The Case for Killing Projects: Budgets, AI, Attention and Token Costs

2026-10-01

Most companies can now start projects faster than they can judge them. Four things changed between 2024 and 2026, and each of them has been measured. There is also a longer line of research on why bad projects are so hard to stop, and on what works when you try. This page collects both, with sources, so you can check the argument before you buy anything.

Did budgets grow?

Yes, and the growth went to AI. In Gartner's survey of CFOs, published in February 2026, three in four expected their technology budget to rise this year. BCG's 2026 survey of IT leaders put expected spending growth at 5.8% and found the money concentrating on one priority: AI.

People went the other way. In the same Gartner survey, planned headcount growth fell from 6% in 2025 to 2% in 2026. So there is more money for AI, and roughly the same number of people deciding what to do with it.

Why are there more projects than before?

Because starting one got cheap. An analysis, a proposal or a prototype that used to take a team a month can now take one person an afternoon. Each one still needs someone to read it and decide.

A lot of it doesn't survive the reading. Researchers from Stanford and BetterUp, writing in Harvard Business Review in 2025, called it workslop: AI-made work that looks finished but doesn't move the task forward. The people who received it spent close to two hours on each case.

Pilots show the same thing at the scale of projects. MIT's NANDA initiative reported in 2025 that about 95% of the generative AI pilots it studied had no measurable effect on profit and loss. S&P Global found that 42% of companies abandoned most of their AI initiatives in 2025, up from 17% the year before, and scrapped on average 46% of their proofs of concept before production. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027.

Those projects are going to die anyway. The question is how much they eat first.

Is management attention really the bottleneck?

Herbert Simon put it in 1971: "a wealth of information creates a poverty of attention." The numbers since have moved one way.

Gallup found that the average manager had 12.1 direct reports in 2025, up from 10.9 a year earlier and nearly 50% more than when Gallup first measured it in 2013. Manager engagement fell from 30% to 27% in 2024, which Gallup named as the main reason engagement dropped worldwide. Microsoft's 2025 Work Trend Index found knowledge workers interrupted by a meeting, email or message about every two minutes. And in 2024 Gartner predicted that through 2026, one in five organisations would use AI to flatten their structure, removing more than half of their middle-management jobs.

Fewer managers, bigger teams, more interruptions, and more things waiting for a decision.

What does a person-hour cost now?

More than the salary. In June 2026 Gartner predicted that by 2028 AI coding costs will be higher than the average developer's salary, because of token use. It found 23% of tech leaders already spending $200 to $500 a month per developer on tokens, and 6% spending more than $2,000. It also heard of a business user, not a developer, who ran up a $32,000 bill in one month.

Nvidia's CEO, Jensen Huang, went further at the company's GTC conference in March 2026. He said he'd be deeply alarmed if an engineer paid $500,000 a year didn't use at least $250,000 of tokens.

So every person on a project now costs a salary plus a token bill, and a project nobody is steering burns both. The bill is coming explains why token prices are unlikely to stay where they are.

Why are bad projects so hard to kill?

Isabelle Royer studied two failed projects, at Essilor and at Lafarge, for Harvard Business Review in 2003 (Why Bad Projects Are So Hard to Kill). The trouble wasn't missing data. It was what she called collective belief: a conviction shared by most of the decision-makers, based on feeling rather than evidence, that the project would succeed in the end. Every bad sign looked like a temporary setback, and commitment went up instead of down. Barry Staw had described the same pattern in 1976 as escalation of commitment: people put more money into a losing course of action when they were the ones who chose it.

Then there is how it looks. In Atlassian's survey of 8,000 office workers, 44% started 2026 carrying zombie projects, and more than one in three said they were afraid of how it would look to pull the plug.

And budgets resist change. McKinsey found that across industries, a business unit's budget correlated 0.92 with the year before. For a third of companies it was 0.99: the same money to the same places, every year. Companies that moved more of their money around delivered higher returns to shareholders.

What works when you need to stop projects?

Give someone the job. Royer's answer was an exit champion, a person whose role is to make the case for stopping. McKinsey described a food company that made "project killer" a full-time job: in three years it cut its portfolio from more than 560 projects to just over 200, and the effect on profitability was, in McKinsey's words, overwhelmingly positive.

Decide the kill point before you start. Annie Duke, in Quit (2022), recommends kill criteria made of a state and a date: if we haven't reached this point by this date, we stop. She also recommends a quitting coach, someone outside the decision who will tell you when it's time.

Reward the people who stop. At X, Alphabet's moonshot factory, Astro Teller has described giving bonuses and recognition to teams that kill their own projects. His explanation is that teams kill their ideas as soon as the evidence is on the table, because they are rewarded for it.

The Inquest is built from these pieces: an outsider in the exit champion's job, a kill date for every project that survives, and decisions made in a room with the people who can make them.