AI & Technical question

Suppose you are scoping a first product for in-house legal teams to review contracts with AI assistance. What requirements would you lock first around target use case, acceptable error rates, human-review steps, citations/provenance, and data handling, and what would have to be true before you let customers use it on real matters?

Practice this question out loud. An AI interviewer asks it, follows up like a real interviewer would, and scores your answer. Type or speak.

What this question tests

AI product judgment on locking the right requirements before letting a high-stakes professional tool touch real client matters.

Lock the target use case narrowly first: contract review for a specific, common category like commercial NDAs or vendor agreements, not general legal review across all contract types.
Define an acceptable error rate benchmarked against human paralegal or junior associate performance on the same narrow task, not an abstract accuracy target.
Require a mandatory human-review step before any output is relied upon, with the tool's role framed explicitly as a first-pass assistant, not a final reviewer.
Require citations and provenance for every flagged clause or risk, pointing to the specific contract language and, where relevant, the legal standard referenced, so a lawyer can verify quickly rather than re-read the whole document.
Set the launch gate: the tool must match or exceed the benchmarked human error rate on a held-out test set of real, varied contracts, reviewed by practicing lawyers, before it touches real client matters.

Narrows the target use case deliberately, a specific contract category, rather than promising broad legal review capability from day one.
Benchmarks acceptable error rate against actual human performance on the same task, giving a credible, defensible standard instead of an arbitrary accuracy number.
Requires per-claim citations and provenance as a hard requirement, directly addressing the verifiability lawyers need to trust and use the tool efficiently.
Sets a concrete, evidence-based launch gate, matching human benchmark performance on a real held-out test set, before allowing use on real matters.

Proposing a broad, general-purpose legal review tool instead of locking a narrow, well-defined first use case.
Setting an abstract accuracy target with no human benchmark or held-out validation before real-world use.

How would you build the held-out test set of real contracts without violating client confidentiality?
What would you do if the tool matched human accuracy on average but failed badly on a specific contract subtype?