The numbers are no longer controversial. MIT's NANDA initiative studied more than 300 enterprise AI deployments and found that 95 percent of generative AI pilots deliver no measurable impact on the P&L. This against 30 to 40 billion dollars of enterprise investment. RAND puts the AI project failure rate above 80 percent, roughly twice the rate of comparable IT projects. Gartner predicts more than 40 percent of agentic AI projects will be canceled by the end of 2027, and names the causes: escalating costs, unclear business value, inadequate risk controls. S&P Global found that 42 percent of companies abandoned most of their AI initiatives in 2025, up from 17 percent the year before.
Read those causes again. Costs nobody modeled. Value nobody defined. Risks nobody owned.
None of that describes a technology problem. Every line of it describes a delivery problem. And that means the person standing closest to the wreckage is not the data scientist or the vendor. It is the program manager.
What the failure actually looks like
It rarely looks like failure. It looks like drift.
The pilot demos well. The steering committee approves scale-up. Then the integration work that nobody scoped starts consuming the schedule. The data turns out to be two years away from ready. The token bill at production volume arrives at five times the pilot estimate. The business case said "productivity" and now nobody can say which line of the P&L was supposed to move.
The status stays green a week longer than it should. Then amber. Then a recovery engagement.
I have spent fifteen years recovering software and data programs in regulated environments. The AI programs I see now fail in a way I recognize completely.
Take the AI label off for a moment
Here is the uncomfortable part for anyone selling AI-specific transformation frameworks.
Scope that entered without a decision. Success that was never defined at the deliverable level. Dependencies two teams each assumed the other owned. A sponsor who held a date publicly that the team stopped believing privately. Reporting that measured effort spent and labelled it progress.
That list describes the failing AI program in your portfolio. It also describes every failing ERP implementation, every failing data warehouse migration, and every failing digital transformation of the last twenty years.
MIT's own researchers said it plainly: the divide between the 5 percent that succeed and the 95 percent that stall is not model quality. It is how organizations integrate and learn. The failure is organizational.
Three things are genuinely different with AI, and they deserve respect. Outputs are probabilistic, so "done" is harder to define and test. Costs scale non-linearly between pilot and production, so the pilot budget tells you very little. And the vendor landscape moves quarterly, so a plan built around one tool ages in months.
Those three differences change the plan. They do not change the discipline. A program that cannot define done, cannot model its costs, and cannot name its owners will fail with or without a model attached. AI just makes it fail faster and more expensively.
Why this recovery is harder than the ones before it
Here is what the failure statistics do not capture.
Since 2022, the tech industry has cut six figures of jobs every single year. Trackers put the cumulative total since 2020 past 900,000 people. Many of those cuts were announced in the same breath as the AI investments now failing. Efficiency. Restructuring. Doing more with less.
Now put yourself in the position of a PM inside one of these programs.
The company has publicly bet on AI. Leadership has told the market this is the future. Colleagues have been walked out under banners that used the same word. And you are the person holding evidence that the flagship AI program is not going to deliver.
Raising that risk does not feel like governance. It feels like volunteering for the next list.
So the risk gets caught in week six and formally raised in week twenty. The report moves from red to amber and nobody questions it. The impossible date gets nodded through in the steering committee and quietly cursed in Slack.
Fear taxes accuracy. In a low-trust environment, every status report pays that tax, and the portfolio view the executives are reading is the sum of everyone's caution. This is why the recovery playbook alone is no longer enough.
Recovery is a technical exercise wrapped inside a human one
The technical half has not changed in twenty years, and it works.
Re-baseline against real capacity, not the capacity the original business case imagined. Define done at the deliverable level and drop percentage complete as a reporting unit. Find the scope that entered without a decision and remove it in public, by name, with the person who asked for it in the room. Assign named owners, not teams. Sequence the milestones so the hard integration work stops hiding behind the easy work.
Any competent delivery leader can run that analysis in two weeks.
The half that decides whether the recovery holds is human, and it is the half most recovery engagements skip.
Someone senior has to absorb the discomfort instead of passing it down. When the executive asks for a date the plan cannot support, the twenty minutes of being disliked has to be spent in that room, by that leader, not converted into weekend work three levels below.
The first person who reports honest red has to be visibly protected, because everyone else is watching what happens to them. One accurate report, received well, resets the truth-telling economics of an entire portfolio. One accurate report punished confirms every fear in the building.
And the people on the program need to hear something said out loud: the program drifted, and no individual failed. In a workforce that has watched four years of layoffs, that distinction is not soft. It is the precondition for getting accurate information out of anyone.
Honest status is an emotional outcome before it is a process outcome. No tool produces it. A leader does.
Call it empathy, call it steadiness, call it leadership. Operationally it is the same thing: a leader makes accuracy safe, and then keeps it safe for longer than one meeting.
Planning the next AI program
For leaders funding the next wave, the discipline is knowable in advance.
Define the P&L metric before approval, not after. If nobody can name the number that moves, the program is research, and it should be funded and governed as research.
Fund data readiness first. Gartner expects 60 percent of AI projects through 2026 to be abandoned for lack of AI-ready data. The unglamorous work is the load-bearing work.
Pick one pain point. The 5 percent that succeed in MIT's data picked one workflow and executed. The 95 percent bought a capability and went looking for uses.
Model production costs at production volume before scale-up, and put the answer in front of the CFO early, while it is still a forecast and not a surprise.
Write the kill criteria into the charter. Agreeing in advance what evidence ends the program converts a future career risk into a governance routine. It is much easier to report the number that triggers a pre-agreed decision than to be the person arguing for cancellation.
And name who is responsible for making bad news safe. Not in the values deck. In the governance structure, with a face attached.
The organizations that get this right will not be the ones with the best models. Everyone has the same models. They will be the ones where a mid-level PM can say "this is not working" in week six and still be there in week seven.
That is not a technology capability. It never was.