GitHub · AI product · Mid · 45 to 60 minutes

What should GitHub's first AI assistant do?

GitHub is a platform where developers host code, collaborate, and use tools across the software development lifecycle. Answer this product management case and get it graded line by line against a weighted rubric.

Product scoping
Prioritization
Experimentation
AI safety judgment
Go to market

The situation

You are a PM on GitHub's emerging AI tooling team. Leadership wants a recommendation for the first productized experience built on large language models for software development. Early demos show the models can generate useful code in some cases, but outputs are inconsistent, sometimes insecure, and often need human review. At the same time, interest is growing from both individual developers and enterprise buyers who want GitHub to help teams code faster. Your job is to define the initial product area for a technical preview, not the long term AI vision. The team could aim broadly, such as helping across multiple steps of the software development lifecycle, or narrowly, such as a focused coding task inside existing workflows. There is also debate about where the experience should live, such as a web surface versus developers' current tools, and whom to target first, individual users or enterprise customers. Engineering warns that every additional use case, interface, and language behavior increases evaluation complexity and safety risk. Assume you need to recommend a plan leadership can approve this quarter, including how to learn quickly from real usage without creating trust or security problems.

What to produce

Write a 1 to 2 page product strategy memo for the technical preview.

Your answer must

Define the target user and the core job to be done, and explain why this should be the first AI problem GitHub tackles.
Recommend the initial scope of AI assistance, where in the developer workflow it should appear, and what should be explicitly out of scope for the first release.
Propose a launch sequence and initial customer segment for the preview, with reasoning about adoption, feedback quality, and operational risk.
Specify success metrics and a learning plan, including at least one qualitative signal and two quantitative signals.
Outline the main safety, security, and trust guardrails for generated suggestions, including how problematic outputs and user feedback will be handled.

Constraints

LLM output quality is promising but inconsistent, and developers must remain able to review and control what is accepted.
The experience should fit into developers' existing workflows rather than requiring a major behavior change.
Security, licensing, and offensive or unsafe output risks must be addressed before broad availability.

Format

Use short section headers. Make one clear recommendation, then briefly note one credible alternative you rejected.

How GitHub actually played it

Summarises how GitHub chose a narrow first problem for Copilot, revisited choices as models improved and filtered risky suggestions before scaling.

Let model ability set the scope
Revisit early assumptions
Launch to individuals before enterprise

Built from GitHub's Learnings from Building Copilot, an Enterprise LLM Application by Matt Saunders (InfoQ).