Most AI agent projects that fail were too big on day one

Last year Gartner predicted that more than 40% of agentic AI projects will be cancelled by the end of 2027, because of rising costs, unclear business value or weak risk controls. This August they put agentic AI at the top of their hype cycle, the "peak of inflated expectations." Which is analyst for "everyone's excited and some of this is going to hurt."

The same Gartner release had a line I liked. Out of the thousands of vendors calling their product an AI agent, they estimated only about 130 actually were one. The rest is what they called "agent washing," a chatbot or an old automation with a new label on it.

Big companies are pulling ahead

McKinsey's State of AI 2026 report found that 40% of companies with more than $1 billion in revenue are now scaling AI agents, up from 27%. Smaller organisations are sitting at 22% and haven't moved. If you run a 10 or 50 person business, that's the number I'd pay attention to.

I don't think small businesses are behind for lack of budget. A lot of the owners we talk to tried something, got burned, and gave up. Usually the first project was too ambitious.

The mistake I see most

Someone decides they want "an AI agent for sales" or "an AI that handles customer service." Those are departments. The agent ends up with access to everything, a fuzzy job description and no clear way to tell whether it's working. Six months later nobody can say what it saved, and it quietly gets switched off.

There's a worse version. In April, a coding agent at a small software company called PocketOS deleted the production database and its backups in about nine seconds. Security firm Cyera went through thousands of public AI incidents and found 188 cases where an AI system did damage in production with no attacker involved. In most of these stories the agent simply had more access than its job needed.

How to pick a first agent that lasts

Start with one job you can describe in a sentence and measure with a number. A few that tend to work well:

• Answer the phone after 6pm, qualify the caller and book them into the calendar. You measure booked appointments.
• Reply to every website form within a minute with the right next step. You measure response time.
• Research each new lead before a salesperson calls. You measure minutes saved per call.
• Follow up on quotes that have gone quiet for a week. You measure quotes that come back to life.

Then keep its permissions tight. An agent that books appointments needs your calendar. It doesn't need to delete anything, and it definitely doesn't need admin access to your database.

And run it next to a person for the first couple of weeks. You'll catch the odd cases early (there are always odd cases), and when it's time to decide whether to keep it you'll have real numbers instead of a feeling.

None of this is exciting. The agents still running a year from now will mostly be the boring ones.