September 7, 2026 · 8 min read
TL;DR: Eighty-three questions in this bank put an AI or machine learning tool inside the scenario, across People, Process and Business Environment. None of them test how the model works. Options that ask the supplier to fix the tool are keyed twice against 21 as distractors; options that retrain it or feed it more data are keyed once against 13. The keyed answer is almost always the thing a model cannot do for you: name who is accountable, define what the output has to satisfy, or test it against evidence the supplier never had.
The July 2026 exam content outline put artificial intelligence into a project manager's working environment, and a lot of candidates read that as a warning to go and learn how models are trained. Then they sit a question about a blood service buying a forecasting tool and find that nothing in the four options mentions training data at all.
We measured the whole slice rather than guessing at it. Eighty-three question stems in this bank name AI or machine learning: 34 in People, 31 in Business Environment, 18 in Process. They cover more than 19 different exam content outline tasks, from managing conflict to planning procurement. Across all of them, the technical option is the one you are being offered so that you will take it.
Almost always a tool that is working correctly and producing a problem anyway.
A ceramics manufacturer runs a machine learning scheduler that assigns each task to whoever is fastest at it. It does that faithfully. The result is that the same few people draw every glaze trial and never touch design work, and the team has split into resentful halves. The options offer sending the aggrieved members to argue with the scheduler's owner, reassigning the next cycle by hand, and publishing the scheduling logic so people can see it is not personal.
The keyed answer adds spread of experience to what the scheduler has to satisfy, not speed alone. Nothing is broken. The objective is incomplete, and a person handing out work would have noticed within a fortnight. That is the shape most of these questions take. The model optimised what it was told to optimise, and nobody checked whether that was the whole job.
Four of them, and they are worth learning by name because they recur across all three domains.
| Option shape | Keyed | Distractor | Why it fails |
|---|---|---|---|
| Ask or require the supplier to fix it | 2 | 21 | The fix lands after the decision you still have to take |
| Retrain the model, or feed it more data | 1 | 13 | Answers accuracy when the problem is the objective |
| Document it, or add a caveat | 0 | 9 | Records the defect instead of removing it |
| Remove or halt the tool | 4 | 7 | Sometimes right, but often avoids a live decision |
The supplier option is the one that catches good candidates, because in a procurement context leaning on the vendor feels like using the contract. A grocery wholesaler's project board has been approving stage transitions on the strength of a monthly readiness score, and the supplier admits it cannot say what produced any particular month's figure. One option asks the supplier to warrant that the score is accurate. It reads like commercial discipline.
It is wrong. A warranty allocates blame after a bad go-live; it does not give the board anything to decide with. The keyed option sets out what the board must be satisfied of before approving go-live, and what evidence would show it. That turns an approval the board was ratifying back into a decision the board is taking.
Common trap: "The tool is inaccurate, so improve the tool." In this bank's AI scenarios, options that retrain the model or gather more data are keyed once against 13 as distractors. The reason is that most of these stems do not describe an inaccurate tool. They describe an accurate tool pointed at a slightly wrong target, or an accurate tool whose accuracy was measured on the supplier's own data. A housing association's damp-risk ranking is the clearest case: it was built from six years of reported repairs, so a home nobody has ever reported stays at the bottom forever. More data of the same kind deepens the blind spot. The keyed answer treats the 3,100 never-visited homes as a new risk the programme itself created and gives them a share of each quarter's surveys.
Three moves, and every one of them is something a model cannot do on your behalf.
Test it against evidence the supplier never had. A regional blood service is buying a donor-attendance forecasting tool whose proposal claims 94 per cent accuracy, measured on the supplier's own evaluation data. The keyed answer holds back two years of the service's own session records and makes acceptance and final payment depend on performance against them. Asking the supplier to explain how it calculated the figure, or writing 94 per cent into the contract with liquidated damages attached, both leave the number resting on data the supplier chose.
Name who is accountable for the output. This is the strongest positive signal in the whole set. Options that state who owns a decision, who answers for an outcome, or where accountability sits are keyed 8 times against 1 across the bank. In an audiology provider rolling out a fitting-support model, the keyed answer rewrites the clinician role's responsibilities with the line manager so they say what the clinician now owns.
Reduce the surface the tool is allowed to touch. A parking-enforcement service has an AI assistant drafting 4,300 individual notices about a single change to permit hours, and two of the first twenty drafts tell a holder their permit is unaffected when it is not. The tempting options tighten the assistant's instructions or restrict it to certain permit types. The keyed answer sends all 4,300 holders one approved notice in the same words. There is one fact to communicate, and 4,300 personalised versions are 4,300 chances to get it wrong.
Read those three together and the pattern is not really about AI. It is the same accountability question the exam has always asked, wearing new clothes. That is worth saying plainly, because it turns an intimidating new topic into a set of questions you already know how to answer.
Does the 2026 PMP exam actually test artificial intelligence? Not as a technical subject. Eighty-three questions in this bank put an AI or machine learning tool inside the scenario, spread across all three domains and more than 19 exam content outline tasks, but none of them ask how a model works. The tool is the situation, and the question is still a project management question about accountability, evidence and scope.
Is "ask the supplier to fix it" ever the right answer on an AI question? Rarely. Across the AI scenarios in this bank, options that ask, require or instruct the supplier to do something are keyed twice against 21 as distractors. It fails because the supplier's fix arrives after the decision the project manager still has to take, and because most of these scenarios describe the tool working exactly as specified.
What about retraining the model or getting more data? Keyed once against 13 distractors. It is the most technically literate wrong answer in the set. The scenarios where it appears usually describe a tool whose objective is wrong rather than whose accuracy is low, and more data aimed at the wrong objective produces the same result faster.
Should I ever pick the option that switches the AI tool off? Occasionally, but treat it as the exception. Removing or halting the tool is keyed 4 times against 7 as a distractor in this bank's AI scenarios. It keys when the tool is producing something the organization cannot stand behind at all, and fails when it is a way of avoiding a decision that still needs making.
PMP Practice's 2,141 questions are re-certified against PMBOK 8 and the July 2026 exam content outline, with every wrong answer explained rather than just marked wrong, and the AI scenarios above are drawn straight from the bank. Start the free 20-question sample — no card, no signup required to try it.
Does the 2026 PMP exam actually test artificial intelligence?
Not as a technical subject. Eighty-three questions in this bank put an AI or machine learning tool inside the scenario, spread across all three domains and more than 19 exam content outline tasks, but none of them ask how a model works. The tool is the situation, and the question is still a project management question about accountability, evidence and scope.
Is 'ask the supplier to fix it' ever the right answer on an AI question?
Rarely. Across the AI scenarios in this bank, options that ask, require or instruct the supplier to do something are keyed twice against 21 as distractors. It fails because the supplier's fix arrives after the decision the project manager still has to take, and because most of these scenarios describe the tool working exactly as specified.
What about retraining the model or getting more data?
Keyed once against 13 distractors. It is the most technically literate wrong answer in the set. The scenarios where it appears usually describe a tool whose objective is wrong rather than whose accuracy is low, and more data aimed at the wrong objective produces the same result faster.
Should I ever pick the option that switches the AI tool off?
Occasionally, but treat it as the exception. Removing or halting the tool is keyed 4 times against 7 as a distractor in this bank's AI scenarios. It keys when the tool is producing something the organization cannot stand behind at all, and fails when it is a way of avoiding a decision that still needs making.