Harvey PM interview process: rounds, domain bar, and what actually clears the screen

Harvey’s PM loop is compact but specific: a 30-minute recruiter screen, a 30-minute hiring manager interview, and a 2-hour final block of four back-to-back rounds. The filter is not legal knowledge. It is whether you can reason about a four-product platform serving 100,000+ lawyers and make decisions that reflect how legal work is actually structured.

Know the platform before any round

Harvey runs four products: Assistant (chat, drafting, document analysis), Vault (bulk cross-document review), Knowledge (legal and regulatory research with citations), and Workflow Agents (multi-step automation, no-code builder), plus 25,000+ custom agents tested across M&A, funds, procurement, labor, IP, and investigations. Candidates who describe Harvey as “an AI assistant for lawyers” are eliminated early. The correct frame: Harvey routes structured legal work to agents. The PM decides which tasks move fully to agents, which keep attorney judgment in the loop, and where proactive AI assistance crosses into interference.

Recruiter screen (30 min)

Platform familiarity and baseline product instinct. Harvey states explicitly you don’t need legal experience. The actual bar: genuine curiosity about how legal teams work, and the ability to ask the right questions to close gaps quickly.

Hiring manager screen (30 min)

Expect a product background question (a decision under uncertainty, something you killed) and a product sense probe inside Harvey’s world. The HM maps answers to three stated values: Decisiveness (act on clear judgment without complete information), Simplicity (the version that scales is the simpler one), and Job’s Not Finished (treat a shipped product as the start, not the end). Framework answers without a concrete judgment call are filtered here.

The final four rounds

Product sense. Sample question: “How would you improve Vault for M&A due diligence?”

strong

"The highest-leverage gap in Vault for M&A is exception escalation. Vault surfaces findings, but the associate still decides what flags the partner. A configurable severity classifier trained on firm-specific precedent would let the agent's output map to the firm's own risk taxonomy, compressing partner review time. That's where billable efficiency shows up in NPS and renewals. Viability: firms pay a premium because it encodes their IP into the product. The tradeoff: the classifier must be tunable per matter type. What's material in fund formation isn't material in employment litigation."

weak

"I'd make Vault easier to use with better summaries." No named user, no job-to-be-done, no tradeoff. Also disqualifying: "We need zero hallucinations before shipping." Harvey's co-founders have stated hallucinations are acceptable in legal AI when value exceeds error rate. Treating hallucination policy as binary fails the domain check.

Execution and metrics. Measurement plan for a new agent workflow, or an RCA on an unexpected metric shift. Harvey published a Legal Agent Benchmark, signaling they treat evals as a product artifact. Strong candidates reason about quality thresholds and error-mode taxonomy as product decisions, not engineering implementation details.

Behavioral. Stories mapped to the three values: a time you acted on clear judgment without full data, a time you cut complexity, a time you treated shipping as a starting point. Pre-packaged STAR answers applicable to any company are recognized and discounted.

Domain depth. Not a legal knowledge test. Harvey’s Applied Legal Researchers (20% of employees are practicing lawyers) sit inside product teams. What’s probed: whether you know what an ethical wall is, why matter-centric organization shapes permissions and retrieval, and how you’d distinguish in-house PM needs (cost per matter, risk exposure) from law firm PM needs (billable efficiency, realization rate).

Compensation and context

Staff PM: $215,000 to $300,000 USD base plus equity, San Francisco, three days in-office. Harvey is roughly 300 employees, $11 billion valuation, targeting a $1 trillion legal market at about 3% technology penetration.

The 2026 bar

Automating legal work at scale is technically feasible. Harvey proves it. The PM role is to decide what moves to agents, what keeps attorney judgment in the loop, and where proactive assistance becomes interference. Viability means partners at AmLaw 100 firms trust the output enough to put it in front of clients. Lovable means the product fits inside matter-centric, ethics-wall-constrained, precedent-driven legal workflows. That distinction is what every round is actually testing.