Intel warns: agentic AI has left its toddler stage and needs a new model of control

Intel suggests thinking of agentic AI as a child that has suddenly stopped crawling and started running. Companies are still debating policies and committees, while autonomous agents are already gaining access to workflows, budgets and critical systems. That means the old approaches to control no longer work.

From chatbots to agents

In a piece for MIT Technology Review, the authors compare the current stage of agentic AI to early childhood. The shift happened fast: in late 2025 and early 2026, a wave of no-code tools hit the market, along with OpenClaw, an open-source personal agent on GitHub. AI used to mostly wait for a prompt in a chatbot interface. Now it works through a chain of actions on its own: it reads data, makes intermediate decisions and pushes the process forward without constant human approval.

This, according to Intel, is where the old governance model breaks. Businesses used to focus on the risks of a model's output: hallucinations, drift, data leakage, data poisoning. But once an agent starts executing a workflow at machine speed, human in the loop loses much of its meaning. For a company, the question is no longer "what did the bot say" but "what did the bot manage to do": which records it changed, which permissions it used and which downstream actions it triggered.

Where control breaks down

The piece's central argument is that it's already too late to govern agents through documents and committees. Guardrails need to be built directly into code and business processes, calibrated to risk level, access rights and potential liability. Otherwise, an agent running on probabilistic logic gets too much freedom in systems where mistakes are expensive. This changes both the architecture and the lines of accountability: if an agent makes a mistake, the model won't be held responsible. The company that gave it access to production will.

"AI does the work, and people carry the risk."

• an agent can chain actions across several corporate systems and gain more influence than any single employee;
• companies quickly pile up service accounts, long-lived API tokens and permissions to modify key files and data;
• a new layer of shadow AI emerges as employees build their own assistants without architecture, support or proper auditing;
• when an employee moves to another department or leaves the company, "orphaned" agents tied to their ID and permissions remain;
• neglected AI pilots and "zombie projects" keep running in the cloud and burning resources with no clear owner.

The authors also draw an analogy to a toddler suddenly handed a toy that's far too powerful. For the enterprise, the lesson is simple: you can't release an autonomous agent into production without observability, the ability to revoke access quickly and a forced-stop mechanism. Companies need discovery, an audit trail, remediation and a clear deactivation procedure. Otherwise, the benefits of automation vanish with the first incident, and cleaning up the aftermath costs more than the deployment was worth.

The price of autonomy

Intel also pushes back on the popular idea that agentic AI is simply a way to cut payroll. In the enterprise model, costs behave differently: instead of a fixed per-user license, companies pay for tokens, compute and external APIs as workflows grow. Labor savings turn out to be too crude a metric, because automation also drives up spending on observability, support, security and financial oversight of the entire agent infrastructure.

The piece cites a December IDC survey commissioned by DataRobot: 96% of companies deploying generative AI and 92% of organizations deploying agentic AI reported that costs were higher or much higher than expected. That's an important signal for business: the problem isn't just the model's per-token price. Money also goes to maintaining a fleet of internal agents, fixing errors, reviewing access rights and staffing the teams that have to keep it all under control.

The unpredictability of usage-based economics makes the problem worse. Unlike classic FinOps, where cloud costs are more or less deterministic, agentic AI behaves probabilistically: long call chains, autonomous loops and planning errors can drive the cost of a single session to extreme levels. The authors even note that some AI-first founders are already facing costs of around $100 000 for a single agent session. Without limits set from the start, an autonomous workflow can easily eat up a budget comparable to hiring another employee.

What it means

For business, agentic AI is no longer an experimental interface on top of an LLM but a new operational layer. The winners won't be the companies that let employees "build their own agent" fastest, but those that first build access rights, auditing, deactivation, budget limits and continuous oversight of what agents do in real systems into their processes.