At a Glance
- Gartner: more than 40% of agentic AI projects will be cancelled by 2027.
- The failures won’t be about model quality. They’ll be about scope, measurement, and unit economics.
The uncomfortable arithmetic
Two out of every five agent programs your leadership team is funding today will be shut down within eighteen months.
That’s not a forecast we made. It’s Gartner’s. And when four in ten of anything gets pulled inside a two-year window, it stops being a technology story. It’s a capital-allocation story.
The instinct is to blame the models. It’s the wrong instinct. Frontier capability has never been the enterprise bottleneck, and the September model releases only sharpen that point. This is where Generative AI services need to be evaluated beyond the model itself, with attention to how the agent is scoped, deployed, and measured.
What actually kills agent programs
Three patterns show up over and over.
The scope nobody policed. Every agent starts as a well-defined task and ends as a general assistant that “also handles a few other things.” The moment scope creeps, the evaluation stops working, and the trust collapses shortly after.
The cost curve nobody owns. Token spend behaves nothing like SaaS seat spend. It spikes with prompt length, retry loops, and tool chains, none of which show up on the invoice as anything a CFO can act on. By the time procurement notices, the pilot is quietly a P&L problem.
The success metric nobody agreed on. “It worked in the demo” is not an outcome. Programs that can’t produce a defensible before-and-after number are the first to get cut when budget season arrives.
Notice what these three have in common: none of them are model problems. All three are program design problems – decided in the first two weeks and rarely revisited.
Atgeir’s take:-
Three things we’re telling clients this week:
Fund outcomes, not agents. The pilots that survive 2027 are the ones with a single named metric they were built to move, a baseline they moved from, and a stakeholder who owns the number. If those three aren’t in the one-pager, the pilot is a demo. For Generative AI services, that same discipline is essential when moving from experimentation to production.
Tighten scope before you scale. A narrow agent that does one task exceptionally well compounds. A broad agent that does ten things adequately is a support ticket generator. The temptation is always to widen; resist it until the narrow version is boring and reliable.
Watch the unit economics from week one. Cost-per-successful-task is the number that decides whether an agent lives or dies at renewal. Track it from the first prototype, not the first invoice shock.
The agent era isn’t slowing down. But the unmanaged agent era is about to end, and the 40% cancellation number is how that ending will show up on your board deck.
The organisations that treat agent programs as normal software programs – with scope, owners, metrics, and unit economics will fund the survivors. The rest will fund the 40%.
Which side do you want to be on?