AI Agents at Work

95% of pilots. 40% of projects. None of it was the technology.

MIT found that 95 percent of corporate AI pilots delivered no measurable return. Gartner expects over 40 percent of agentic AI projects to be canceled by the end of 2027 - for escalating costs, unclear business value, or inadequate risk controls.

Read that list again. Not one of those three is a statement about the technology. Each is a decision somebody did not make before switching it on.

Two things that already happened

In February 2024 a payments company announced its AI assistant was doing the work of 700 people, and it handled roughly 2.3 million conversations in its first month. Fifteen months later the company was hiring humans again, and its chief executive was explaining that the AI had been cheaper and the work worse.

The same month, a tribunal ruled an airline responsible for what its chatbot told a customer. The airline had argued the chatbot was a separate legal entity, responsible for its own actions. That argument lost, and has been losing ever since.

What your agent says, you said. That is not a technical constraint - it is one you have already agreed to.

The P.I.L.O.T. Method

• P - PICK. Which work is agent-shaped: repetitive, rule-checkable, reversible, bounded, verifiable. Never the task that annoys you most - and this chapter explains why that trap is so convincing.
• I - INSTRUCT. A brief in five sections, edge cases longest by far - plus the diagnosis that catches more broken projects than anything else here: if you cannot write the brief, the problem was never the agent.
• L - LIMIT. Three budgets - what it may see, what it may spend, and what it may say on your behalf. Almost nobody sets the third, and it is the one that produces the incidents.
• O - OBSERVE. Four numbers, weekly, in fifteen minutes - defined before you switch anything on, because afterwards you will always find a story in which it succeeded.
• T - TRANSFER. Three kinds of handoff, and what travels with each so the person receiving it does not conclude it would have been faster to do it themselves.

Six functions, 24 prompts each

Customer support, sales and outbound, finance and admin, marketing and content, recruiting and HR, operations and scheduling. Each with the trap specific to that function, a case where it went wrong, and a full worked walk-through of all five steps on one real task.

Nothing in it expires

No code, no product recommendation, no framework named in the method. The tools change every few months; the decisions do not. What is this allowed to touch? How will we know whether it worked? Who is answerable when it gets something wrong? None of those gets easier when the technology improves.

Written for you if

• Somebody has asked you to "look at agents" and you do not code
• You cannot tell whether what you are being sold is an agent, an assistant, or a rebrand - roughly one vendor in a hundred was judged genuinely agentic
• You have a pilot running and no idea how you would defend it in a review

Also included

Five fill-in instruments: a nine-question task screen, a brief template with a worked example, a three-budget permission matrix, a twelve-point pre-launch checklist, and a one-page incident note - plus a chapter on surviving the sales meeting and a thirty-day plan that starts Monday.

George Chu holds a Master of Engineering in Software Engineering from Peking University and is a Senior Member of the IEEE.