Abandoned airplane in a vacant lot, metaphor for killing AI projects that no one closes

Publications

Killing AI Projects on Time: The Cemetery No CIO Wants to Dig

Hector Pincheira

Hector Pincheira

CIO | CTO
Digital Strategy & Technology Innovation

items

A few weeks ago, I reviewed the AI initiative dashboard of a CIO I’ve known for years. Sixteen active projects. Color-coded dashboards, progress bars, traffic lights almost all in green. I asked him how many had generated a measurable business result. He was silent for a long moment and then said, «two, maybe three.» The other thirteen were still there. Consuming licenses, hours from a scarce data team, and, above all, the attention of people who could be doing something that actually matters.

No one had killed them. No one wanted to be the one to do it. That discomfort is repeated in hundreds of executive committees. And it explains why the following is almost never discussed aloud: kill AI projects Being on time is not a sign of failure. It is, arguably, the most underrated leadership skill of this decade.

The board that nobody wants to clear

When is it worth killing an AI project? Before the political cost of shutting it down exceeds the cost of keeping it alive. Gartner projects that more than 40% of agentic AI projects will be cancelled before the end of 2027. The discipline of finishing, not the discipline of starting, is a CIO’s real competitive advantage.

That scene repeats itself in executive committees across half of Chile and half the world, with different names and the exact same pattern: AI portfolios that only grow. Pilots are approved with enthusiasm and presented to the board as evidence of innovation. And there they stay, indefinitely. It is a limbo: they do not fail fast enough to justify shutting them down, nor do they work well enough to scale them.

The problem is not the number of pilots. It is that almost no CIO has a real mechanism to end them, and that absence is disguised as strategic patience when it is actually avoidance. A portfolio that only adds and never subtracts is, sooner or later, a portfolio that no one fully understands or controls anymore.

To pitch is free; to kill costs your job

Here is the uncomfortable idea: the CIO who measures their AI leadership by how many initiatives they have running is measuring the wrong variable. The one the board should care about is not how many AI projects exist in the organization, but how many were killed this quarter and why.

Launching a pilot is the easy part of the job. Any manager with a budget and an enthusiastic vendor can announce one at the next committee meeting. Killing AI projects on time is the decision that truly requires leadership. It involves publicly admitting that something you yourself pushed for is not working. And doing so before the sunk cost becomes an excuse to keep funding it.

I have been in that room. I know what it feels like to propose ending something you brought to the committee eight months ago. It feels like losing, even if it's the most responsible thing you can do.

Dashboard with 16 active AI initiatives and only 3 with measurable results, the scene that forces the killing of AI projects

The number your directory is not asking for

Recent evidence confirms that this is no longer a theoretical problem. The 2026 State of AI Cost Governance Report Mavvrik surveyed 396 business organizations between April and May of this year. One in four delayed or canceled an AI initiative due to unanticipated costs. Nearly half reported spending surprises that escalated to the board of directors. And a third had to implement emergency budget freezes, according to CIO Dive coverage.

Gartner projects that more than 40% of agentic AI projects will be canceled before the end of 2027. The reason is not technical: skyrocketing costs, unclear business value, and insufficient risk controls. The survey State of the CIO 2026 CIO.com surveyed 662 IT leaders and 249 business users. Only 19% said their AI initiatives had met or exceeded their goals.

As I mentioned a few weeks ago when reviewing the July tech radar, AI governance continues to lag behind adoption. And a portfolio without closure criteria is exactly that: governance that’s behind the curve—again.

With that scenario, one would expect cancellation to be the norm. It is not. What is rare is not that so many AI projects fail: we already knew that. What is rare is how many are still alive after failing, taking up budget that should be funding the next bet.

The Royal Room: Why No One is Signing the Cancellation

The reason is not technical; it is political. Canceling an AI project in a committee almost always means exposing someone. The sponsor who enthusiastically defended it. The vendor that promised results that never arrived. Or the CIO themselves who prioritized it over other initiatives. In that room, killing the project is seen as admitting an error in judgment. Very few executives want to be remembered for that. Even though the alternative—letting it limp along out of inertia—is worse for the company.

Thus are «zombie platforms» born: AI systems that nobody really uses and that nobody thoroughly audits. Nor does anyone dare to turn them off. Turning them off requires someone to raise their hand and say, «this was a bet that didn't pay off.» It's the same logic I described a few months ago when talking about the shell companysystems that exist on paper, in the budget, and in the presentation to the board of directors, but not in actual operations.

And here it connects with something I already brought up when discussing who signs when the AI decidesIf no one has the authority —nor the stomach— to sign off on the launch with real accountability, no one will have it to sign off on the closure either. They are the same empty chair, viewed from the other end of the project. The same void explains something I already raised with the last knocker-upper. So many promises of management automation die in silence, instead of dying in a committee meeting minutes.

What the pharmaceutical industry knows about finishing on time

There is an industry that solved this problem decades ago. We almost never look at it as a management reference: the pharmaceutical industry. Norstella It tracks drug development globally. According to its data, the probability that a compound entering Phase I will ultimately be approved fell to 6.7%, based on figures from 2014–2023. Most compounds fail in Phase II, where only 28% manage to advance.

No serious pharmaceutical company treats those deaths as team failures. It treats them as the system working correctly. The budget, the milestones, and the internal culture are designed from day one assuming that almost everything tested will die along the way. The discipline to kill it quickly, before committing Phase III capital, is precisely what funds the next compound that does work.

A CIO, CTO, and CAIO with years of experience summarized it like this in this year's CIO.com survey: «we kill about a third of what we start, and that's healthy». Their model runs on staged funding tied to outcome milestones—not deliverables—with reviews at 90, 180, and 270 days. If a project fails two consecutive reviews, it is shut down. No special committee, no drama, no one losing their position for proposing it. The shutdown is built into the process design; it does not depend on an executive's individual courage on a random Tuesday.

That is the exact difference between managing an AI portfolio and collecting pilots: in one case, killing AI projects is a scheduled event with no political cost; in the other, it is a personal battle that almost nobody wants to fight.

Staged funding model with reviews at 90, 180, and 270 days to kill AI projects with no political cost

When killing fast is the mistake

And here I have to concede something to the CIO who resists this logic, because they are right on a point that is usually overlooked: not everything that fails to show an ROI in 90 days deserves to die. There is AI investment that is, by nature, of a long horizon: a data platform, a governance model, the cleanup of an ERP poorly implemented ten years ago. None of those will show a direct financial return at a quarterly checkpoint. Their value is enabling everything that comes after, not generating a number of their own.

Applying the same fail-fast trigger to those structural bets is a different, but equally costly, mistake: it kills exactly what the rest of the portfolio needs to function. I know companies that canceled their data governance initiative due to a «lack of visible results.» Eighteen months later, they still couldn't trust a single AI-generated report. The foundation was never finished. The kill rate as a metric can also become a dangerous fad if applied without distinguishing between a tactical pilot and an infrastructure bet.

The discipline of ending, then, is not killing quickly just for the sake of killing. It is having explicit, and different, criteria for each type of bet. Defined before starting. Not invented halfway through, when someone needs a political excuse to save (or sink) an uncomfortable project.

The Movement: How to Kill AI Projects Without Losing Your Job

The concrete move is this: before approving any AI initiative, your committee should require—along with the business case—the answer to a question that is almost never asked: under what exact condition does this get shut down, and who has the authority to trigger it without asking for special permission? If no one can answer that question in the same meeting where the budget is approved, you don't have an AI portfolio. You have a collection of bets with no expiration date.

Three figures from Gartner, CIO.com, and Mavvrik regarding the cancellation and failure of artificial intelligence initiatives

The next time you sit down to review your AI initiatives dashboard, don’t start by asking how many are currently running. Ask how many should have been shut down two months ago but are still there because no one wanted to be the one to pull the plug. That number—not the other one—is what truly measures whether you’re the one managing the portfolio, or whether the portfolio is managing you.

Most Read

QlikSense Hierarchy: parent-child chart of accounts expanded by levels
QlikSense Hierarchy: the Chart of Accounts That Doesn't Add Up
QlikSense Sum TOTAL: share of each family over the total
QlikSense Sum TOTAL: the Percentage That Changes When You Filter
SAP Business One Inventory Revaluation - featured image of the tutorial
SAP Business One Inventory Revaluation: Fix Wrongly Loaded Costs
Empty chamber with wooden benches: the counterweight that required being present to be used
The Veto You Had to Say in Person: Your Decisions With No Counterweight
SAP Business One customizations - featured image of the pre-upgrade audit tutorial
SAP Business One Customizations: Pre-Upgrade Audit
SAP Business One Customer Concentration - featured image of the tutorial
SAP Business One Customer Concentration: Measure Your Risk with SQL

RELATED ARTICLES

Empty chamber with wooden benches: the counterweight that required being present to be used
Leadership in times of AI

The Veto You Had to Say in Person: Your Decisions With No Counterweight

There was once a post whose veto counted only if you said it in the room, in the flesh, at the exact minute. Nobody repealed it: the one deciding ended up keeping it. Your committee has just repeated the manoeuvre without noticing, and that is how decisions with no counterweight are born, the kind no governance framework ever catches.

A harbor pilot advising the captain on a ship's bridge, a metaphor for the CIO's responsibility without authority
Leadership in times of AI

The Pilot on the Bridge: the Responsibility Without Authority the CIO Accepted

On a ship's bridge, the one who knows the channel best advises and the one who answers commands. In your company it runs backwards: the business buys the AI and the CIO signs the inquiry. Three 2026 data points, one uncomfortable concession and three moves to take responsibility without authority apart before the bill arrives.

Traveler looking at an airport departures board, a metaphor for AI vendor dependency
Leadership in times of AI

Someone Else's Calendar: The AI Vendor Dependency Nobody Negotiates

Your model vendor publishes, on a web page, the dates it switches versions off. Nobody in your company has that page bookmarked. AI vendor dependency is not measured in exit cost, but in how many of your decisions carry an expiry date set by a third party. What to sign beforehand, and what to take to the board.