LinkedIn · Execution · Senior · 45 to 60 minutes

Launch a trusted AI assistant for jobs and posts

LinkedIn is a professional network where people build profiles, follow content, and search for jobs and career opportunities. Answer this product management case and get it graded line by line against a weighted rubric.

execution planning
quality evaluation
platform strategy
tradeoff judgment
cross-functional leadership

The situation

Your GM for Consumer Product has asked you to lead the launch plan for a new LinkedIn assistant that helps members explore feed posts and job listings by asking follow-up questions, assessing fit, and getting advice. Early demos built with large language models look promising on simple queries, and leadership wants to move quickly because member expectations are shifting fast. But this product touches high trust moments like job fit and career guidance, where hallucinations, harsh wording, or slow responses could damage confidence in LinkedIn. Several engineering teams want to build different use cases in parallel, but that creates risk of fragmented UX, duplicated tooling, and uneven quality standards. Assume you have one quarter to ship a limited beta on web and mobile, a fixed GPU budget, existing internal data exposed mostly through RPC APIs, and a requirement that the experience feel useful from day one. You need to recommend how to scope the first release, how to evaluate quality before ramp, and how to organize platform and feature work so the product can improve quickly after launch.

What to produce

Write a 2 page product execution memo.

Your answer must

Recommend the launch scope for the beta, including which member use cases to include first, which to defer, and why.
Define an evaluation plan with clear pre-launch gates and post-launch success metrics, covering factuality, safety, tone, coherence, and usefulness in multi-turn conversations.
Propose how the assistant should access LinkedIn data and outside information, including how you will handle structured tool or API calls and failures.
Set targets and tradeoffs for latency, cost, and capacity, and explain how those targets should influence model and feature choices.
Describe the team operating model, shared platform investments, and rollout plan needed to keep the experience consistent across surfaces and use cases.

Constraints

Most differentiated LinkedIn data is available through existing internal RPC APIs that were not designed for LLM use.
The beta must support web, iOS, and Android with a reasonably consistent assistant experience.
GPU capacity is limited and expensive, so extra model calls and long generations have meaningful cost and throughput impact.
Because the product affects career decisions, low quality or unsafe answers are not acceptable even if that means narrower initial scope.

Format

Use short section headings and include one small table for launch gates or tradeoffs.

How LinkedIn actually played it

LinkedIn engineers share what it took to ship an assistant for jobs and posts. Reaching the first 80 percent was fast, while evaluation, structured output and latency took the rest of the time.

Plan for the last stretch of quality to be slow
Build evaluation early
Balance speed, cost and capacity

Built from Musings on Building a Generative AI Product by Juan Pablo Bottaro and Karthik R..