The 40% figure is a forecast tied to post incident governance gaps, not a measured failure rate, and the deployment model it tracks is the one most enterprise rollouts are still missing.
The 40% of enterprise AI agents that Gartner predicts will be demoted or decommissioned by 2027 aren't being pulled because the underlying models disappoint. They're being pulled because nobody built the off-switch first.
That framing comes from a June 12 ZDNET analysis by Mark Samuels summarizing practitioner discussions at the Snowflake Summit in San Francisco. The Gartner figure is a forecast, not a measured failure rate. It projects that, by 2027, four in ten organizations will abandon agent rollouts because governance gaps surface only after the agent ships to production and breaks something the evaluation harness never simulated. The distinction matters. A failure rate tells you the technology failed. A forecast tied to post-incident governance gaps tells you the deployment model failed.
The practitioners Samuels interviewed at Snowflake's vendor event offered three guiding lessons: focus on governance and frameworks, work with experts, and set clear outcomes. That is sound advice, and it is also the advice most enterprises have already heard. The reason it does not change the Gartner curve is that "governance," "expertise," and "outcomes" are abstractions until they are wired into the runtime. An agent that cannot be paused, rolled back, or audited is not governed. A workflow with no human checkpoint between the agent and a regulated action is not expert-in-the-loop. A metric with no owner is not an outcome.
The operating core that the Gartner forecast is actually measuring looks like five concrete things, and most enterprise agent rollouts still ship without most of them.
An evaluation harness that replays recorded production traffic against a frozen version of the model. Without it, the team discovers regressions only when a real user triggers one. The harness has to cover the failure modes an agent will actually face: ambiguous inputs, tool errors, partial system downtime, and the long tail of edge cases that appear in week two of a deployment and never surface in the demo.
Human-in-the-loop checkpoints at the action boundary the agent crosses, not at the chat boundary. The decision that needs human review is the one where the agent is about to commit an irreversible action: a refund, a code merge, a record update. A human approval flow on the prompt is theater.
A kill switch and a tested rollback path. "We can turn it off" is not a rollback path. A rollback path is a rehearsed procedure, with a named owner, that restores the prior state inside a defined time window. If the off-switch has never been pulled in production, the team does not know whether it works.
Owner-on-the-hook accountability for the workflow, not for the model. The person whose name is on the agent is the person who answers when the agent does the wrong thing. In most organizations, that role does not currently exist in any defined org chart. The platform team owns the agent. The line of business owns the workflow. Compliance owns the risk. The actual decision lives in the gap between them.
ROI measured against a defined workflow, with a stop condition. If the agent cannot articulate which workflow it improves, by how much, and over what measurement window, it is not an investment. It is a demo that survived procurement.
These five items are not a maturity model. They are the floor. The Gartner forecast is a prediction that, by 2027, most organizations will still not have built them. That is a deployment-model problem, not a model-quality problem.
The Snowflake context sharpens this rather than softens it. The Summit's product surface, including the Snowflake CoCo framework discussed alongside the agent lessons, comes from a vendor with a direct commercial interest in agent adoption succeeding at scale. Practitioner advice delivered on that stage is not automatically wrong, but it is delivered in a setting that rewards optimism about agents and silence about the failure modes. The reporter's job is to keep the forecast and the framework separate, and to ask which of the five items above the practitioners Samuels quoted actually had a named owner and a tested rollback in their production deployments. The ZDNET piece does not answer that question, and that gap is the story.
The watch item for the next two quarters is whether the agent platforms themselves start shipping the operating core as a product. The vendors that survive the Gartner curve will be the ones that make evaluation harnesses and kill switches default, with owner-bound metrics baked into the platform. The vendors that treat governance as a services line will be the ones whose customers end up in the 40%.
Practitioner sources from ZDNET's Snowflake Summit reporting: Matt Luizzi, VP of analytics at Whoop; Madeleine Want, VP of data at Fanatics; Sriram Sitaraman, CIO at Synopsys.