I get called when programs are already on fire — over budget, past deadline, under-governed, with a steering committee that has just discovered the green status they'd been shown for two quarters was a polite fiction. The first thing I tell the sponsor is the thing they don't want to hear from someone selling AI: no model is going to rescue this. Recovery is a human act. It runs on hard conversations, owned decisions, and someone willing to put their name on an unpopular call.

What AI changes is not the rescue. It's the timing. Almost every program I've recovered was knowable as a failure long before anyone declared it one. The signals were in the system the whole time. Nobody was reading them, because reading them was a full-time job nobody had.

Failure is loud, long before it's official

A program in trouble emits signal constantly. Tickets start reopening. Estimates drift in a consistent direction. The same three dependencies get re-planned every sprint. Commit activity clusters around a few heroes while the rest of the team goes quiet. The tone of the status threads changes — more hedging, more "should be," fewer commitments. None of this is hidden. It's just spread across ten thousand data points that no human reads end to end.

This is precisely where machines have an edge. Natural-language analysis of project communications can surface shifts in team sentiment and rising churn in requirements — the soft signals that precede the hard slip. Continuous monitoring beats the periodic review by design: the review happens monthly, the program fails daily.

The review happens monthly. The program fails daily. AI lives in the gap.

The numbers on early detection are not marginal. In one widely cited case, a large health insurer used AI to catch integration issues 45 days earlier than conventional reporting, saving $4.7M and a four-month delay. Forty-five days is the difference between a course correction and a recovery engagement. It's the difference between a hard conversation and a write-off.

The honesty problem AI quietly solves

Here is the part that makes executives uncomfortable. The reason failing programs stay "green" isn't bad data. It's human incentive. Nobody at layer three wants to be the one who reports red to layer four. So yellow becomes green on the way up, and the truth only arrives when it's too big to hide.

An agent reading the system of record has no such incentive. It does not want the promotion. It does not fear the sponsor. When it reports that reopened-ticket rate has climbed for six straight weeks, that's not an accusation — it's arithmetic. That neutrality is worth more than the algorithm. It gives a leader cover to ask the hard question early, backed by a number rather than a hunch.

Where the hype gets dangerous

Now the caution, because selling AI as a crystal ball is how you get burned. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, and a lot of that is people who trusted a confident output they hadn't earned the right to trust. Three rules I hold to:

A prediction is a prompt for judgment, not a verdict. The model flags the dependency. A human decides whether it's the one that ends the program. The flag starts the conversation; it doesn't end it.

Garbage in is now automated garbage out. An early-warning system inherits the quality of your data. If your Jira hygiene is poor, your AI risk signal is poor — just faster and more confident. Fix the foundation before you trust the forecast.

Detection without the will to act is theater. The hardest part was never seeing the risk. It was being willing to act on it when acting means an awkward conversation with someone senior. AI removes the excuse of not knowing. It cannot supply the spine.

What this means for a sponsor
  • Use AI to watch the soft signals — sentiment, ticket churn, re-planning frequency — not just the burn-down chart.
  • Trust the machine's neutrality, not its omniscience: it removes the incentive to hide, not the need to judge.
  • Forty-five days of early warning is the difference between a correction and a rescue. Buy the days.
  • Pair detection with a leader who will act on red. Without that, you've bought a smoke alarm and disconnected it.

The honest pitch

So no — I won't tell you AI saves failing programs. I'll tell you it shortens the distance between "something's wrong" and "we're doing something about it" from months to days. The saving is still done by people. AI just makes sure the people find out while there's still a program left to save.