The AI Agent You're Buying Probably Isn't One — Here's What You're Actually Getting
More than 40% of corporate AI agent projects will be cancelled by the end of 2027 — a number boards are quoting without the date attached. Those projects aren't failing because agents don't work. They're failing because most of what companies bought was never an agent. The map of the enterprise automation stack, and the three questions that tell a real agent from a repaint before you sign.
The Promise
- The stack underneath the label is real and worth understanding. RPA, business process automation, integration platforms and the newer orchestration layer Gartner calls BOAT are genuine technologies, and consolidation onto a single platform is where enterprise automation is actually heading — Gartner projects 70% of enterprises there by 2030, up from 5% today.
- The dividing line is testable before you sign. An agent proactively executes multi-step work, decides inside guardrails you set, and persists across sessions. Anything that cannot re-plan when the situation changes is automation with a friendly voice — and you can prove which one you're being sold in a live demo.
The Risk
- Agent washing is the mechanism underneath the cancellation statistic: existing chatbots, RPA and workflow tools relabelled without the capability the word implies. Of the thousands of vendors selling agentic AI, Gartner puts roughly 130 in the real category.
- Every relabel compounds on a meter. Agent pricing is consumption-based, and Gartner projects 60% of organisations running AI will hit unbudgeted cost overruns through 2029. A washed product on a consumption contract is the worst combination available — none of the autonomy, all of the open-ended cost.
The number, and the date nobody checked
More than 40% of corporate AI agent projects will be cancelled by the end of 2027. The figure is everywhere this summer — boards quoting it, vendors quoting it. Gartner made that prediction in June of 2025. A Forbes piece put it back in circulation on 7 July, and most of the coverage since has dropped the date and presented a year-old forecast as breaking news.
The number is a lesson in itself. But the part that matters is why those projects fail. Not because AI agents don’t work. Because most of what companies bought was never an agent to begin with.
The map that protects you
Most enterprises spent the last decade building the same four layers, usually without a map. At the bottom, RPA — software robots doing rule-based tasks: clicking, copying, moving data between systems that were never designed to talk. Above it, business process automation, which automates a whole multi-step path someone designed in advance. Then iPaaS, the plumbing connecting systems that don’t natively connect. And on top, the layer vendors are selling hardest and boards have rarely heard named: BOAT — business orchestration and automation technologies, Gartner’s term for consolidating all of it under one roof.
Every one of those layers is deterministic. It follows rules a human designed ahead of time. Change the situation in a way the designer didn’t foresee and it stops, or it breaks.
So use Gartner’s own dividing line. An assistant reacts to a prompt. An agent proactively executes multi-step work toward a goal, decides inside guardrails you set, and persists across sessions instead of forgetting the moment you close the window. The test underneath all three: when the situation changes, can it resequence its objectives, weigh the trade-offs, and adapt its plan? RPA can’t. A scripted chatbot can’t. A workflow engine can’t.
Adoption is the surface the label sticks to. Gartner’s 2026 survey found 17% of organisations have actually deployed agents, while more than 60% intend to within two years — the steepest adoption curve Gartner has measured for any emerging technology. When intent runs three times ahead of experience, most people evaluating an agent have never operated one.
Three questions before the signature
One: when the situation changes and the script no longer fits, does the system re-plan on its own, or stop and wait for a human? That is reasoning versus rules.
Two: run the same task twice on different inputs, in front of me. Did it adapt, or did it repeat? That is agency versus automation.
Three: what does this cost when the agent runs ten times more often than it did in the demo? That is the question the meter contract is counting on you not to ask.
For a board director the question was never whether the company has an AI agent strategy. It’s sharper than that: when a vendor said agentic, what did our own team verify underneath it? Ask it before the invoice does.