Part of what we do for a living is automating processes with AI. It's in our interest that you hire us for an agent project. And yet we're opening with this: Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027, due to escalating costs, unclear business value or insufficient risk controls. Precisely because this is how we make a living, we'd rather tell you why those projects die before yours becomes one of them.
The interesting question isn't whether AI agents work. They work, and they keep getting better. The question is why so many projects die after starting with a flawless demo and an enthusiastic steering committee. The short answer: it's almost never the model's fault.
The numbers that never make it into the demo
Three data points, from two different sources, tell the same story from different angles:
- 1.Gartner, June 2025: over 40% of agentic AI projects canceled by the end of 2027. In the same note, a less-quoted and more uncomfortable figure: of the thousands of vendors marketing themselves as "agentic AI", Gartner estimates only about 130 are the real thing. The rest it calls agent washing: the same old chatbots, RPA and assistants, relabeled.
- 2.MIT, August 2025 (The State of AI in Business report): after reviewing over 300 public initiatives, 52 interviews and 153 executive surveys, it concluded that 95% of generative AI pilots showed no measurable return, with $30–40 billion already spent. Honest caveat: the study measured P&L impact six months after the pilot, so diffuse productivity gains didn't count. Even so, the pattern among the 5% that did work is telling: deep integration into the workflow, not generic tools bolted on top of the process.
- 3.Gartner again, May 2026: by 2027, 40% of enterprises will demote or decommission autonomous agents because of governance gaps discovered after a production incident. Read that slowly: after.
And so nobody reads this as "AI doesn't work": the same Gartner expects that by 2028, at least 15% of day-to-day operational decisions will be made autonomously (it was 0% in 2024) and a third of enterprise software will ship with agentic AI built in. The technology is here to stay. Badly framed projects are not.
Why they die (and it's not the model)
First: they automate a broken process. An agent on top of a process nobody has mapped fixes nothing; it hands you back the same broken process, just faster and with fewer people watching. Half the value of an automation project is in the process map it forces you to draw before writing a single line. If that map doesn't exist, the project isn't ready — whatever AI happens to be fashionable that quarter.
Second: they mistake the demo for production. The demo shows the happy path: the well-scanned document, the well-written request, the data where it belongs. Production is everything else: the upside-down invoice, the ambiguous email, the system that responds late. That "everything else" is where most of the real effort goes, and it's exactly what the demo never shows. The MIT report points right there: the pilots that survived were the ones integrated into the real workflow, ugly cases included.
Third: binary governance. That's Gartner's May warning: most organizations treat agents as all-or-nothing. Either the agent is so locked down it adds nothing, or it's fully trusted and one day it acts where it shouldn't. The recommendation —which we subscribe to— is proportional governance by autonomy level: you don't govern an agent that only reads and summarizes the same way as one that recommends with human review, or one that executes actions. The first needs little more than data access control; the last needs scope limits, review, and a kill switch someone knows how to use on a Sunday.
Four myths that projects pay for
- ✗"The agent replaces a person." It replaces bounded tasks within a person's job. Projects sold internally as headcount cuts tend to be badly scoped, breed resistance, and end up on the 40% list.
- ✗"More autonomy means more value." Autonomy isn't value: it's a governance and risk cost you buy when the return pays for it. Many processes perform better with an agent that recommends and a person who decides.
- ✗"It's a technology project." It's a process project with a piece of technology inside. If the conversation starts with the tool instead of the process, you already know how it ends.
- ✗"If the pilot works, production works." MIT measured exactly the opposite: 95% die between the pilot and the P&L. The pilot proves the technology can; production demands that the process, the data and the governance keep up.
How we approach it
For years we've applied a deeply boring rule to infrastructure that now turns out to be the key to agents: least privilege. Just as we don't give a monitoring tool an administrator token —we give it a read-only role—, an AI agent starts by observing. It has to earn the right to recommend with metrics. To act on its own, even more so. In our AI and automation service that's the order of things, and we don't change it just because the demo is spectacular.
The second thing is measuring before automating. A concrete process, with a baseline: how many hours a month it costs, how many errors it produces, how long it takes. Without a baseline there's no provable return at six months, and without provable return the project dies in a steering committee — which is exactly what the MIT study describes. And third, human review wherever an error costs money or reputation: the goal isn't to remove the person, it's to remove the part of their job a machine does better.
And the usual house rule: we don't sell licenses for any AI platform. If the right answer for your case is "don't automate this yet", that's the answer you'll get. Our incentive is for the project to live and defend itself with numbers, not for the contract to get signed.
The six questions before signing anything
- 1.Which process, exactly, and who really knows it? If nobody in-house can draw it on a whiteboard, it's not ready to be automated.
- 2.What does it cost today, in hours and errors? Without a baseline there will be no return to show six months from now.
- 3.What level of autonomy does it need: read, recommend, or act? Start one level below what you think.
- 4.What data and systems does it access, with which credentials, and who reviews them? An agent with admin access "so it works" is the same mistake we've been watching for years with API tokens.
- 5.Who validates the output and what happens when it gets things wrong? It will. The question is whether the error costs a click or a customer.
- 6.How do you switch it off? Who, with which button, and what happens to the process in the meantime. If the answer starts with "we'd have to call the vendor", there's no project yet.
In short
AI agents are here to stay, and used well they save real hours. The difference between the 40% that will get canceled and the rest isn't in the model: it's in whether someone mapped the process, measured the baseline, sized the autonomy and had the kill switch ready before the demo, not after the incident. It's less glamorous than the demo. It's also the only thing that survives the six-month committee review.
Sources (verified): Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027" (Jun 25, 2025) — gartner.com; Gartner, "Applying Uniform Governance Across AI Agents Will Lead to Enterprise AI Agent Failure" (May 26, 2026) — gartner.com; MIT, "The State of AI in Business 2025" (project NANDA, Aug 2025), coverage at — Fortune/Yahoo Finance. The methodological criticism of the MIT study is acknowledged in the text.
Got a process begging for automation (or a pilot that won't take off)?
At everyWAN we automate processes with AI with the same discipline we apply to infrastructure: least privilege, measurement and reversibility. And if your case isn't automatable yet, we'll tell you just as plainly.
Talk to everyWAN
Microsoft 365
ES 
CAT 
EN