager trustworthy with autonomous systems, and it is legible in a session in a way it is not legible in a document.

Read the evidence >