What AI changes about delivery: the labour moves, the accountability does not

The instinct in most PMOs is to point AI at the visible pain: the status report is late, so automate the status report. That is the right pile but usually the wrong first step, because the status report is the artifact with the most political weight in your function. Get it wrong once in front of a sponsor and the whole idea is dead for a year.

The better read is to sort the labour pile by two things only: how many hours it eats across the whole function, and how much damage a bad first draft can do. Start where the hours are high and the blast radius is low. Meeting notes into actions. Reformatting the same plan for three audiences. Collating four tools into one view. Those are unglamorous, they are also where a delivery function quietly loses a day a week per PM.

When your PMs argue about whether something belongs in the labour pile, these four questions settle it faster than a framework:

• Does it produce a document, or a decision? Documents move. Decisions do not.
• Is the input already written down somewhere? If the raw material lives in tickets, threads and past reports, it is collation. If it lives in someone's head or in a corridor conversation, it is judgment.
• Would a competent stranger produce roughly the same thing? If yes, it is labour. If the answer depends on knowing that Elena V. hates adjectives and Dan R. always under-promises, it is not.
• What happens if the first draft is wrong? A wrong draft of meeting notes costs five minutes. A wrong committed date costs a quarter.

Then give them the boundary in one sentence: "AI drafts the artifact, the team supplies the facts, a named person signs it." If your PMs can repeat that back to you, most of the adoption arguments in the next six months resolve themselves without you in the room.

Notice what is missing from that list: nothing decides anything. No date is committed, no message is sent, no risk is closed. That is not caution for its own sake - it is the same boundary PMI's standard draws, and it is the boundary that keeps the first three uses safe enough to actually adopt.

Public reporting on AI in project management points at a squeeze that most heads of delivery will recognise from their own week. Roughly half of respondents say AI already has an impact on how projects get managed. But only about a fifth of project managers claim good practical experience with it, and about half claim little or none.

Read the shape of that, not the decimals: high pressure from above, thin practical skill below, and nothing written down in the middle. That middle is you.

• Shadow adoption. With no policy, the keen PMs use it anyway, on whatever account they have, with whatever data is on their screen. You find out during an audit.
• Uneven quality. Two PMs, same programme, wildly different artifacts. One pastes raw output, one rewrites everything. Neither can tell you which is which, because there is no rubric.
• The credibility tax. One fluent, empty status report at a steerco and your sponsor stops trusting the whole reporting line, including the reports that were fine.
• Wasted licences. The most common outcome of a tool-first rollout is a per-seat spend with a usage chart that dies after week three.

The standard sets out eight equally weighted principles. Reciting all eight in a leadership session is a good way to lose the room, so take the four that change what you do on Monday.

• Governance and compliance. AI governance belongs inside your existing governance, not beside it. The standard asks for "clear structures, roles, authority, and guardrails" - which is a description of a RACI and a policy, not a new committee.
• Data quality. A principle in its own right, and the one your PMs will feel first. A status pack drawn from a stale context pack is confidently wrong about last month's programme, which is far harder to catch than obvious nonsense.
• People and culture, and ethics. If your PMs think this is a headcount exercise, the artifacts get quietly worse and nobody tells you why. And the accountable name does not move to a vendor. Ever.

The other four - strategic value, risk, stakeholder engagement, optimization and innovation - matter, and they are where sessions a4 to a6 land. Equal weighting is the point: there is no principle you get to trade away because the quarter is tight.

This is the single most useful line in the standard for a head of delivery, and it is worth reading twice. The human is not in the loop to catch the machine's mistakes. The human is in the loop because that is where the judgment lives. Reviewing is not the point. Deciding is.

The practical test is whether your reviewer has anything to intervene with. A reviewer who can only say yes or no is a rubber stamp. A reviewer who knows Marcus L. is the single point of failure on the CDC job, who was in the room when decision D-14 descoped real-time POS to nightly batch, and who has the authority to move the M3 gate - that person is adding something the draft could never contain.

• Context beats prompting. When a draft comes back generic, the fix is almost never a cleverer instruction. It is a missing document. If your PMs are trading prompt tricks instead of loading the charter, plan and risk register, they are optimising the wrong thing - and session a2 is about what they are allowed to load.
• The dates come from the team. Ask for a plan and you will get dates. They will be evenly spaced, plausible, and entirely fictional, because nobody was asked and nothing was capacity-checked. AI drafts the structure of a plan: milestones, gate criteria, dependencies, the questions to ask. The commitments come from the people who will be held to them.
• A named human signs it. One accountable name per artifact, and that person read it. This is not a formality; it is exactly what PMI means by oversight with real triggers.