Agile for AI work
Adapting agile practices to AI features: hypothesis-driven backlog items and time-boxed spikes, sprint goals based on evaluation results, a definition of done for AI, making quality visible, handling work that doesn't fit a sprint neatly, and building a team that learns AI together.
๐ฏ By the end of this lesson you will be able to:
- Write AI backlog items as hypotheses and use time-boxed spikes for uncertain work
- Apply an AI-specific definition of done and make quality trends visible
- Help teams learn AI skills together and adopt AI tools with good working agreements
In this lesson
Why agile suits AI, with adjustments
Agile was designed for uncertainty: short cycles, frequent feedback and changing plans based on what you learn. AI work is even more uncertain than usual software, because you often don't know whether something will work until you try it on real data. The principles fit well, but a few practices need adjusting.
| Usual practice | Adjustment for AI work |
|---|---|
| Stories describe features to build | Many items are experiments with a hypothesis and a measure |
| Estimates in points for known work | Time-boxes for uncertain work; estimate building only after the spike |
| Done = code complete and tested | Done includes evaluation thresholds, failure paths and monitoring |
| Velocity shows progress | Quality trends on the evaluation set show progress too |
| Demo shows working features | Demo shows typical and failure cases plus the evaluation results |
Hypotheses and spikes
A hypothesis-driven backlog item
We believe that adding the product manual section to the prompt context will raise the "answer correct" score from 72% to at least 80% on the 150-case support evaluation set. Time-box: 3 days. If < 76%, drop it and try reranking instead.
Clear hypothesis, a measure, a time-box and a decision rule. The team can't "fail"; it can only learn.
- Spike: a time-boxed investigation that answers a question, such as "Can the model extract invoice fields from scanned PDFs with โฅ 90% accuracy?"
- Plan spikes before committing to the build stories that depend on them.
- Make the spike's output a decision and evidence, not production code.
A definition of done for AI
- Evaluation set run; results meet the agreed thresholds with no slice regressions.
- Zero-tolerance checks pass (personal data, unsafe actions, harmful content).
- Failure paths tested: low confidence, nothing found, AI service down.
- Prompt, model and configuration versions recorded.
- Cost per request measured and within budget.
- Monitoring and feedback capture in place.
- Users know it's AI and can edit, reject or report outputs.
Sprint goals and planning
- Set sprint goals as outcomes: "Reach 80% correct on billing questions" rather than "Improve prompts".
- Expect to change the plan mid-sprint when an experiment fails; that's the process working.
- Balance experiments with delivery work so the team ships something usable regularly.
- Account for waiting time: expert labelling, approvals and data access often sit outside the team's control.
Building a team that learns AI
- Learning time: protect a small amount of time each sprint for AI skills, such as this course.
- Show and tell: short sessions where people share a prompt, a tool or a failure they learned from.
- Pairing: pair people with AI experience with those without, across roles.
- Shared prompt library: keep useful, tested prompts where the team can find and improve them.
- Working agreements: agree how the team uses AI tools (next section).
Working agreements for AI tools
A team's AI working agreement
1. We use only company-approved AI tools and never paste restricted data. 2. We review all AI-generated code and tests as if a new colleague wrote them. 3. The author is responsible for anything they commit or send, AI-assisted or not. 4. We mention significant AI help in pull requests so reviewers know where to look harder. 5. We share useful prompts in the team library and failures in show-and-tell. 6. We review this agreement every quarter.
Clear, short and owned by the team, not imposed from outside.
What this means for your role
Developer
Write experiments with hypotheses and time-boxes, and share quality results at every review, including what didn't work.
Tester / QA
Own the evaluation dashboard the team reviews each sprint, and keep the AI definition of done honest.
Business Analyst / PO
Write backlog items with hypotheses and measures, and keep acceptance criteria tied to evaluation thresholds.
Tech Lead / Architect
Help the team break uncertain AI work into spikes and decide when evidence is strong enough to build.
Project / Delivery Manager
Report progress as quality trends and evidence, and expect plans to change as experiments succeed or fail.
Scrum Master / Agile
Coach the team on hypotheses, spikes, the AI definition of done, learning time and working agreements, and make AI adoption safe for everyone.
Try it: adapt your team's practices
- Rewrite one AI backlog item as a hypothesis with a measure, a time-box and a decision rule.
- Draft three additions to your team's definition of done for AI features.
- Draft a five-line AI working agreement and bring it to your next retrospective.
Key takeaways
- Agile principles fit AI well; adjust backlog items, estimates, done criteria and demos for uncertainty.
- Write AI work as hypotheses with measures, time-boxes and decision rules, and use spikes before committing to builds.
- Use an AI definition of done and show quality trends at every review.
- Build team learning and clear working agreements, and make AI adoption safe and fair for everyone.
Key terms
- Spike
- A time-boxed investigation to answer a question or reduce uncertainty before committing to build.
- Hypothesis-driven development
- Framing work as a testable belief with a measure and a decision rule.
- Definition of done
- The shared checklist a piece of work must meet to be considered complete.
- Working agreement
- Rules a team agrees for how it works together, including how it uses tools.
Check your understanding
Answer all 5 questions. Score 3 or more to complete the lesson.
1. Which is a well-formed AI backlog item?
2. What should a spike produce?
3. Which item belongs in an AI definition of done?
4. An experiment fails mid-sprint. What is the right response?
5. Who is responsible for AI-assisted code that someone commits?