On Gartner's own forecast, your company will abandon roughly six out of every ten AI projects it approves this year, because the firm expects organizations to scrap 60% of AI projects through 2026 wherever they aren't supported by AI-ready data and a sound integration infrastructure.

The companion number is harder to swallow, since McKinsey's most recent global survey on the state of AI shows that while nearly nine in ten organizations now use AI regularly, more than 80% see no measurable impact on enterprise-level EBIT, which means plenty of pilots and plenty of invoices but very little reaching the P&L.

Faced with that gap, the instinctive boardroom response is to ask for more use cases, more models and more budget for the next vendor promising magic, and that is precisely the move that deepens the problem rather than solving it.

The number that should rebalance your budget

A Gartner figure published in April 2026 changes the conversation, because organizations reporting successful AI initiatives invest up to four times more, measured as a share of revenue, in what Gartner calls foundations: data quality, governance, AI-ready people, and change management. That multiple doesn't buy more models; it concentrates on the ground those models stand on.

The same study, run across 353 data, analytics and AI leaders between November and December 2025, surfaces another line worth reading twice, since only 39% of technology leaders are confident their organization's current AI investment will have a positive impact on financial performance, which means six in ten already doubt it from inside the house.

The reason is that money tends to flow upward, toward the thin visible layer where model licences, copilots and the showy board pilot live, while the layer that actually decides whether any of that pays back sits underneath, wide and unglamorous, and usually ends up starved.

Why this is a governance decision, not a technology one

The reclassification that matters starts here, because when an AI project fails the committee tends to file it as a technical problem (the model wasn't good enough, the vendor underdelivered, engineering talent was thin) when the real cause almost never sits in that territory.

Gartner measured exactly this in a survey of 248 data management leaders, where 63% of organizations either did not have, or were unsure whether they had, the right data management practices for AI. A project rarely falls over because of the model; it falls over because nobody defined, governed or made accessible the data feeding it, and because nobody with business authority owned the outcome.

That turns it into a leadership-team problem, and where regulatory or capital exposure is in play, a board one, since deciding who owns the data, what gets measured and who can stop a project and on what criteria is not something a vendor solves, but rather the governance you put in place before you sign.

"If the model is the only thing your company is willing to fund, you already know why six in ten projects never reach the P&L."

The good news is that all of this is actionable without a hyperscaler budget, because it requires no proprietary model and no data centre, only a decision filter and the discipline to apply it at every approval.

The filter: six questions before you approve a euro

What follows is something a CEO or CFO can take into the next committee and turn into an approval standard, so that no AI project gets funded until the sponsor has answered these six questions in writing, and any project that fails one of them is reworked or dropped instead of approved.

1. Is there a business owner with a P&L, not an IT owner? When the sponsor is the CIO or an innovation lead with no profit-and-loss, the project is born an orphan, because the owner should be whoever wins or loses money on the affected process and who puts their name on the outcome.

2. Is the data feeding this ready, or are we assuming it is? "Ready" means defined, governed, accessible and quality-measured rather than merely "we have it somewhere," so if the honest answer is "we assume so," the project's first deliverable becomes preparing the data well before anyone touches the model.

3. Is it tied to a business metric with a baseline measured today? Any magnitude the business understands will do, whether revenue, unit cost, cash cycle or process time, but if you can't state this week's starting number you won't be able to prove the return a year out, and the project will end up coasting on "it seems to be going well."

4. Who can kill this project, when, and on what criterion? You need a named person, a specific review date (90 days is reasonable for the first checkpoint) and a specific threshold below which it is cancelled without debate, because until someone holds that authority every pilot tends to become permanent.

5. Have we funded the foundation layer at least to the level of the model? This covers data, governance and change management, and if 90% of the budget goes to model licences and consulting while only 10% sustains the base, you've bought a race car for an unpaved road, which is the opposite of what the Gartner data suggests when it shows the winners tilting the balance toward the foundations.

6. Does it clear the European regulatory filter? The EU AI Act's obligations for high-risk systems apply from 2 August 2026, with some uncertainty around a possible partial delay via the proposed "Digital Omnibus" package that a prudent committee should not bank on, and since penalties run up to 35 million euros or 7% of worldwide turnover, it pays to classify the use case before you build it rather than after you deploy it.

These six questions take barely a quarter of an hour of committee time, and the cost of skipping them is exactly the 60% of abandoned projects Gartner was talking about.

What to do with the pilots already in flight

Almost every mid-market company already carries a portfolio of pilots nobody approved through a filter like this, and ordering it doesn't require a quarter-long audit: running questions 2 and 4 against them is enough, since those two expose the truth fastest.

Question 2 tells you which ones are built on sand, while question 4 flags which ones have gone months with nobody holding the authority to stop them, and the intersection of those two groups forms your immediate cancellation list, almost always longer than the team would like to admit.

One caution before you swing the axe, because a pilot failing question 2 isn't always worthless when the idea is sound and only the data is missing. The discipline there is to rewrite it rather than merely defend it, converting "deploy the model" into "make the data ready, then revisit" with the same owner, the same kill date and a smaller cheque, so that what you cancel is the original funding decision rather than the ambition behind it.

Freeing that budget works as a reallocation rather than an accounting cut, since you move money from the layer that returns nothing toward the layer that decides whether anything will, and that rebalance is precisely what separates the 39% who are confident from the 60% who abandon.

The filter applied: a fifteen-minute case

It helps to see this on a concrete case, so picture the most typical proposal landing in a leadership committee today: a customer-service copilot that will "cut response time and take load off the support team." It reads reasonably on paper, yet the filter puts it through three tests that change the conversation.

Owner. The sponsor is the head of innovation, who doesn't hold the customer-service P&L, so the proposal fails the first question before it even starts. The way forward isn't to kill the idea but to require the customer operations director to co-sign and own the outcome as the condition for it to move.

Data. The copilot needs ticket history, the knowledge base and product data, and because nobody in the room can confirm those three sources are defined, clean and accessible, the signal is clear: the first deliverable stops being the copilot and becomes preparing those sources, with its own budget and an assigned date.

Metric and executioner. "Cut response time" doesn't count as a baseline while nobody can state this week's average, in minutes and by channel, because without that starting point there is no way to prove return. And the most uncomfortable question remains: if at 90 days the time hasn't dropped past the agreed threshold, who signs the cancellation? When the answer is "we'll see," the project walks straight into purgatory.

In barely a quarter of an hour the committee has turned a vague proposal into three concrete decisions (a co-signer with a P&L, a first budgeted data milestone and a cancellation criterion with a date), none of which the vendor could have made and all of which shift the odds that the money comes back.

Who has to own this

The filter loses all its value if the department asking for the budget is the one applying it, so it has to be held by someone with cross-functional authority, whether the CEO, the CFO or the board where regulatory or capital exposure justifies it. The reason is simple, because questions 1, 4 and 5 force you to say no to projects with internal champions, and nobody says no to their own boss as readily as an outsider would.

When a company has already burned an AI budget with nothing to show, the problem was rarely the model but the fact that nobody at the strategy table set a filter before signing the cheque, and that conversation, the one about where AI capital allocation is actually decided and who holds the authority to stop it, is exactly where we work: see our programs.

Next time an AI proposal reaches the committee, rather than asking which model it uses, ask who owns it, whether the data is ready, what metric it moves and who can kill it, because if those answers don't exist, the cheque shouldn't either.