AI-Focused Retrospectives for Team AI Working Agreements

AI already appears in daily software and product work. It drafts notes, explains code, summarizes research, suggests tests, and helps people get past blank-page work.

Across a team, access to an AI tool can spread well before anyone writes shared rules that explain how people should use it in shared work.

One person uses AI to draft the customer-facing copy they are responsible for shipping. Another avoids it under unclear rules. Someone else pastes sensitive context and misses the risk. A useful prompt pattern might help one person while staying hidden from teammates who could reuse it in their own work.

An AI-focused retrospective turns those scattered habits into a practical team AI working agreement.

When to run an AI retrospective

Run one when AI use is outpacing the team's norms. That can happen after a new tool rollout, a policy change, a quality issue, or a sprint where AI clearly changed how work moved.

Use the session to decide how this team will use AI in this workflow and where human ownership sits.

Start with real examples

Opening with "What do we think about AI?" invites abstract opinions. Ask people for recent examples from work the team completed during the current sprint or project.

• Where did AI save time this month?
• Where did AI produce an output that someone on the team had to correct before the work could continue?
• Where did AI make review harder?
• Where are the rules unclear enough that people feel unsure whether AI use is allowed?
• Which useful AI case should the team know?

Group the answers into three buckets: encouraged uses, guardrailed uses, and off-limits uses, so the team can see which behavior belongs in each workflow. The buckets keep the conversation grounded in behavior and specific workflows.

Clarify human review

The first AI working agreement is simple: AI can assist while humans own the outcome.

Make that ownership explicit. Decide where AI may draft, summarize, transform, suggest, or explore, then define who reviews the work before it ships.

• AI can draft release notes, and a human reviews factual accuracy before publishing them.
• AI suggests test cases. QA owns final coverage and risk assessment.
• AI can explain unfamiliar code while engineers verify the described behavior directly in the codebase.
• Keep secrets, private keys, and unapproved customer data out of every AI tool the team uses during its work.

Write the team AI working agreement

End the retro with a short agreement people can use during the work. Five to seven rules are enough. Long policies slip from memory quickly, and people start skimming past the rules when they need an answer during work that day.

• Encouraged uses: where AI should become normal.
• Guardrailed uses: where AI is allowed after human review or through an approved tool in the team's workflow.
• Prohibited uses: AI stays out.
• Review expectations: name the person who checks accuracy, privacy, quality, and decisions before the work moves forward.
• Sharing habits: how prompts and lessons travel.
• Next review date: when the team revisits the rules.

Make the agreement visible

An AI retro that ends with "we should use AI better" leaves the workflow untouched. Turn each part of the agreement into a small, owned action that can move through the team's normal work.

Example: a QA team uses AI for first-pass test ideas. Humans own risk analysis, coverage, and final test design.

Track whether the agreement improves the workflow: fewer review loops, clearer handoffs, safer use of internal context, and more confidence from people who were unsure what was allowed.

Use HeyRetro to build custom AI-retro prompts, collect anonymous input, and turn agreement updates into action items. Adapt a template with columns for opportunities, risks, standards, and experiments.

Teams that need a stronger action habit can read how to make retro actions stick first.