Why most AI agent projects never make it past the pilot

In June 2025 Gartner predicted that over 40% of agentic AI projects will be canceled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls. It also estimated that only about 130 of the thousands of vendors selling "agentic AI" offer the real thing; the rest rebrand chatbots and scripts. None of that says agents do not work. It says most projects skip the part between the demo and production.

Five gaps between a pilot and production

• No job to measure. "Use AI for support" is not a goal. "Answer order-status questions without a person, within two minutes" is.
• Tested on happy paths. The pilot used ten clean examples. Real traffic sends typos, angry customers, half-finished forms and questions nobody planned for.
• Too much access. An agent holding admin keys is one bad instruction away from a deleted table or a refund it should never have issued.
• No record of what it did. When a customer complains, nobody can see which tool the agent called, or why.
• Nobody owns it after launch. Models change, prompts drift and data shifts. An agent nobody reviews gets worse quietly.

The checklist we run before an agent goes live

• One named job, and one number that says it is working.
• A test set of real past cases, awkward ones included, run again after every change.
• Permissions limited to what the job needs, with approval for anything that moves money or deletes data.
• Every tool call and decision logged where your team can read it.
• A hand-over to a person when it is unsure, and it says it is an AI.
• A monthly cost and response-time limit, with alerts.
• A named person who reads transcripts every week for the first month.

Your agent speaks for your business

Moffatt v. Air Canada (2024) is the reminder: a tribunal held the airline responsible for what its website chatbot told a customer about fares. Whatever your agent says or does, your business said or did it. That is why the checklist is part of the build, not a phase you can skip to hit a launch date.

Start smaller than you want to

The agents that survive start with one workflow, run on a slice of real traffic, and earn more access as the logs prove them right. That is slower than a launch announcement, and much faster than a canceled project.