Core point
No one is building the structural learning loops. Who watches the agent? Who decides on interventions? Who writes outcomes back into the journey? As long as "Agent Steward" is a side task of tool admins instead of a role, every platform investment is structurally still a pilot.
An agent isn't a tool — it's an employee without a manager. A human in the same position would have: an onboarding plan, a steward (mentor/lead), performance reviews, escalation paths, data-protection training, clear decision rights. The agent gets: an API key.
The research is unambiguous here. KPMG "AI governance for the agentic AI era" (2025) makes the point: classical AI governance is built for predict/generate models. Agentic systems take actions — and need their own structures for audit trails, identity, liability. Most organisations simply don't have that layer. Mayer Brown adds the legal angle: under the EU AI Act, agent identity, authorisation scopes and audit trails are no longer best practice — they're required.
The market reaction confirms the gap. Vendors like SAS bake stewardship layers into their platforms (SAS Viya, April 2026) — because the line role is missing in most companies. That's symptom, not cure.
"The primary barrier to widespread adoption is no longer a lack of capability — it is a lack of trust."a21.ai · From Ignore to Execute (2025)
And the DACH practitioner voice says the same. Thomas Maxeiner (Beyondbuzzwords, Feb 2026): "Mittelstand companies need an agentic governance council." Roover.de puts it more bluntly: "The biggest hurdle is rarely the model — it's operations."
This isn't hype. It's the direct consequence of 9.5 % at scale, 95 % pilot failure, and the 40 % cancellation forecast. The companies that anchor stewardship as a line function are the 5 % that will still be running in 2027.