Rolling it out to a PMO: one programme, one artifact, six weeks, one volunteer

People reach for the exciting artifact - the plan, the business case, the risk model. Pick the boring one. The weekly status report wins on three properties that nothing else in the PM canon has at once.

• Highest frequency. It happens every week on every programme, so six weeks gives you six real data points per PM instead of one. A charter pilot gives you one event and no trend.
• Clearest rubric. The six checks are objective enough that two reviewers agree: is the RAG rule written in the report, does every action have one name and one date, is there a traceable id behind each claim. You cannot score a charter that cleanly.
• Lowest blast radius. A bad drafted report is caught on Wednesday by the person who signs it on Friday. A bad drafted plan gets committed to and lives for a year. Start where the mistakes are cheap and visible.

Public reporting says the same thing from the other direction: roughly half of project professionals say AI already touches their work, about a fifth have real hands-on experience, and status reporting is consistently the first place it lands. You are not choosing the frontier here. You are choosing the place your function is already drifting, and putting a rubric under it before it becomes habit.

The single most common pilot failure is a nominated participant. Nominated people produce compliant behaviour, compliant feedback, and no information. You want someone who put their hand up, because only they will tell you the part that did not work. Pick a credible PM rather than the keenest one - if it succeeds, the sceptics have to believe the result, which the department enthusiast cannot deliver.

Pick a programme that is mid-flight and slightly messy, because a pristine one proves nothing. Northwind at week 9 - a gate slipped from 10 August to 24 August, a vendor gone quiet, quality at 96.4% against a 99% gate - is exactly the sort of week worth testing on. Do not pick a programme in crisis: nobody learns anything in a fire, and the tool gets the blame for it. And protect your volunteer explicitly, out loud, in front of others: if the pilot produces nothing, that is a finding and not a performance issue. Otherwise they will find something, and you will have bought yourself a lie.

Pick a fixed share of artifacts, review those properly, and leave the rest alone. The exact rate matters far less than the fact that it is fixed, published and unpredictable in its selection.

• Start around one in five, chosen at random, for the first quarter. Then drop it as scores stabilise. A share that never falls is a police force wearing a sampling badge.
• Review the artifact, never the person, and publish only the aggregate. The finding is "this report scored 4 of 6, the ask and the delta are missing", not "this PM is weak on reporting". The function sees "we are averaging 4.6 and the ask is our weakest check", and nobody sees whose report was whose.
• Sample the gaps blocks too. A report that scores 6 of 6 with an empty gaps block every single week is not a triumph, it is a PM who has learned to stop declaring what they could not source.

When a sampled artifact fails, the response is graded and it never starts with the PM:

• First failure of a check: fix the shared prompt or the house rules. Most failures are systemic - if four PMs all miss the ask, the template is missing a section, not the people. Second time, same PM, same check: fifteen minutes with them, on the check, not on their competence.
• A failure that reached a sponsor with a wrong number in it: that is a real incident. Reconstruct how the number got in, whether the signing human read it, and whether the pack was stale. Almost always it is a stale pack or a claim with no id, and both are fixable in the cadence.

Do not build a parallel structure. PMI's 2026 standard asks for clear structures, roles, authority and guardrails inside existing governance, with real intervention triggers - not a separate AI committee with its own reporting line and no power. In practice that is four small amendments to documents you already have.

The house rules file gains the six checks and the RAG rule. The PMO assurance routine gains the sampling rate, as one more artifact type rather than a new process. The steerco pack gains one line: the aggregate report score and the weakest check this month, which makes quality visible without naming anyone. And the escalation path gains one trigger - a wrong figure reaching a sponsor is an incident and is logged like any other, which is what turns oversight from a claim into a mechanism. Four amendments beat a new framework nobody reads, and it is the version your risk and audit colleagues will actually accept.

• Hours moved off artifact labour. Ask each pilot PM for one number in week 1: how long the weekly pack takes them, door to door. Ask again in week 6. It is self-reported and imperfect, and it is still the number your finance director will care about most. Be honest that it is an estimate, and never annualise it into a headcount saving unless you intend to make one.
• Report score against the rubric. Out of six, from your sample, averaged across the function. This is the quality half, and it is the number that stops the hours metric from turning into "fast and worse".
• Recurring findings in the gaps block. Count how often the same gap appears across three weeks. A gap that keeps returning is a structural hole in the programme - an unowned workstream, a metric nobody produces - and finding those is arguably the biggest prize in this whole track.
• Decisions reached per steerco, and days from slip detected to escalation raised. Count the decisions: if the ask is working, that number goes up and the meeting gets shorter. The escalation lag matters most and gets measured least - if your Tuesday sweep is working it shrinks, so baseline it from three past slips before you start or you will never be able to claim it.
• And two rules for whatever you report upward. Report the shape, not false precision: "the weekly pack went from about half a day to about an hour, across three PMs, over six weeks" is defensible, "41% efficiency gain" is not, and the first person who tests it will find that out in front of you. Always pair an hours number with the quality number in the same sentence, so nobody can quote half of it.

• Opt-in beats mandate, for longer than feels comfortable. Mandates produce compliance, and compliance produces artifacts that technically pass and teach you nothing. Keep it voluntary through the pilot and the first wave after it. If the thing works, the second wave asks to join.
• Show one PM's Friday afternoon coming back. Not a slide about productivity. A named colleague saying what they did instead: went home, or spent it on the vendor who had gone quiet. That story travels through a PMO faster than any number you will ever produce.
• Never measure individuals against each other, and make the failed pilot safe. The moment a leaderboard of report scores exists, the scores become the work and the gaps blocks go empty, so aggregate up, always. And say the stop condition before you start, then mean it - a pilot that can only succeed is not a pilot, and everyone in the room knows it.

And the thing to say out loud, in the first session, in these words or close to them: nobody's accountability moved. The plan is still theirs. The date they commit is still theirs. The report with their name on it is still theirs. The only thing that changed is how long the first draft takes. That sentence does more for adoption than the demo, because the fear underneath the resistance is usually not about the tool. It is about whether the thing they are accountable for is about to be decided by something they cannot argue with.