11:58pm. You just inherited Project Meridian.
The previous PM left this morning. Meridian — a customer-portal rebuild with a fixed launch commitment to the client — is now yours. Your org has an AI co-pilot enabled on the project workspace: it scores risks, drafts schedules, and proposes estimates, all in fluent, confident, well-formatted prose. The first message is already waiting.
It reads like the work is done. That feeling — fluent, decisive, pre-formatted — is exactly the trap. The 2026 PMP exam is heavily situational and weighted toward agile and hybrid work: it grades whether you take the most appropriate first action under ambiguity, not whether you can recite a definition. The same discipline is what stops you rubber-stamping a plausible-but-wrong co-pilot recommendation.
Track 1
Graded the PMI way. Every round rewards the action a calibrated PM would take first — analyze before acting, delegate without abdicating, match the risk strategy to the threat, pick predictive vs. adaptive to fit the situation.
Track 2
Tests your AI judgment. Each recommendation is yours to keep, interrogate, or override. Get it right and trust climbs; rubber-stamp a flawed one and the meter — and your schedule — pay for it later.
One project, carried forward. The trust meter, the schedule, and the risk register persist from round to round. A decision in Round 1 changes what Round 2 looks like. Nothing resets between rounds — because real projects don’t.
Calibrate the reflex: keep, interrogate, or override?
Before the rounds begin, calibrate the muscle this whole lab trains. The co-pilot says the remaining build is 9 days, 92% confidence. Don’t accept it and don’t reject it — slide your own gut estimate of how many days it really takes, then reveal the assumption the AI quietly baked in.
Keep
The recommendation is low-stakes, reversible, and inside the AI’s competence. Trusting it is efficient, not lazy.
Interrogate
Plausible but consequential. Ask what it assumed, what it couldn’t see, and re-run with the missing input before you commit.
Override
The call needs human judgment — stakeholders, ethics, scope, or a fact the model got wrong. You own it.
Round 1 — Kickoff staffing
The co-pilot has auto-assigned Dana, your strongest engineer, to the entire critical path — and flagged that Sam, a capable mid-level engineer, is “under-utilized.” This is a People-domain call. What is your first move? Pick one, watch the consequence ripple into team-trust state, then commit it forward.
Round 2 — Triage the AI’s risk register
The co-pilot posted four threats with pre-filled response strategies. Sort each into the correct threat strategy — Avoid, Transfer, Mitigate, or Accept. Two of the AI’s pre-scores are miscalibrated; the carried residual-risk meter only stays contained if you override the bad calls. Click an item to cycle it through the bins, then check.
Round 3 — the estimate, then the fork
1. Interrogate the “too-good” estimate
Now interrogate that “9 days, 92% confidence” estimate properly. Below is the co-pilot’s forecast paragraph. Click every flawed assumption — the things that make a confident number untrustworthy — then check. The more flaws you catch, the more trust you earn for catching what the badge was hiding.
Click each phrase that signals a flawed estimating assumption, then press Check.
2. The fork — the sponsor wants the AI’s date
The sponsor saw the co-pilot’s “9 days, 92%” banner and wants it announced to the client at the 9 a.m. standup — today. You now know that number ignored a holiday and carried no reserve. This is the pivot of the project: your choice here opens a genuinely different Round 4. Pick your path, see where it leads, then commit it.
Round 4 — method, ownership, recovery
1. Methodology under pressure
Round 4 opens with the situation your fork created. (Make your decision-fork choice in Step 4 to set this.)
Mid-project, the client requests a meaningful change to the portal’s dashboard. Meridian is being run as a hybrid with an adaptive delivery team. The co-pilot offers two ways to handle the change. Which response is stronger for this situation — and which would the exam mark correct?
2. Which decisions are the AI’s to make?
Keep / interrogate / override isn’t a vibe — it maps to ownership. Some calls the co-pilot can decide, some it can only assist on while you act, and some are human-only. Click a decision on the left, then the ownership level it belongs to on the right. Link all four, then check.
3. Order the recovery
A status review shows Meridian has slipped. The co-pilot offers a tempting one-click fix. Arrange the recovery steps into the order a calibrated PM runs them. The exam’s favourite trap lives here: acting before analyzing. Use the arrows to reorder, then check.
Brief the sponsor, then debrief yourself
1. Build your AI-augmented status report
Time to brief the sponsor on Meridian. Assemble the status report by toggling components on. A credibility meter rewards transparency about what the AI generated versus what you own and verified — and it drops for components that hide the AI’s hand or over-claim confidence. Build a report you could defend in the room.
2. Your decision trail
Here is the trail you left through Project Meridian — the carried state from every round. Your trust meter moved both ways depending on whether you kept, interrogated, or overrode. Write the one judgment principle you’d carry into a real project, then reveal a model PM’s reasoning.
How a calibrated PM read Meridian
“I treated every co-pilot recommendation as a draft, not a decision. I kept the cheap, reversible, in-competence calls; I interrogated the confident estimate and the risk scores — that’s where the single-point ‘9 days, 92%’ and the mis-rated vendor threat were hiding; and I overrode the calls that belonged to me: staffing as a coaching opportunity, the date conversation with the sponsor, and anything that touched the client. I escalated with data, never emotion. And I never let the AI own a people, scope, or disclosure decision.”
Notice the pattern: the trust meter rose when you interrogated and overrode well, and fell when you rubber-stamped a confident output or abdicated a call you owned. That is the exact muscle the exam grades — the most appropriate first action under ambiguity — and the exact muscle that keeps a plausible-but-wrong AI recommendation from driving your project.
3. The override cheat-sheet
One screen to keep. Everything Project Meridian taught, at a glance.
The keep / interrogate / override rule
1Threat strategies
- Avoid — eliminate the threat or its cause
- Transfer — shift impact + ownership (insurance, contract)
- Mitigate — reduce probability or impact
- Accept — low enough to live with; set a reserve
2Delegate vs. abdicate
- Delegation = real ownership + a support cadence + a checkpoint
- Abdication = handing it off and walking away
- Servant leadership grows the person, not just the task
3Estimating discipline
- A single-point number is not a commitment
- Give a range + a contingency reserve
- Check the calendar (holidays) and the velocity’s source
4Method to situation
- Predictive — stable scope, formal change control
- Adaptive — changes reprioritized into the backlog
- Don’t default to predictive in an agile context
5Ownership ladder
- AI-decides — trivial, reversible
- AI-assists / human-acts — forecasts, signals
- Human-only — scope, ethics, client disclosure
6The traps
- Acting before analyzing (one-click rebaseline)
- Automation bias — rubber-stamping confident output
- Anchoring on the AI’s number / confidence badge
You ran Meridian the exam way.
You took the most appropriate first action under ambiguity, matched each risk strategy to its threat, delegated without abdicating, picked the method that fit the situation, and decided round by round whether to keep, interrogate, or override the co-pilot. That disciplined judgment is what the 2026 PMP grades — and what keeps a plausible-but-wrong AI recommendation from running your project.