Intel warns: agentic AI has moved beyond its "childhood" stage and requires a new control model

Intel warns: agentic AI has left the 'childhood' stage and requires a new control model

Intel proposes viewing agentic AI as a child who suddenly stopped crawling and started running. Companies are still arguing about policies and committees, but autonomous agents are already gaining access to workflows, budgets, and critical systems — meaning old control approaches no longer work.

From chatbots to agents

In an article for MIT Technology Review, the authors compare the current stage of agentic AI to 'childhood'. The transition happened quickly: 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. If previously AI mostly waited for a request in a chatbot interface, now it runs through a chain of actions on its own: reads data, makes intermediate decisions, and moves the process forward without constant human confirmation.

This is where, according to Intel, the previous governance model breaks down. Previously, business concentrated on the risks of the model's response: hallucinations, drift, data leaks, data poisoning. But when an agent starts executing workflows at machine speed, the meaning of human in the loop drops sharply.

For a company, this is no longer a question of "what did the bot say," but "what did the bot manage to do" — for example, which records it changed, which permissions it used, and which actions it triggered further down the chain.

Where control breaks

The main point of the article is that managing agents through documents and committees is already too late. Constraints should be built directly into the code and business processes with consideration for risk level, access rights, and potential liability. Otherwise, an agent with probabilistic logic gets too much freedom in systems where a mistake is costly. This changes both architecture and the zone of responsibility: if an agent makes an error, it is still the company that will be responsible, not the model, for giving it access to the production environment.

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

• an agent can assemble a chain of actions across multiple corporate systems and gain more influence than a single employee;
• service accounts, long-lived API tokens, and rights to change key files and data quickly accumulate in a company;
• a new layer of shadow AI emerges when employees create their own assistants without architecture, support, and proper audit;
• when an employee is transferred to another department or after dismissal, "orphaned" agents remain, linked to their ID and permissions;
• neglected AI pilots and "zombie projects" continue running in the cloud and burning resources without a clear owner.

The authors draw a separate analogy with a toddler who was suddenly given too powerful a toy. For a corporate environment, the meaning is simple: you cannot release an autonomous agent to production without observability, the ability to quickly revoke access, and a mechanism for forced shutdown. You need discovery, audit trail, remediation, and a clear deactivation procedure, otherwise the benefit of automation disappears at the moment of the first incident, and analyzing the consequences turns out to be more expensive than the benefit of implementation.

The price of autonomy

Intel also disputes the popular idea that agentic AI is simply a way to cut the payroll fund. In the enterprise model, expenses behave differently: this is not a fixed license per user, but consumption of tokens, compute, and external APIs as workflows grow. Savings on people turn out to be too crude a metric, because alongside automation, costs for observability, support, security, and financial control of the entire agent infrastructure grow.

The article cites a December IDC survey commissioned by DataRobot: 96% of companies implementing generative AI and 92% of organizations implementing agentic AI reported that costs turned out to be higher or much higher than expected. This is an important signal for business: the problem doesn't boil down to the price per token. Money also goes to supporting many internal agents, fixing errors, reviewing access rights, and maintaining teams that must keep all of this under control.

The problem is compounded by the unpredictability of usage-based economics. Unlike classical FinOps, where cloud spending is more or less deterministic, agentic AI behaves probabilistically: long chains of calls, autonomous cycles, and planning errors can drive the cost of a single session to extreme values. The authors note that some AI-first founders are already facing expenses on the order of $100,000 for a single agent session.

If you don't set limits from the beginning, 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 companies that will win are not those that let employees "assemble their own agent" faster, but those that earlier embed access rights, audit, deactivation, budget limits, and ongoing control over what the agent does in real systems into their processes.