Thousands Sell AI Agents. Gartner Counts About 130 Real Ones. Ask These Three Questions Before You Sign

Every software vendor now has an "AI agent" on the pricing page. Gartner looked at that crowd in June 2025 and put a number on it: of the thousands of vendors describing their product as an agentic AI provider, only about 130 have substantial agentic capabilities. The rest, in Gartner's phrase, are "agent washing," the rebranding of existing assistants, chatbots and robotic process automation without the ability to plan and act on their own.

That number is the starting point for this page. Our earlier piece on enterprise AI agent adoption versus cancellation covered how fast the market is growing and how many projects will not survive it. This page is about the moment before any of that: the day a business owner sits across from a vendor and has to decide whether the thing being sold is real, and whether it will still be running in two years.

Two forecasts that point the same way

The first forecast comes from the same June 2025 release. Gartner expects over 40 percent of agentic AI projects to be canceled by the end of 2027, and it names three reasons: escalating costs, unclear business value, and inadequate risk controls. The analyst quoted in the release, Anushree Verma, added that most current agent projects are "early stage experiments or proof of concepts that are mostly driven by hype and are often misapplied."

The second forecast came almost a year later, on May 26, 2026, and it is about the projects that do reach production. By 2027, Gartner predicts, 40 percent of enterprises will demote or decommission autonomous AI agents because of governance gaps that were identified only after a production incident. The root cause, according to analyst Shiva Varma, is that companies treat agent governance as binary: an agent is either locked down or fully trusted, and the same controls get applied to every agent regardless of how much it is allowed to do.

Put the two releases together and a pattern appears. Projects die early for cost, value and risk. Projects that survive get pulled later because nobody matched the controls to the agent's actual scope. Both failures are decided before the contract is signed, in the questions the buyer did or did not ask.

The demand is real, which is exactly the problem

Gartner's 2026 Hype Cycle for Agentic AI reports that only 17 percent of organizations have deployed AI agents so far, while more than 60 percent expect to within two years. The same page calls it the most aggressive adoption curve among all emerging technologies in the survey, and notes that most deployments remain narrowly scoped and that fully autonomous agents are not ready for the majority of enterprise use cases.

A market where 60 percent of buyers intend to purchase something that only 17 percent have tested is a seller's market. Vendors know it. That is why the word "agent" is on every slide, and why the buyer, not the vendor, has to do the sorting.

Three questions that map to the three failure causes

Each of Gartner's cancellation reasons has a question that exposes it early. We use these three with our own clients, and we answer them about our own products.

One. What does a single completed task cost, all in? Not the seat price and not the monthly platform fee. The cost of one booked appointment, one qualified lead, one resolved ticket, including the model calls, the integrations and the human time when the agent hands off. A vendor who cannot give you a per-task number has not run the agent at volume, and escalating costs will find you after launch instead of before.

Two. What will we measure in ninety days, and who signs off that it worked? Pick one metric before the pilot starts: calls answered after hours, no-shows recovered, hours of back-office work removed. Name the person who will look at the number. Gartner's phrase for the second cancellation cause is "unclear business value," and value stays unclear for exactly as long as nobody agreed on the number in advance.

Three. What can the agent touch, who approves an action, and how do we switch it off? This is the governance question, and it is the one the May 2026 release is about. Ask which systems the agent can read, which it can write to, which actions require a human approval, and what the off-switch looks like on a bad day. An honest vendor will describe different controls for a scheduling agent than for one that moves money, because the scope is different. A vendor who offers one setting for everything is selling the binary governance Gartner calls the root cause.

What a real answer sounds like

A real agent vendor can show you a log of actions the agent took last week, with the approval points visible. It can tell you the cost per completed task from live traffic, not from a projection. It can show you the inventory of what the agent is connected to and the button that disconnects it. None of that requires a technical buyer. It requires a buyer who asks.

What we build at PATech

PATech builds and runs AI agents for service businesses, starting with the phone line, and we build them to answer these three questions on day one: a per-task cost, one agreed metric, and a scope with a human approval point and an off-switch. If you are comparing vendors right now, talk to PATech and bring the three questions. We will answer them about our own agent before we ask you to sign anything.

This material is for informational purposes only and is not legal, financial, or other professional advice.