AI in project delivery.

I’ve spent 30 years recovering failing programmes. Not one of them failed because the scheduling maths was wrong.

That’s worth holding in mind as AI arrives in project management.

AI is genuinely good at a lot of project work. It synthesises huge document sets in minutes. It drafts reports. It spots patterns across data a human would miss. It will take real cost and time out of the controls function – and that is a good thing.

But here is what I keep coming back to. AI is brilliant at the part of project management that was never the problem.

Programmes do not fail because the critical path was calculated slowly, or because the status report took too long to write. They fail in the human layer – weak sponsorship, suppressed bad news, unclear accountability, optimism that no one challenges, governance that meets but does not decide.

AI does not touch that layer. It cannot make a sponsor available. It cannot make a board want to hear bad news. It cannot create accountability that was never designed in.

And there is a sharper risk.

AI amplifies whatever culture you drop it into.

An organisation that already suppresses bad news will now produce more polished, more confident, more frequent green reports – automated optimism, generated faster than anyone can question it. The tool does not fix the reporting culture. It accelerates it.

So my position is not anti-AI. I use it. It is a genuine asset to the controls function, and to anyone trying to make sense of a chaotic programme quickly. But it is a co-pilot for the mechanical layer – not a substitute for judgement in the human one.

The hard parts of delivering a major programme are still hard. They are still about people, governance, and honest conversation. AI will not save a programme that is failing for the reasons programmes actually fail.

If you are bringing AI into your project function – good. Just be clear about which problem you are solving. And do not mistake a faster report for a healthier programme.

You used to be able to spot a weak project report by reading it.

Vague language. Hedged conclusions. A risk section that said nothing. The gaps gave themselves away. AI has ended that – and owner organisations are the ones who need to notice.

Every owner organisation delivering a major programme needs an intelligent client function. It is the internal capability that holds your interests when someone else does the work – the people who read a contractor’s commercial report critically, challenge a schedule, test a risk position, and know when the answer they have been given does not add up.

You cannot outsource accountability for a programme. The intelligent client function is how you keep it.

That function has always rested on one core skill: interrogating what you are given. AI has just made that skill harder.

Here is the problem. AI-generated analysis is fluent, confident, well-structured. It reads as authoritative – whether or not it is. The traditional tells of weak work – clumsy language, visible hedging, obvious gaps – are exactly what AI smooths away. The report you receive now looks polished regardless of whether the thinking behind it is sound.

The packaging has been decoupled from the substance.

So the intelligent client needs a new literacy. Three habits in particular.

Ask what it is grounded in. Not “what does this conclude” but “what data, what assumptions, what reasoning is this built on?” Fluent output can rest on very little.

Do not be disarmed by polish. Confidence is not evidence. A confident AI-generated forecast and a confident contractor are interrogated the same way – on the substance, not the delivery.

Insist on human accountability. When an AI-assisted analysis is submitted to you, a named person should stand behind it. “The model produced it” is not an answer an intelligent client accepts.

None of this is anti-AI. AI is now part of how programme analysis gets done – on your side of the table and the contractor’s. But it raises the bar for the intelligent client. It does not remove it.

The organisations that struggle will be the ones that mistake a polished output for a sound one. The organisations that do well will keep asking the oldest question in the intelligent client’s toolkit: what is this actually based on?

The hardest part of recovering a failing programme is not fixing it. It is finding out where it actually is.

Every troubled programme tells two stories. The one in the reports – schedules, status updates, risk registers, board papers. And the true one. By the time a programme is genuinely in trouble, those two stories have pulled apart, often badly.

The recovery cannot start until you close that gap. You cannot fix a programme whose real position you do not know.

Traditionally, finding the real position is slow, painful detective work. You read everything – months of reports, change logs, correspondence, contract papers – looking for the place where the story stops adding up. The risk that quietly disappeared from the register. The percentage complete that does not reconcile with the cost. The change discussed in emails but never formally raised.

That work could take weeks. A failing programme does not have weeks. It loses money and options every day the truth stays buried.

This is where AI has genuinely changed my work. AI is exceptionally good at one specific task: surfacing inconsistency across a large, messy documentation set. Point it at six months of programme records and it will find – in hours, not weeks – where two documents disagree, where the reporting went quiet on a live problem, where the numbers do not reconcile.

That is the gap between reported and real. AI gets me to it faster than any method I have used in 30 years.

But here is the limit, and it matters.

AI can show you that reporting and reality have diverged. It cannot tell you why.

And the why is the recovery. The gap did not open by accident. It opened because of pressure, or optimism, or a culture that could not absorb bad news. AI surfaces the symptom at speed. Diagnosing the cause, rebuilding the programme, restoring trust – that is human work, and always will be.

One more thing worth saying. In an earlier post I noted that AI can automate optimism – make a weak reporting culture produce more polished, more confident green reports. The same capability, pointed the other way, cuts straight through them.

If AI can expose the gap between reported and real in hours during a recovery, it can do the same during routine assurance – long before a recovery is ever needed.

The best recovery is the one that was never required. Used well, AI makes more of those possible.