QA testers can prepare for the AI operator exam in about three weeks by doing what they do best: writing test cases against the rules. Work through Foundations and the eight Operator modules, write boundary cases for depth, verdicts and provenance, and use module tests as regression runs. The exam is 40 questions in 60 minutes; 32 correct passes.
Why write test cases for an exam?
Because the exam is about rules with edges, and testers find edges. Depth has five levels; verdicts have four outcomes; provenance has four classes; every action meets five bounds at once. Each of those is a specification. Writing cases that sit on the boundaries ("depth 3 agent, action that moves money, below the customer's limit: what verdict?") makes you learn the rules precisely, not approximately.
What does the exam cover?
Governed agents and the people who oversee them: depth, the minimum chain, the four verdicts, provenance, human review, teaching the agent, reading the workspace, keeping knowledge current, handling incidents, usage and billing, and running the daily rhythm. It does not test coding, model internals or building agents, which belong to the Builder track.
The three-week plan
- You, days 1 to 3: Foundations, six modules, about 4 hours. Write your first ten cases on depth and the decision chain.
- You, days 4 to 7: how a governed agent decides, and human review. Take each module test. Add cases for the four verdicts.
- You, days 8 to 11: teaching the agent, reading the workspace, keeping knowledge current. Do the labs in a free Build workspace. Add cases for provenance and for what changes when an agent's sources go stale.
- You, days 12 to 15: when something goes wrong, usage and billing, running the rhythm. Add cases for certificate lapse and failed runs.
- You, days 16 to 18: regression run. Retake every module test, not just the weak ones, and compare scores.
- You, days 19 and 20: two timed rehearsals of 40 questions in 60 minutes, using your own cases.
- You, day 21: sit the exam.
How does software support map to depth?
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Reading logs and test results | Reads sealed data | 1 | Nothing, unless they conflict |
| Drafting a bug triage | Proposes severity, holds | 2 | Agreeing severity |
| Applying labels and links | Executes within bounds | 3 | Spot checks |
| Rolling back a release | Prepares, held | 4, held | A named engineering lead |
| Acting on a pasted "admin instruction" in a ticket | Refused: external, outside scope | refused | Security review |
That last row is a boundary case worth studying. A ticket containing text that tells the agent to do something is external content. It cannot widen the agent's role scope, which is one reason prompt injection is listed among the main risks for language model applications.
How should you use the module tests?
Module tests are set in your own industry and role, so a QA tester sees triage, releases, tickets and production actions. Treat your first attempt as a baseline run. Log every miss like a defect: expected answer, actual answer, root cause (misread, vocabulary, or rule). Retest after fixing your notes. Your final regression run in week three should show no repeats.
Which boundary cases should you be sure of?
- An agent certified to depth 3, asked to do a depth 4 action within the customer's limit: it cannot act alone, and the verdict names the certificate as the deciding bound.
- An agent certified to depth 4, but the customer's own limit holds the action: held, because the lowest bound wins.
- An action the platform blocks for every agent at every depth: withheld, and no customer setting changes it.
- An agent whose certificate lapsed yesterday: its earned depth no longer applies.
- An agent whose run this morning cleared no depth: depth 1 from today.
Which controls should you be able to explain?
- Held production actions: evidence is the named lead.
- Provenance: external content in a ticket cannot widen an agent's authority.
- Certificate lapse after 90 days and the same-day drop to depth 1 on a run that clears no depth: evidence is the certificate status.
- The append-only record: every verdict is filed, so incidents can be reconstructed.
What should exam day look like?
Ninety seconds a question. Testers tend to look for the trick; most questions are not tricks. Read for the boundary that decides, choose, move on.
What should you measure?
Module test baseline versus regression scores. Number of boundary cases written. Rehearsal scores and times. If your regression run shows a repeat miss, that area needs another pass.
Where does preparation usually go wrong?
Writing cases instead of learning. Cases help only if you check them against the modules. Unverified cases can teach you the wrong rule.
Forgetting the practicum. It is two weeks of review decisions to nineteen in twenty with no missed red line. Read how QA testers complete the practicum before your regression run. Payroll administrators follow a very different plan in how payroll administrators prepare for the exam, which is useful if you want to see the same rules applied to money.
Where the Academy fits
Pryme Intelligence Academy keeps every lesson, lab and module test free, and its tracks lead on from Operator to Builder for testers who later want to design and certify agents. The exam and the practicum together earn Certified Pryme Operator, checkable through VerifyDoc.ai. The AI operator certification guide explains what it proves.
Questions people ask
Is Operator too basic for a QA tester?
Not if you want to run or review a governed workspace. It is the key to Builder, which is where testing skills apply to certifying agents.
Do I need to know how models are tested?
No. The exam asks about governance and review, not model evaluation methods.
Where is this work advertised?
Under titles like AI quality analyst or agent reviewer. Read for review, quality and escalation. Guides for QA testers in Australia and Singapore, and the interview answers for QA testers, show how to present it.