Gartner forecasts over 40% of agentic AI projects canceled even as a third of software ships with agents

The two figures count different things. The 40% is a share of projects ^[2] (initiatives a company starts), the one-third is a share of applications ^[1] (products a vendor ships), and one buyer can land in both groups at once: a canceled in-house pilot on one side, an agent feature switched on inside software it already licenses on the other. On Quartz's reading, agent capability reaches users through software they already have, while a large share of the spending on custom-built versions ends up written off ^[5]. The Quartz piece does not say how many of the canceled projects were custom builds, how many projects Gartner counted, or what they cost, so the write-off cannot be priced from this source.

Take the cancellation figure at face value and the exposure is easy to show. Above 40% canceled leaves fewer than 60 of every 100 projects standing ^[6]. Split a budget evenly across 100 projects, release it on day one, and at least 40 of those shares go to work that is later stopped ^[7], with the money out the door before any of the causes Gartner cites had shown up ^[3]. A tranche structure stops payment at the first failed gate, so a doomed project costs one tranche and not the whole line.

The first tranches write themselves from the deployments Quartz describes as typical today: a customer-service agent that resolves a billing question but escalates any refund above a set threshold, and a procurement agent that sends a purchase order while a person signs off before payment ^[11]. Each has a threshold and a named human, which is exactly what a gate can test. Quartz adds that companies deploy before they understand what the software needs to run safely ^[12].

We think the budget that should not exist is a single multi-year line for a broad rollout. The case for staging is strongest for custom builds, or rather for the builds Quartz says get written off. The counter-thesis is that staging costs time. If embedded agents arrive on Gartner's schedule, a company that releases money slowly buys the same features at the vendor's pace anyway, having paid for gates it did not need. Waiting has its own limit: even the 2028 forecast leaves roughly two-thirds of applications without agentic capability ^[8], so most categories will still be buyable later, and nothing forces a rush.

We would be wrong on two counts. If the cancellations cluster in projects that already had audit logs, thresholds and owners, the gates do not predict survival. And if the final count for 2027 lands well under 40%, the loss that staging protects against is smaller than the delay it adds.