Gartner expects around forty percent of agentic AI projects to be cancelled by the end of 2027, mostly for poor planning and unmet expectations. Having watched a number of these up close, the pattern is more specific than that, and more avoidable.
Failure one: no governed knowledge
The project starts with a model and a channel, and the knowledge is whatever documents someone could find. Six weeks in, the AI is confidently wrong about a refund policy, a senior person sees it, and trust collapses faster than it was built. Nothing recovers a programme from that.
The fix is boring and it has to come first: one governed knowledge source with ownership, versioning and an approval step, and retrieval that answers only from it.
Failure two: a pilot that cannot become production
The pilot runs on exported data, a test tenant and a friendly subset of traffic. It performs well, everyone is pleased, and then production turns out to need identity resolution, rate limits, audit logging, failure handling and a support rota. The pilot proved the model and nothing about the operation.
Run the pilot on live traffic in the real environment, smaller in scope rather than lighter in rigour. A narrow slice of production teaches you more than a broad slice of sandbox.
Failure three: measuring the wrong thing
Deflection is the metric that kills programmes, because it rewards the system for being unhelpful. Containment is the same trap wearing a different name. Both look excellent while repeat contact climbs and satisfaction falls, and by the time anyone notices, the numbers that justified the budget are the numbers nobody believes.
Failure four: no owner on the business side
If the programme lives entirely in IT, the approval boundaries never get set, because the person who can decide that the AI may issue a credit under a certain value does not sit in IT. These projects do not fail loudly. They stall, get re-scoped to something safer, and are quietly not renewed.
The tell, in the first month
Ask two questions. Who owns the knowledge base, by name? And what is the AI allowed to do without asking anyone? If either answer is vague in month one, the project is already in the forty percent.