Gartner has put a number on something a lot of operators have felt for a while. By the end of 2027, more than 40% of AI agent projects will be cancelled. Not paused, cancelled. The reasons it gives are not new ones but age old challenges for business improvement and transformation projects: cost overruns, unclear business value, and controls that were never really in place.

It is tempting to read that as a verdict on the technology. It isn’t. The models are the most capable they have ever been, and they keep getting cheaper, with the major labs cutting prices again only last month. Adoption is climbing fast at the same time, which is what makes the cancellation figure so striking. Everyone is moving, and close to half will walk away with nothing to show for it.

The projects that die rarely die because the AI couldn’t do the task. They die because the business around the AI wasn’t ready to hand it one. That distinction matters, because it changes what you do about it.

What actually goes wrong

When we get called in after a pilot has stalled, the story is usually the same. Someone bought a capable tool, pointed it at a real problem, and watched it produce something that looked sharp in a demo and came apart in daily use.

Dig into why and it is almost never the model. It is the ground the model was standing on.

The process the agent was meant to run had never been written down. It lived in a few people’s heads, with exceptions nobody had captured, so the agent had no way to learn the parts of the job that were never made explicit.

The data it needed sat across four systems that don’t talk to each other, in formats that made sense to the teams who owned them and to nobody else.

No one had agreed what good looked like. There was no number the project was meant to move, so when someone finally asked whether it was working, there was no honest way to answer.

And often there was no owner. The pilot belonged to everyone, which meant it belonged to no one, and when it hit its first real obstacle there was nobody whose job it was to push through.

None of that is an AI problem. It is an operations problem that AI happened to expose.

Why so many AI agent projects fail

The 40% stops being surprising once you notice how most of these projects begin. They begin with the technology. A capable tool arrives, it is exciting, and the question becomes “what can we do with this?” rather than “what do we most need done, and is this the right way to do it?”

That order is backwards, and it produces a predictable kind of failure. You automate a process that was shaky to begin with, and an agent running a shaky process just breaks faster and at greater scale than a person ever could.

There is a lot of noise in the market too. Plenty of tools are sold as agentic that are really a chatbot with a longer memory. Some of the cancelled projects were never viable, because the thing being bought couldn’t do what the label promised. Gartner has a name for this, “agent washing,” and it is worth being alert to when you are the one being sold to.

What the survivors do differently

The projects that make it through are not the ones with the cleverest model. They are the ones that did the dull work first.

They started with a process that was well understood and worth improving, and wrote it down properly before automating a line of it. The data went into a state the agent could actually use. Everyone agreed, up front, what the project was meant to change and how they would know. And it had an owner with the authority to make calls when reality drifted from the plan.

That work is not exciting, which is exactly why it gets skipped. It is also the whole difference between a project that ships and one that becomes a line in Gartner’s 40%.

The encouraging part is that this is the part you can get right without a research lab or a global consultancy on the payroll. Documenting a process, tidying your data, and agreeing a measure of success is ordinary operational discipline. It just has to happen before the technology, rather than after it has gone wrong.

Where this leaves you

If you are weighing up an agent project, the useful question is not which model to use. It is whether the process underneath it is solid enough to hand over. If it isn’t, no model will rescue it, and you will probably end up in the 40%. If it is, the technology tends to be the easy part.

The businesses that get real value from AI over the next two years won’t be the ones that moved fastest on tooling. They will be the ones that did the groundwork the other 40% skipped. In our experience that work has far less to do with AI than people expect, and it is usually the best place to start.