PSM-AI common mistakes
PSM-AI · Understand the traps and how to avoid them.
Answering that an AI agent or tool can hold a Scrum accountability - e.g. 'the AI can act as Product Owner', 'the bot is a team member', or 'the model owns the Sprint Goal'.
Vendor marketing and conversational interfaces make tools feel like colleagues, so candidates carry an anthropomorphic mental model into scenario questions about roles and accountabilities.
What to do instead
Anchor on the rule that Scrum accountabilities are held by people; AI is a tool people use, and a named human stays accountable for anything an AI produces or recommends. Re-read every role option and ask 'who is accountable if this is wrong?'.
Choosing options that assume a model remembers yesterday's conversation, 'knows' facts about the product, or gives the same answer twice to the same prompt.
Candidates reason about AI as if it were a database or a knowledgeable colleague, so they over-trust outputs and under-plan for verification, grounding and non-determinism.
What to do instead
Learn the basics well enough to predict behaviour: probabilistic generation, finite context window, no memory between sessions by default, plausible-sounding errors, and output variability. Prefer options that add grounding (retrieval of real data) and human verification over options that assume reliability.
Accepting an AI-generated artefact - a retrospective summary, a facilitation plan, a coaching script - as the outcome itself, rather than as an input the Scrum Master and team still inspect and reshape.
Efficiency framing ('save time on ceremonies') tempts candidates toward answers that keep the artefact and drop the human sense-making that gives the event its purpose.
What to do instead
For any AI-assisted event, pick the option where AI compresses preparation and drafting while humans keep inspection, adaptation and the conversation. If an option removes the team's own reflection or decision, it is wrong regardless of how efficient it sounds.
Not making it transparent that AI was used to produce artefacts the team relies on - e.g. not disclosing that backlog items, estimates or stakeholder updates were AI-drafted, or hiding the tool's involvement to avoid scrutiny.
Candidates focus on the output's quality and forget that transparency is a Scrum pillar; they treat AI use as a private productivity habit rather than something that affects what others inspect.
What to do instead
Choose options that make AI involvement visible to the people inspecting the work (marking AI-drafted items, stating it in the update, agreeing team norms). Where transparency is missing from an otherwise attractive option, treat that as the defect the question is testing.
Deciding for the team - answering that the Scrum Master should mandate an AI tool, or ban one, without the team experimenting and choosing; or the mirror error of refusing to raise AI at all because 'the team decides'.
Candidates collapse 'self-managing' into 'the SM has no role', or collapse 'SM enables' into 'SM selects tools', and both produce answers that remove the team's ownership.
What to do instead
Look for the option where the SM creates the conditions: surfacing the opportunity, teaching experimentation and empiricism, removing impediments, and letting the team inspect results and decide. The SM owns the environment and the learning, not the tool choice.
Leaving the Definition of Done and quality practices unchanged while the team adopts AI-generated work - e.g. accepting AI-written code or content with no review step because 'the tool is usually right'.
Novelty and automation bias make AI output feel pre-validated, and candidates answer as if quality standards are about the tool rather than about the increment.
What to do instead
Expect answers that keep or strengthen the Done criteria as AI use increases - review, testing, provenance, human sign-off on anything shipped. Reject both extremes: no change at all, and a blanket ban on AI use.
Treating organizational AI adoption as a single big-bang rollout - choosing 'standardize one tool across all teams', 'mandate usage from next quarter', or a full process redesign before anything is tried.
Adoption questions look like programme-management questions, so candidates reach for scale, consistency and top-down control instead of the empiricism the rest of the exam rewards.
What to do instead
Prefer options that run small, time-boxed experiments with explicit measures, inspect the results, and expand or stop based on evidence. Use Scrum itself - transparency, inspection, adaptation - as the adoption approach.
Leaving systemic AI impediments unaddressed - no data-handling policy, no security or procurement path, no licensing, no training budget - while coaching the team to 'just experiment'.
Candidates stay inside the team boundary because that is where the Scrum Master usually operates, and miss that the scenario is testing whether they will engage the organization.
What to do instead
Pick the option where the SM makes the impediment visible to the people who can remove it, and works with leadership on policy, access and capability. Enabling the organization is part of the accountability, not outside it.
Putting confidential material into a public AI tool because it improves output quality - pasting customer data, unreleased product details or internal strategy into a general-purpose chatbot without classification or an agreed policy.
Candidates optimise for the visible quality gain and treat data handling as an IT concern, so confidentiality conditions in the stem get discounted as background detail.
What to do instead
Before choosing, check the option against data classification and the organization's usage policy. Correct answers preserve confidentiality, use approved or on-premise tooling where required, and involve the people accountable for data protection.
Running the ethics check as 'is it legal, and do the vendor terms allow it?' - and then letting AI inform or make decisions about people (hiring, performance, evaluation) or customer-facing outcomes without human accountability, bias checks or an explanation path.
Compliance framing is familiar and feels rigorous, so candidates stop at lawfulness and never test fairness, explainability and who answers for harm.
What to do instead
Add three questions to every AI option: who is affected, could bias or error cause harm, and which human answers for the decision. Reject options that remove human accountability for consequential decisions, even when nothing about them is illegal.