Product design question
Design a citation system so lawyers can fully trust Harvey's outputs.
- Harvey
- Product design
- Hard
Practice this question out loud. An AI interviewer asks it, follows up like a real interviewer would, and scores your answer. Type or speak.
Start a mock interview on this question · Mock interview from a job description
What this question tests
Tests product design for trust: making an AI system's outputs independently verifiable to a professional audience that cannot afford to be wrong.
How to approach it
- Define the requirement precisely: every factual or legal claim Harvey generates must trace to a specific, checkable source document or case citation.
- Identify the failure mode being solved: lawyers cannot bill a client or file a brief based on a claim they cannot verify themselves.
- Design the citation mechanic: inline citations linked directly to the exact paragraph or page of the source, not just a document title.
- Add a verification layer: a one click way to open the source in context, so checking a citation takes seconds, not a manual search.
- Handle the no source case explicitly: if Harvey cannot ground a claim, it should say so rather than presenting an ungrounded statement as fact.
- Confirm with the interviewer whether the focus is case law citations, client documents, or both, since sourcing differs.
What a strong answer includes
- Treats citation as a core trust primitive, not a nice to have, matching how lawyers already work with Bluebook style pinpoint citations.
- Proposes visibly flagging low confidence or unsupported claims rather than hiding uncertainty, since a false citation is worse for trust than an admitted gap.
- Suggests a random sample audit process where firms can check Harvey's citation accuracy rate over time, building institutional trust beyond any single answer.
- Notes that citation format should match what lawyers already use, like Bluebook, so the output fits directly into existing workflows.
Common mistakes
- Designing citations as an afterthought or footnote rather than the central trust mechanism of the product.
- Not addressing what happens when no reliable source exists for a claim.
- Ignoring that lawyers need citation formats that match their existing professional conventions.
Likely follow-up questions
- How would you measure citation accuracy at scale?
- What would you do if Harvey cited a real case that does not actually support the claim?
- How would this differ for transactional work versus litigation?
More product design questions
- How would you improve Harvey's Vault for bulk document review at large law firms?Harvey · Product design · Hard
- Design a Workflow Builder that lets non-technical lawyers automate recurring tasks.Harvey · Product design · Medium
- A top law firm says Vault’s AI search is promising, but attorneys still avoid using it in high-stakes matters. How would you diagnose whether the trust gap comes from retrieval quality, citation and explanation UX, permissioning concerns, workflow fit, or change management, and how would you prioritize the first product changes?Harvey · Product design · Hard
- Several top law firm customers say they will not expand Harvey without stronger identity and access management. How would you prioritize and scope SSO, SCIM, RBAC, and audit logging for the first two releases of Command Center, and how would you balance enterprise security requirements against implementation complexity and admin usability?Harvey · Product design · Hard
- Lawyers will rarely ask for 'better queuing' or 'stronger isolation,' but they will care deeply about accuracy, latency, reliability, auditability, and security. How would you uncover those needs and translate them into clear infrastructure product requirements and prioritized engineering work? Be specific about discovery methods, requirement framing, and how you would separate must-haves from nice-to-haves.Harvey · Product design · Hard
- You have 4 weeks with a strategic customer in a new professional-services vertical to identify one workflow for an AI-agent pilot. How would you run discovery, score candidate workflows, and decide whether to solve the problem by configuring Harvey’s existing capabilities, building a reusable extension, or not pursuing a bespoke solution at all?Harvey · Product design · Hard
More questions from Harvey
Learn the skill behind it
Chapters of the AI PM course that teach what this question tests.
- Chapter 4: Discovery and strategy for AI products
- Chapter 7: AI UX and human oversight: design for a system that is wrong sometimes
- Chapter 14: Get the job: the AI PM interview loop