80% of the Fortune 500 run AI agents. Gartner expects 40% of projects to die

MIT Technology Review Insights published a piece this week saying more than 80% of Fortune 500 companies have adopted agentic AI. The piece is sponsored content, produced with the software vendor NiCE, and the only person quoted in it works there. Gartner, which sells nothing to these buyers, expects more than 40% of agentic AI projects to be cancelled by the end of 2027.

The vendor's argument

Arun Chandra, chief operating officer at NiCE, puts the failures down to plumbing rather than models. Companies bolt agents onto processes that were already broken, then wonder why the results disappoint.

“The efficacy of these AI agents is purely a function of the context, the knowledge, and the data they can ingest and use.”

— Arun Chandra, Chief Operating Officer, NiCE, MIT Technology Review Insights, 3 September 2026

Quote source: MIT Technology Review, 3 September 2026

He also warns against applying AI to an outdated or inefficient workflow. The point holds, and it is the standard argument of vendors who sell the workflow layer underneath the agent.

The independent numbers

Gartner put the cancellation figure at over 40% by the end of 2027 and named the causes: escalating costs, unclear business value, inadequate risk controls. Anushree Verma, senior director analyst there, described most current projects as early stage experiments driven by hype and often misapplied.

The same research names a supply-side problem. Gartner counts roughly 130 real agentic AI vendors among the thousands claiming the label and calls the rest agent washing: assistants, chatbots and robotic process automation rebranded without the underlying capability. A buyer reading that 80% adoption figure is reading a number that includes a lot of rebranded chatbots.

Why the pilots stall

Gartner lists inadequate risk controls, and agents make that concrete. One that browses, calls tools and moves data across systems widens the attack surface in ways a chatbot never did. We covered how MCP servers turned into a new attack surface, and the running tally of times agents have been used to breach real companies now sits at 17.

Governance is where those facts meet procurement. A pilot answers whether the agent can do the task. Production asks who is accountable when it does the task wrong at three in the morning across four systems.

Two questions before scaling

• Can you name the workflow the agent replaces, and did it work before the agent arrived? If the process was broken, the pilot measured the wrong thing.
• What happens when the agent acts on bad input, and who finds out? If the answer starts with what the model should do, that is not a control.

The 80% number and the 40% number describe different things. One counts companies that have started, the other counts projects that will stop. Neither the vendor nor the analyst has published the number in between: the share of those 80% that reached production and stayed there.

This piece is informational, not a recommendation to buy, sell, or hold any asset.