Key takeaways
- A story is testable when someone who was not in the conversation could write the check. Ask the model what it assumed before you accept its draft.
- Split a story by what the user gets. Left alone, a model splits by technical layer, and a layer is not a story.
- Feedback synthesis with AI works when every theme carries verbatim quotes tagged to their source, and you still read the outliers yourself.
- The Certified AI Product Manager / Product Owner credential is marked on this exact work by an assessor who did not teach your class.
A request from sales arrives as one sentence. A story has rolled through three iterations without closing. Four hundred feedback items sit in a spreadsheet nobody has read end to end. A model can help with every part of this, and it can mislead on every part in a way that looks fine until review. The difference is judgment, which is what the Certified AI Product Manager / Product Owner credential assesses.
The vague request
Ask the model to draft a story from the sentence you were given, then ask a second question before you read the draft: what did it assume about the user, the trigger and the outcome? A model does not know your users, so it fills the gaps with a plausible average. Replace each assumption with something you know or can go and ask. The test for the finished story is plain. Could a tester who was not in the room write the check without coming back to you? If not, the story is still a wish.
Splitting a story too big to move
The story that keeps rolling over is usually several stories wearing one title. Ask the model for split candidates along the lines a Product Owner would use: by workflow step, by user type, by business rule, by the data variant causing the argument. Then hold each candidate to one question. Can it be finished in an iteration and still give the user something?
Left alone, a model tends to split by technical layer instead. Build the table, then the service, then the screen. None of those can be shown to a user, and a piece that cannot be shown cannot be accepted. No AI credential we could find asks the learner to split a real story and has the result marked. The AI-PMPO brief does: one story from your own backlog, split into pieces that can each be finished in an iteration.
Acceptance criteria that survive review
Criteria survive review when nobody can misread them. The most useful thing a model does here is attack its own draft. Ask it to list the ways a developer could satisfy each criterion and still miss the point, then the edge cases the criterion is silent on. Repair what it finds, then read the result as an editor would, because a model will happily produce criteria that are precise about the wrong thing.
The rubric names this directly. The Judgment criterion passes when you changed the AI output and can say why. An unedited draft submitted as your own work lands in Not yet.
Reading hundreds of feedback items
Feed the items in batches and ask for recurring themes, each backed by verbatim quotes tagged to the item they came from. Requiring the quotes is the step that matters. A model asked for themes without evidence will produce tidy patterns no customer actually wrote.
Then the honest caveat. Models average. They surface consensus, and consensus flattens conflict. If most users like the new flow and a handful abandoned it in visible frustration, the summary reads as broadly positive, and the handful is often where the roadmap insight is. Synthesis now takes an afternoon instead of a week, which buys you the hour to read the outliers yourself. Spend it.
How the credential assesses this work
The class is one day, live online, $495, capped at 12 people, and run by our authorized training providers. Five modules:
- Where AI goes and where it must not (45 minutes)
- Prompt patterns as a working library (60 minutes)
- Weak versus strong: reading AI output like an editor (75 minutes)
- Building your evaluation set (90 minutes)
- The role brief (75 minutes)
Of 390 minutes, 300 are hands-on on your own material. After class you submit a six-section pack (cover block, prompt set, raw AI output, final version, change log, boundary note), due 14 days after the last day of class. An assessor who did not teach you, and who never sees your name or employer, reads it against the rubric published before you buy a seat (see the standard) and returns it inside 10 working days. The criteria are Judgment, Specificity, Boundary awareness and Reusability, each banded Not yet, Pass or Strong. A Not yet comes with written feedback and one free resubmission.
The credential is for life, with no renewal fee, a printed Certificate ID and a digital badge anyone can check at Verify a Credential. Pass rates are published once the first 30 submissions in a stage have been marked, and never before. To book a seat, find a class.
Sources: American School of AI catalogue and certification standards
Frequently asked questions
What do I submit for the Certified AI Product Manager / Product Owner credential?
One story from your own backlog that is too big to move, split into pieces that can each be finished in an iteration, with the acceptance criteria rewritten on two of them. It goes in the six-section pack, due 14 days after class, and comes back marked inside 10 working days.
My employer will not let me use real backlog material. Can I still be assessed?
Yes. A fallback dataset exists for this case. Work built on it is fully admissible and marked to the same rubric. You declare the dataset on the cover block of your pack.
Can I fail, and what happens if I do?
Yes. A Not yet on any criterion has not passed. You receive written feedback and one free resubmission inside 14 days of the mark. After that there is one paid re-mark, and if that also lands in Not yet the credential closes to you for six months.