Gartner says 4 in 10 AI-agent projects will be scrapped by 2027 — and most of the “agents” you're pitched aren't real

The research firm Gartner put a number on the AI-agent gold rush last year, and it's not the one the vendors quote at you. More than 40% of agentic AI projects, it reckons, will be scrapped before the end of 2027 — killed off by runaway costs, value nobody can point to, and risk controls that were never built. That's not a sceptic on a podcast. That's the firm whose reports get read in boardrooms to decide what to buy.

If you run a cleaning crew, a massage room or a trades business, you're not in a boardroom — but you're being sold the exact same thing, in a smaller box, this year. So it's worth understanding what that 40% is actually telling you, because it's not "AI agents are rubbish." It's something more useful.

There's now a word for the con: “agent washing”

Here's the part that should make you narrow your eyes at the next demo. Gartner reckons that of the thousands of companies calling themselves agentic-AI vendors, only around 130 are the real thing. The rest are doing what the industry has quietly named agent washing — taking a chatbot, a canned automation, or a plain old assistant, slapping the word "agent" on the box, and charging more for it.

It's the same move as "cloud" a decade ago and "AI-powered" a couple of years back. The label races ahead of the substance, because the label is what sells. And the smaller the buyer, the less time they've got to check under the bonnet — which is exactly the buyer agent washing is aimed at.

Why nearly half the real ones still get canned

Set the fakes aside for a second — even the genuine projects fail at a startling clip. Gartner's three reasons are worth reading slowly, because they're the same three that sink a small-business tool, not just an enterprise rollout:

• Cost that creeps. The demo is cheap. Running the thing across every enquiry, every day, is where the meter adds up — and a lot of projects only discover the real bill after they've committed.
• Value nobody can name. "We put an AI agent in" is not a result. If you can't say which job it took off your plate and how many hours that saved, there's nothing holding the subscription up when money gets tight.
• Risk controls that were an afterthought. An agent that can act on its own, with no human gate and no undo, is one bad afternoon away from a mess — and that's usually the reason the plug gets pulled, not the tech.

Notice none of those is "the AI wasn't smart enough." The models are fine. The projects die on the boring stuff — cost, proof, and guardrails — which is the stuff the hype skips and the stuff you can actually check before you pay.

The three questions that cut through it

You don't need to audit anyone's architecture. You need to make the vendor answer three plain questions, and watch whether they can:

• “What does it actually do on its own — and what does it cost me when I lean on it?” A real answer has a number in it. A wave of the hand or an “unlimited” badge means the bill is coming later, by surprise.
• “Which specific job does this take off me?” If the pitch is a vibe — “boosts productivity,” “AI-powered everything” — and not a task you'd recognise from your own week, that's agent washing talking.
• “Where does a human sit, and can I undo what it did?” If the honest answer is “it just runs,” that's not confidence, it's the missing risk control that's about to become someone's cancelled project.

A tool that can answer those crisply is in the 130, near enough. A tool that gets slippery on all three is in the 40% — you'll just be finding out on your own dime instead of Gartner's.

The shakeout isn't the tech dying — it's the froth clearing

It'd be easy to read all this as "so AI agents are a bust, wait it out." That's the wrong lesson, and the same Gartner has the other half of the numbers: it still expects a chunk of everyday work decisions to be made by agents within a few years, and a third of business software to have real agentic features baked in by 2028, up from basically none in 2024. The direction is real. What's not real is most of the current sales pitch.

That's what a shakeout is. The capability keeps climbing while the pretenders wash out — and the people who lose money are the ones who bought the loudest label right before it collapsed. The people who come out ahead bought the boring, checkable version: a specific job, a known cost, a human on the off switch.

Where we land on it

We'd rather you use those three questions on us than take our word for anything. Inside Dispatch, the AI does a named job — it reads your enquiries and drafts the replies, quotes the standard stuff off your own price list — and the send button sits behind you, so there's always a human and an undo. The AI parts are priced as add-ons with a clear allowance and metered usage past it, not folded into an “unlimited” badge, because we know what they cost to run and we'd rather you did too.

None of that is exciting. It's the deliberately unglamorous version of AI — the one built to still be here in 2027, when 40% of the flashier stuff has been quietly switched off. When the next tool tells you it's an “AI agent,” don't argue with the label. Ask what it does, what it costs, and where you sit. The real ones have answers. The washed ones have adjectives.