Agentic AI Projects Canceled: A Running Ledger of Public Cases and Reasons

In June 2025, Gartner predicted that more than 40% of agentic AI projects would be canceled by the end of 2027. Sixteen months later, we could verify only 3 named, public cancellations or reversals of AI automation projects made since that prediction. Not one of them gave cost as the reason. That gap between the survey data and the public record is the most useful thing to know about agentic AI projects canceled so far.

The direct answer. Gartner predicted in June 2025 that over 40% of agentic AI projects would be canceled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. Surveys since then suggest roughly half of large companies have already slowed or narrowed agent rollouts. Named public cancellations remain rare, and the ones on record cite quality failures rather than cost.

What the Gartner agentic AI prediction actually said

The prediction is quoted far more often than it is read. The release, dated 25 June 2025, said more than 40% of agentic AI projects would be canceled by the end of 2027 "due to escalating costs, unclear business value or inadequate risk controls". Those three reasons are the frame this ledger uses to sort each case.

Gartner analyst Anushree Verma described most agentic AI projects at the time as early stage experiments or proofs of concept, mostly driven by hype and often misapplied. That framing matters. The prediction was about a population of projects that were, by Gartner's own description, mostly not in production yet.

The same release introduced a phrase that has since become standard: agent washing. Gartner said many vendors were rebranding AI assistants, robotic process automation and chatbots as agentic AI. It estimated that only about 130 of the thousands of vendors claiming agentic products were real.

The poll behind the prediction

The release drew on a January 2025 Gartner poll of 3,412 webinar attendees. Of those, 19% said their organisation had made significant investments in agentic AI, 42% conservative investments, 8% none, and 31% were waiting or unsure.

A webinar audience is self-selected. People who attend a Gartner session on agents are more interested in agents than the average buyer, so treat the investment split as directional. The 40% figure itself is an analyst prediction, not a measurement, and it should be judged by whether the public and survey record starts to match it.

The public ledger of agentic AI projects canceled so far

We set out to collect every agentic AI project cancellation disclosed in a filing, an earnings call or named press coverage. The list is short. Here it is, including two cases that predate the prediction because they set the pattern the later ones follow.

Only the last three rows fall after the 25 June 2025 prediction. Sources: NRN (McDonald's, Taco Bell), Entrepreneur citing Bloomberg (Klarna), Information Age (Commonwealth Bank), The Next Web citing Reuters (Meta). Full references at the end of this post.

How an entry qualifies

Each entry needed three things. A named organisation, a project that was ended, reversed or materially narrowed, and either a company statement or a report built on named documents or executives. Anonymous "one large bank" anecdotes from vendor blogs did not qualify, however vivid.

We also left out outages and single-feature pauses. A vendor's AI agent going down for an afternoon because its model provider had an incident is a reliability event, not a cancellation. The same goes for a feature switched off pending a security fix. Neither tells you anything about whether the project was worth doing.

Sorting each AI project failure by the reason given

Gartner gave three reasons. A fair test of the prediction is to sort each public case into them. When we did, two patterns showed up straight away.

Quality failures dominate the public record

Every case lands in the middle column. McDonald's never published a reason, but the accuracy gap was the reported context: low 80s against a 95%-plus bar. Klarna's chief executive Sebastian Siemiatkowski told Bloomberg in May 2025 that the cheaper AI service resulted in "lower quality" output, about 15 months after the company said its assistant did the work of 700 agents.

Taco Bell's chief digital and technology officer Dane Mathews put it most honestly. "Sometimes it lets me down, but sometimes it really surprises me," he said, as reported by Nation's Restaurant News in September 2025. Taco Bell kept voice AI and started deciding where not to deploy it. That is a scale-back, and it is probably the most common shape a cancellation takes.

Broken business cases and incidents

Commonwealth Bank is the cleanest example of the "unclear value" reason. The bank cut 45 roles on the basis that a voice bot had reduced call volumes. The Finance Sector Union said volumes were rising. In August 2025 the bank conceded its assessment "did not adequately consider all relevant business considerations" and called the redundancies an error, according to Information Age.

That is a business case failure, not a technology failure. The bot may have worked as designed. The numbers used to justify removing humans did not hold up when someone outside the project checked them. In our view, this is the failure mode most enterprise buyers should fear, because it survives every internal review.

Taco Bell and Meta also show the risk column. Taco Bell's system was gamed by customers in public, including a widely shared order for 18,000 cups of water. Meta's internal data, covered below, showed a 40% rise in major technical and security incidents.

Why no named cancellation blames cost

The survey data and the public record disagree, and both are probably right. They are measuring different things.

KPMG's Q2 2026 Global AI Pulse asked 2,145 senior leaders in 20 countries whether they had changed course on AI agents because expected costs began to outweigh the value generated. Of those, 24% had scaled back or narrowed a deployment and 25% had delayed or paused further rollout. Another 22% had questioned the decision without changing course. Only 7% said their organisation had reached established return on investment.

Cost failures are silent by design

A project killed on cost rarely gets a press release. The renewal is not signed, the pilot budget is not extended, and the team is moved to something else. Nobody outside the company notices because nothing customer-facing changes. The mechanics of that slow death are covered in the analysis of how sunk cost keeps AI pilots funded.

Quality failures are the opposite. They happen in front of customers, in drive-thru lanes and support queues, and they get filmed. Labour reversals involve unions, who have every reason to publicise a win. So the public ledger over-represents loud failures and under-represents expensive ones. If you read only the headlines, you would conclude that agents fail on quality. The survey data says they fail on economics first.

We think the survey is closer to the truth here. A company will happily admit its bot got an order wrong. Admitting it spent money on something whose value it never measured is a governance confession, and few boards sign off on that.

Meta's Project OT: the one true agentic AI cancellation

If you apply Gartner's own definition strictly, only one ledger entry is an agentic project. The others are voice ordering and chat assistants, which Gartner would likely call agent washing. Meta's Project OT is different in kind.

According to a Reuters investigation by Katie Paul, reported by The Next Web in August 2026, Project OT ("organization transformation") was set up in January 2026. Meta explored cutting the size of many teams by as much as 60% and handing their work to AI agents overseen by smaller groups. A first wave of about 8,000 job cuts, roughly 10% of staff, went ahead in May. On the night of 19 May, hours before that wave, Mark Zuckerberg called off the planning for a second wave in November.

What Meta's own numbers showed

Reuters could not establish why the second wave was dropped. The internal data reported alongside it is suggestive. A June post by chief technology officer Andrew Bosworth said code changes to internal platforms rose 220% year on year, while changes that reached users as new or upgraded features rose 36%. Major technical and security incidents rose 40%, and time spent firefighting rose 70%.

Read together, those numbers describe agents producing far more activity without a matching rise in delivered value, plus a higher incident load. That maps onto two of Gartner's three reasons at once: unclear business value and inadequate risk controls.

Meta disputes the framing. It said the 60% figure covered scenarios for certain teams only and was never a company-wide plan. "Ultimately, we didn't move forward with every scenario from the exercise, and it was never assumed we would," the company said. It also said thousands of employees were redeployed to new AI-focused teams. That is a fair point, and it is why the ledger lists the second wave, not the whole programme, as canceled.

The wider question of whether AI-justified headcount cuts are really AI decisions at all is taken up in the piece on AI layoffs as an overhiring correction.

Where this ledger is weakest

This post makes a count, and counts are easy to attack. Here is where ours is soft.

Our search method misses things

The ledger was built from web searches of press coverage, company statements and named-source reporting during one research session in October 2026. We did not run a systematic text search of every earnings call transcript or every 10-K. A cancellation mentioned in passing on a mid-cap earnings call could easily be missing. A count of 3 is a floor, not a census.

Several primary sources also refused automated access, including Gartner's own newsroom and the Reuters and Guardian originals on Meta. The figures from those sources were checked against syndicated or secondary coverage that quoted them. The KPMG global percentages come from PPC Land's reading of the report, because the report PDF itself was no longer available at its original address.

Most entries are not agents, and a scale-back is not a cancellation

Strictly, McDonald's, Klarna, Commonwealth Bank and Taco Bell ran conversational automation, not agents that plan and act across systems. Including them stretches the category. We kept them because buyers, vendors and the press all called these deployments AI agents at the time, and because Gartner's prediction counts projects that organisations believe are agentic.

The other stretch is status. Klarna and Taco Bell narrowed their programmes rather than ending them. If you count only full cancellations of genuine agentic systems, the public ledger since June 2025 has 1 entry, and even Meta contests it. A sceptic could fairly say the prediction is untestable from public data. We would agree, and that is the argument for reading the survey data rather than waiting for headlines.

What the 40% will look like when it arrives

Gartner has since added two predictions that describe what the cancellations will probably look like. Neither will produce many press releases.

In February 2026, Gartner predicted that by 2027, half of companies that cut customer service staff because of AI will rehire for similar work, often under different job titles. The same release reported that only 20% of 321 customer service leaders surveyed in October 2025 had actually reduced staffing because of AI. The rehire will be quiet too: new titles, new requisitions, no announcement. What that does to support headcount is worked through in the analysis of support staffing after deflection.

In May 2026, Gartner predicted that by 2027, 40% of enterprises will demote or decommission autonomous agents because of governance gaps found only after production incidents. "Demote" is the important word. Most failing agents will lose permissions and become assistants again, which no external observer can see. The control model that would catch those gaps before an incident is laid out in the agent governance maturity model.

So our prediction is this. By the end of 2027 the surveys will show the 40% has arrived, and the public ledger will still have fewer than 20 named entries. We will keep updating this page as cases appear, and we will add any we missed with the date they were added.

Frequently asked questions

Where to start

Pull the list of every agent or AI assistant your company pays for or has in pilot. For each one, write down the number that justified it and who outside the project last checked that number. If the answer is nobody, you have found your Commonwealth Bank risk, and it costs an afternoon to fix. If you need thresholds to check against, start with the six kill criteria for AI projects.

Then do the thing the public record suggests nobody does. If you narrow or end a project, write one paragraph on why, with the cost figure included, and file it where the next project sponsor will read it. The industry's cancellation data is anonymous because every company keeps its own failures private. Yours does not have to be.

Related reading