Gartner cools the AI agent rush: autonomy now has to pass audit

The Register reports a Gartner forecast that cuts through one of the year’s loudest AI narratives: four in ten AI agents could be demoted or thrown out. That is not a verdict that agentic AI is useless. It is a sharper signal that companies do not yet have strong enough governance for software that can plan, choose steps and trigger actions rather than merely answer prompts.

The decisive word is not intelligence. It is authority. A conventional chatbot can produce a bad answer. An agent can turn a bad answer into a workflow, pass it through a tool, open a ticket, modify a record or chain several tasks together. That is why Gartner’s reported prediction lands in the exact zone where demo-stage enthusiasm meets audit, security, legal review and process ownership.

The tension is especially awkward because agentic AI is often sold on the promise of autonomy. But autonomy without boundaries creates questions many organizations have not settled: who owns the decision, where the automation chain stops, how each step is logged, when a human must approve an action and what happens when the model confidently takes the wrong path. Frameworks such as the NIST AI Risk Management Framework already push the discussion toward measurable risk rather than vague productivity claims.

So Gartner’s forecast should not be read as a dramatic turn against AI. It looks more like a sorting process. Agents with narrow jobs, clear inputs, limited permissions and reliable audit trails are more likely to survive. Agents sold as universal digital employees, without hard controls over tools and data, are the obvious candidates for reduced authority.

The regulatory backdrop makes improvisation even harder. The European AI Act and similar pressure elsewhere do not reward a polished model story alone. They demand accountability, risk management and evidence that a system operates inside defined limits. In that environment, an agent that acts across multiple applications is no longer just an IT experiment. It becomes part of a controlled business process.

For AI platform vendors, this is an uncomfortable but useful test. The market has learned to talk quickly about agents, orchestration and automated workflows. It has learned more slowly to talk about approvals, logs, rollback, access limits and failure modes. Documentation for projects such as the OpenAI Agents SDK shows how practical agent design depends on defining tools, boundaries and supervision, not simply letting a model “do work.”

If four in ten agents are indeed pulled back, that will not mean agentic AI has collapsed. It will mean the first wave of deployments was overpromised and under-governed. The next phase will be less theatrical: smaller agents, narrower tasks, stricter controls and many more questions before software is allowed to press the button on its own.