Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, for three reasons: escalating costs, unclear business value and inadequate risk controls. None of these is really about the technology. Each comes from choices made before the first agent runs: which process, how value is measured, who carries the cost and how the work is governed.
The three causes
Gartner's analysts describe most agentic AI projects today as early experiments or proofs of concept driven by hype and often misapplied. They add that many use cases positioned as agentic do not need agentic implementations at all.
They also warn about "agent washing": vendors rebranding assistants, robotic process automation and chatbots as agents. Gartner estimated only about 130 of the thousands of agentic AI vendors are real.
Cause one: costs that keep growing
Agent projects often start as a pilot with a small model bill. Then volumes rise, prompts grow, retries multiply and someone has to watch it all. The bill arrives after the business case was approved.
How to avoid it: decide who carries the running cost before you start. If a provider is paid per verified outcome, model spend, retries and monitoring are their cost, not yours. If you build in-house, put budget caps and routing to cheaper models in from day one, and track cost per completed item, not per call.
Cause two: value nobody can show
A pilot that "saves time" without a measured baseline has nothing to show when the budget review comes. Hours saved that are not redeployed or removed do not show up anywhere.
How to avoid it: pick a process where the result lands in a system of record, like a posted invoice or an issued certificate. Measure the baseline from that system before you start. Write the outcome definition in advance. Then value is a number, not an opinion.
Cause three: risk controls added too late
Agents that act in real systems need controls: who approves what, what gets logged, what happens when the agent is unsure. When those questions come up after the pilot, legal and compliance teams stop the rollout, and they are right to.
Governance is the most common blocker. In Omdia's polling of MSPs and IT decision-makers, 47% named governance and compliance as the top barrier to agentic AI, far ahead of technical skills at 16%.
How to avoid it: design the controls with the process. Route low-confidence items to people. Keep licensed and financial decisions with named approvers. Record every step, model version and approval in a tamper-evident log from the first day.
A checklist before you start
- Does the process produce a result recorded in a system?
- Is there a baseline pulled from that system?
- Is there a written definition of a completed outcome?
- Who pays for model usage, retries and monitoring?
- Who approves decisions that need a license or move money?
- Is every step logged in a way an auditor could follow?
If you can answer all six, you have removed the three reasons Gartner gives for cancellation. Our guide to scoring a process for AI agents turns this into a scorecard, and how it works shows how we apply it.