AI & Technical question

Suppose Harvey is building an Outlook or Word integration that drafts grounded work product from matter documents and email history. How would you partner with engineering to make architecture tradeoffs among retrieval depth vs latency, precomputed vs on-demand indexing, model quality vs cost, and strict permissioning vs implementation speed? What would you ship first, and why?

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

Tests technical product judgment on architecture tradeoffs for a grounded drafting feature, and the ability to sequence what ships first.

Clarify the target workflow first: drafting a memo from a specific matter's documents and recent email thread, not open ended search.
Choose precomputed indexing for the matter's core documents, since latency matters more than freshness for a drafting assistant used mid work.
Use on demand retrieval only for recent email, since it changes daily and a stale precomputed index would hurt trust.
Prioritize strict permissioning from day one, since a wrong document surfaced across a matter wall is a worse outcome than a slow or weaker draft.
Ship the Word integration first if attorneys spend more drafting time there, deferring Outlook until the retrieval and permissioning core is proven.
Pick a mid tier model with strong grounding over the largest model, if cost and latency at drafting scale become the bottleneck.

Ranks permissioning above speed to market explicitly, because a permissions leak in legal data is a trust ending event, not a bug to fix later.
Chooses precomputed indexing for stable matter documents but on demand for fast changing email, rather than picking one approach for everything.
States a clear ship first workflow, for example one document type in one integration, instead of trying to launch both integrations at once.

Optimizing purely for retrieval depth or model quality while treating permissioning as an add on to be handled later.
Trying to ship both Word and Outlook integrations simultaneously instead of proving the core retrieval and permission model first.

How would you validate that the permissioning logic actually respects ethical walls before launch?
What would you do if precomputed indexing turns out too stale for how attorneys actually work?