Question 1

How do you explain Generative AI to a non-technical stakeholder in under a minute?

Accepted Answer

I describe it as a prediction engine, not a truth engine. It has read an enormous amount of text and predicts the most plausible next words for a given prompt. That makes it fast and fluent, and it also means it can be confidently wrong, because plausibility and accuracy are not the same thing. So I frame the operating rule that follows from it: fast drafting, human verification, and no unverified output in anything that carries a decision or a customer commitment.

Question 2

What is the first rule you teach a team about using public AI tools?

Accepted Answer

Treat the prompt window as a public whiteboard: no proprietary source code, customer data, credentials, or unreleased commercial terms. But a rule nobody can follow gets ignored, so I pair it with a supported path — an approved enterprise tool with a no-training-on-our-data agreement — and I get the boundary written into the team's working agreements and the Definition of Done, so it is inspected rather than remembered.

Question 3

Your organisation wants to add three AI coding agents to a Scrum Team of five Developers. Are the agents members of the Scrum Team?

Accepted Answer

No. Agents are capacity, not members. Scrum accountabilities are held by people, because accountability means someone can be asked to explain a decision and can change their behaviour afterwards — an agent can do neither. Practically, that means the five Developers remain accountable for the Increment and for everything an agent produced on their behalf. The useful question is not whether the team is now eight, it is whether five people can verify the volume of work that three agents generate.

Question 4

An AI agent generated a pull request that a Developer approved. It caused a production outage. Who is accountable, and what do you change?

Accepted Answer

The Developer who approved it, and the team who owns the Increment. That is the answer regardless of who wrote the code. But blame is not the interesting part. The interesting part is why a review passed something this dangerous, and the usual cause is review volume: when agents produce more change than humans can meaningfully read, approval degrades into rubber-stamping. So in the retrospective I look at review load per person, and the fix is usually a limit on how much unreviewed generated change can be in flight, plus stronger automated gates for the risk classes a human reader reliably misses.

Question 5

How does agent-assisted work change Sprint Planning and the Sprint Backlog?

Accepted Answer

The Sprint Goal and the Sprint Backlog do not change in nature — they are still the Developers' plan and commitment. What changes is where the uncertainty sits. Generation time collapses; review, integration, and validation time does not, and that is now the dominant cost in most items. So during Planning the team estimates the verification effort rather than the typing effort, and they decide item by item where an agent is appropriate. Items touching security, money, or data migration are usually not good candidates, regardless of how simple they look.

Question 6

Your team burns through its monthly AI token budget by the second week of a four-week period. How do you handle it?

Accepted Answer

I make the constraint visible to the team, because a budget nobody can see is a budget nobody can manage. First I get consumption broken down by activity, and the pattern is usually concentrated — a small number of workflows, often agents retrying against poorly specified tasks, consume most of it. Then the team treats it like any other capacity constraint in the retrospective: which uses earn their cost, which should move to a cheaper model tier, and where the spend is buying rework rather than progress. I would rather the team make that trade-off than have finance impose a cap they do not understand.

Question 7

Since adopting AI coding tools, your team opens twice as many pull requests but cycle time has got worse. What is happening?

Accepted Answer

Writing code got cheap; verifying it did not. So the queue moved downstream and the team is now bottlenecked on review, testing, and integration. I would show the team where the wait time actually accumulates — usually items sitting in review, not in development — and then treat review as a workstation with a limit rather than an afterthought people do between other tasks. In practice that means capping concurrent items awaiting review, making review a planned activity, and giving generated changes smaller batch sizes so a human can genuinely read them.

Question 8

Leadership wants to use an AI model to predict delivery dates more accurately than 'yesterday's weather'. What do you advise?

Accepted Answer

I would push back, gently. You do not need machine learning to forecast delivery — you need throughput history and a Monte Carlo simulation, which gives you a probabilistic answer like 'eighty-five percent confidence by the twelfth' from a spreadsheet's worth of data. A learned model trained on a few dozen noisy Sprints mostly manufactures false precision, and it is far harder to explain when it is wrong. The real forecasting problem in most organisations is not the algorithm anyway. It is unstable scope and undersized batches, and no model fixes that.

Question 9

After an AI tooling rollout, your DORA metrics improve sharply. Deployment frequency is up, lead time is down. How confident are you that things are better?

Accepted Answer

Moderately, and I would want two things before I believed it. First, the counter-metrics: change failure rate and time to restore. It is easy to improve frequency and lead time by shipping smaller, less-verified changes, and that shows up later as incidents. Second, I would check whether batch sizes shrank or whether work was simply split more finely to game the count. Once a number is used to judge an investment, it stops being a good measure of that investment, so I present the set together and I resist reporting any of the four in isolation.

Question 10

Some items take three times longer than the team's average. How do you use AI to investigate?

Accepted Answer

I pull the outliers and use an AI tool for what it is genuinely good at — clustering unstructured text and surfacing correlations across a hundred tickets faster than I could read them. It might surface that items touching a particular service or requiring one specific approval consistently run long. That is a hypothesis, not a conclusion, so I bring it to the retrospective as data and let the team explain it. The value is that the team argues about a pattern instead of an anecdote.

Question 11

A manager suggests running sentiment analysis on the team's chat history to spot morale problems before retrospectives. What do you say?

Accepted Answer

I would decline, and I would explain why in terms the manager cares about. Psychological safety is the precondition for the honest signal they are trying to buy. The moment a team learns their conversations are being scored, the conversations change — you destroy the measurement by taking it, and you get quieter, safer chat instead of better morale. Anonymity does not save it either; at a team of seven, anonymised sentiment is rarely anonymous. If the goal is a real read on morale, I would use something the team opts into and owns: a short pulse check they run themselves, where they see the results first and decide what to raise. The data belongs to the team, not to management.

Question 12

How can AI help you find organisational impediments the team has stopped complaining about?

Accepted Answer

I look at the work, not at the workers. Workflow data already records where items wait: state transitions, time in each queue, ageing items, how long approvals and environment requests take. AI is useful for reading the unstructured parts at volume — clustering blocked-item comments to show that a third of them mention the same access request. That gives me a defensible number to take to management, and it comes from the system's behaviour rather than from anyone's private conversations. Teams stop reporting impediments they believe are permanent, so the data often finds what the retrospective no longer mentions.

Question 13

A stakeholder suggests replacing the Daily Scrum with an AI-generated status digest posted to chat. How do you respond?

Accepted Answer

The suggestion assumes the Daily Scrum is a status meeting. It is not. It is the Developers inspecting progress toward the Sprint Goal and adapting their plan for the next day, and no digest can do the adapting. That said, the underlying complaint is often correct: if the fifteen minutes are being spent reciting yesterday's activity, that is a real anti-pattern. So I would take the digest and use it well — let a bot post the activity summary beforehand so nobody recites it, and use the event itself for what only humans can do, which is deciding what to change today.

Question 14

Stakeholder attendance at Sprint Review is falling. How might AI help, and where is the limit?

Accepted Answer

AI can lower the cost of the artefacts — assembling demo recordings into a short highlight summary for people who genuinely could not attend, or turning the session into a readable digest afterwards. That is worth doing. But it does not address why people stopped coming, which is usually that the Review has become a demo broadcast rather than a working session where stakeholders influence what happens next. So I would use the tooling for reach, and separately fix the event: fewer slides, real feedback, and a visible connection between what stakeholders say and what enters the Product Backlog.

Question 15

The team keeps landing on the same shallow root cause in retrospectives. How would you use AI in the session?

Accepted Answer

I use it as a divergence device, never as the answer. When the team stalls, I will put the problem statement to a model and ask for five causes we have not raised, then read them out as candidates to argue with. Most are wrong, which is fine — the value is breaking groupthink and giving quieter people something to push against. The team still decides which cause is real and what they will try. If I ever found myself reading the model's conclusion as the outcome of the retrospective, I would have facilitated badly.

Question 16

How should the Definition of Done change on a team using AI code generation?

Accepted Answer

The Definition of Done is the team's commitment to what makes an Increment usable, so it changes where the risk changed. On a generating team, the risk is unverified output, so I would expect items covering it: generated code has been read and understood by a named human, tests exercise behaviour rather than merely existing, dependencies introduced by generation have been licence-checked, and security scanning has run. The word doing the work is 'understood'. A Developer who cannot explain a change should not be able to mark it Done, whatever produced it.

Question 17

Your team has started shipping AI-generated code without really reviewing it, and the defect rate is climbing. How do you handle it?

Accepted Answer

I would bring the defect data to the retrospective and let the team draw the line themselves, because a rule I impose lasts one Sprint. But I would also ask why review got skipped, and the answer is usually structural rather than lazy — the team is being measured on output, or the volume of generated change simply exceeds what they can read in a Sprint. If it is the second, telling them to review more carefully is asking them to do the impossible, so the fix is smaller batches and a limit on unreviewed change in flight, not an exhortation.

Question 18

A senior developer refuses to use AI coding tools and says they do not trust them. How do you handle it?

Accepted Answer

I would start by taking the objection seriously, because senior engineers who resist a tool often have a specific reason — licence contamination, an unreviewable diff, a codebase the model handles badly — and that reason is useful information for the whole team. If it turns out to be discomfort rather than a technical concern, I would propose a bounded experiment on low-risk work for one Sprint, framed around what it does for them rather than for the company. If they still decline afterwards, that is fine. The team's commitment is to the Definition of Done, not to a tool.

Question 19

A team member's written communication is consistently blunt and is damaging relationships with a partner team. Would you use AI to help?

Accepted Answer

I would give the feedback myself, in person and specifically, because that is the part that cannot be delegated. Where AI helps is afterwards, as a private rehearsal tool they control — drafting a message and asking a model how it might land, or how to say the same thing collaboratively. That builds their own judgement over time. What I would not do is have AI write their messages for them, or send AI-mediated feedback about them, which turns a coaching conversation into a performance note.

Question 20

A management dashboard shows one developer contributes far less AI-assisted code than peers. Management wants to manage them out. What do you do?

Accepted Answer

I would challenge the measure before defending the person. Volume of code has never been a measure of value, and it is a worse one now that generation is cheap — the metric rewards accepting suggestions uncritically. Then I would show what that developer actually contributes, which in my experience is often the reviewing, mentoring, and design work that keeps everyone else's output safe. That work is invisible to a contribution dashboard and expensive to lose. If there is a genuine performance concern, it belongs in a conversation with their manager based on outcomes, not in a dashboard.

Question 21

Your junior developers now solve problems with AI in minutes but cannot explain the solutions. How do you respond?

Accepted Answer

This is the most serious long-term risk of AI-assisted delivery and it does not show up in this quarter's numbers. Juniors used to learn by struggling; if the struggle is removed, so is the learning, and in two years you have a team that can produce code nobody on it understands. I would make understanding explicit in how work is done — pairing on generated code, having the author walk someone through a change in review, and treating 'I can explain this' as part of Done. I would also protect some deliberately unassisted work, not for productivity, but because that is where capability is built.

Question 22

The Product Owner is drowning in refinement. How do you help them use AI without losing quality?

Accepted Answer

I coach the technique rather than doing the splitting for them. A useful starting move is asking a model to split a large item using a known pattern — Mike Cohn's SPIDR is the one I usually teach — which produces a draft in seconds. The Product Owner then does the actual work, which is deciding value and ordering, and the team still refines with them. The gain is that they arrive at refinement with a starting point instead of a blank page. The risk is a backlog that grows faster than the team can understand it, so I watch for that.

Question 23

Developers see documentation as overhead. How can AI reduce that friction responsibly?

Accepted Answer

AI is good at the documentation nobody wanted to write by hand — interface descriptions, change summaries, and 'what this function does' text generated from the code itself, with a Developer reviewing before it lands. Where it is unreliable is the documentation that carries intent: why a decision was made, what was rejected, what constraint forced the design. Generated text will invent plausible reasoning there. So I would automate the descriptive layer and keep architecture decision records human-written and short.

Question 24

Which AI features in work-tracking tools are actually worth turning on for a Scrum Team?

Accepted Answer

The ones that reduce duplicated effort and surface connections a human would miss: similarity detection that flags when a new item resembles a bug fixed last quarter, and linking related work across teams. Those pay for themselves. The ones I am wary of are automated estimation and automated prioritisation, because they take work away from exactly the conversations that create shared understanding. The test I use is simple: does the feature remove clerical effort, or does it remove a conversation the team needs to have?

Question 25

Developers are constantly interrupted by stakeholders asking for status. Can AI help?

Accepted Answer

Yes, for the low-value half of it. A bot answering 'what is the state of this item' from the work-tracking data removes a genuine interruption cost and nobody misses those exchanges. What I would not automate away is the other half — stakeholders asking questions that are really about scope, risk, or trust. Those interruptions are a signal that the Sprint Review or the transparency of the work is not doing its job, and hiding them behind a bot leaves the actual problem in place.

Question 26

Across eight teams, dependencies are discovered late and repeatedly. How could AI help, and what will it not fix?

Accepted Answer

Detection helps at the margin — text analysis across teams' backlogs can flag that two teams are describing the same integration in different words, which a manual read-through misses. I would use it as an input to planning conversations. But late dependency discovery is usually a symptom of how the organisation is split: teams organised around components rather than around value will keep generating dependencies faster than any tool can find them. So I would use the analysis to make the pattern visible, then argue for the structural change it points at.

Question 27

Your team works in a regulated environment. An auditor asks how you can prove a human reviewed AI-generated code before release. What do you show them?

Accepted Answer

I would show evidence produced by the work itself, not a policy document: review records tied to specific changes with a named approver, an origin marker on generated changes so the reviewed population is identifiable, pipeline gates that block release without a passing review, and the Definition of Done that specifies human understanding as a release criterion. The principle is that compliance evidence should be a by-product of how the team already works. If we have to reconstruct it for the audit, we do not actually have the control.

Question 28

Your client forbids AI tooling on their codebase for data-residency reasons, while a competing vendor is promising AI-driven delivery speed. How do you handle it?

Accepted Answer

I would separate the client's actual constraint from the blanket ban. Usually the concern is code and data leaving a jurisdiction, not AI as a category, and that leaves real options: an approved model deployed inside their environment, or using assistance on the parts of the work that never touch their data, such as test design and documentation. Then I would compete on what they are actually buying, which is predictable delivery and low defect rates, not tooling. If a competitor is promising speed by putting client code into an unapproved service, that is a risk the client should know they are being sold.

Question 29

If AI can schedule events, take notes, chase actions and produce metrics, why does this company need a human Scrum Master?

Accepted Answer

Because almost nothing on that list is the job. Those are the visible artefacts of the job, and I am glad to hand them over. What remains is the part that requires standing in a room and being accountable: telling a leader something they do not want to hear, holding a boundary when a team is being pushed into unsafe delivery, coaching a person through resistance, and changing how an organisation makes decisions. An agent has no standing to do any of that. If a Scrum Master's role in this company genuinely is scheduling and note-taking, then the honest answer is that the role was already being wasted.

Question 30

Tell me about an AI initiative on your team that did not work. What did you do?

Accepted Answer

Structure this as situation, what you expected, what actually happened, what you changed, and what you now do differently. Pick a real failure with a measurable cost — a tool that increased rework, an automation the team quietly abandoned, an agreement that did not survive delivery pressure. The credibility comes from naming your own contribution to the failure rather than the vendor's, and from the specific change you made afterwards. Avoid the disguised success where the failure conveniently made everything better.

AI Scrum Master Interview Questions and Answers

How do you explain Generative AI to a non-technical stakeholder in under a minute?

A strong answer

Answers that lose you the room

What is the first rule you teach a team about using public AI tools?

A strong answer

Answers that lose you the room

Your organisation wants to add three AI coding agents to a Scrum Team of five Developers. Are the agents members of the Scrum Team?

A strong answer

Answers that lose you the room

An AI agent generated a pull request that a Developer approved. It caused a production outage. Who is accountable, and what do you change?

A strong answer

Answers that lose you the room

How does agent-assisted work change Sprint Planning and the Sprint Backlog?

A strong answer

Answers that lose you the room

Your team burns through its monthly AI token budget by the second week of a four-week period. How do you handle it?

A strong answer

Answers that lose you the room

Since adopting AI coding tools, your team opens twice as many pull requests but cycle time has got worse. What is happening?

A strong answer

Answers that lose you the room

Leadership wants to use an AI model to predict delivery dates more accurately than 'yesterday's weather'. What do you advise?

A strong answer

Answers that lose you the room

After an AI tooling rollout, your DORA metrics improve sharply. Deployment frequency is up, lead time is down. How confident are you that things are better?

A strong answer

Answers that lose you the room

Some items take three times longer than the team's average. How do you use AI to investigate?

A strong answer

Answers that lose you the room

A manager suggests running sentiment analysis on the team's chat history to spot morale problems before retrospectives. What do you say?

A strong answer

Answers that lose you the room

How can AI help you find organisational impediments the team has stopped complaining about?

A strong answer

Answers that lose you the room

A stakeholder suggests replacing the Daily Scrum with an AI-generated status digest posted to chat. How do you respond?

A strong answer

Answers that lose you the room

Stakeholder attendance at Sprint Review is falling. How might AI help, and where is the limit?

A strong answer

Answers that lose you the room

The team keeps landing on the same shallow root cause in retrospectives. How would you use AI in the session?

A strong answer

Answers that lose you the room

How should the Definition of Done change on a team using AI code generation?

A strong answer

Answers that lose you the room

Your team has started shipping AI-generated code without really reviewing it, and the defect rate is climbing. How do you handle it?

A strong answer

Answers that lose you the room

A senior developer refuses to use AI coding tools and says they do not trust them. How do you handle it?

A strong answer

Answers that lose you the room

A team member's written communication is consistently blunt and is damaging relationships with a partner team. Would you use AI to help?

A strong answer

Answers that lose you the room

A management dashboard shows one developer contributes far less AI-assisted code than peers. Management wants to manage them out. What do you do?

A strong answer

Answers that lose you the room

Your junior developers now solve problems with AI in minutes but cannot explain the solutions. How do you respond?

A strong answer

Answers that lose you the room

The Product Owner is drowning in refinement. How do you help them use AI without losing quality?

A strong answer

Answers that lose you the room

Developers see documentation as overhead. How can AI reduce that friction responsibly?

A strong answer

Answers that lose you the room

Which AI features in work-tracking tools are actually worth turning on for a Scrum Team?

A strong answer

Answers that lose you the room

Developers are constantly interrupted by stakeholders asking for status. Can AI help?

A strong answer

Answers that lose you the room

Across eight teams, dependencies are discovered late and repeatedly. How could AI help, and what will it not fix?

A strong answer

Answers that lose you the room

Your team works in a regulated environment. An auditor asks how you can prove a human reviewed AI-generated code before release. What do you show them?