Microsoft President: AI Is Regulated Less Than Milk and Washing Machines

Microsoft President Brad Smith said that artificial intelligence is currently regulated less than milk and washing machines. The Guardian reported his remarks on 12 October 2026. Smith proposed two specific mechanisms: emergency "kill switches" for AI systems and a legal duty to the "victim" if a rogue AI agent breaches their system.

The source does not include the full text of his speech, so the details of these mechanisms are not yet known. But the way he framed the issue is telling.

Why milk

The comparison is well chosen. Milk goes through processing standards, quality control and labeling. A washing machine gets electrical safety certification before it reaches store shelves. An ordinary buyer has a clear set of guarantees: if a product is dangerous, there is a standard to point to and a company to hold accountable.

AI agents have no such familiar framework. A model with access to email, a cloud console or a corporate database can take actions with real-world consequences. Yet in most jurisdictions, the requirements for testing agents, shutting them down and investigating incidents remain fragmentary. The EU has the AI Act, adopted in 2024, but it is built around risk categories rather than agentic scenarios.

What Smith proposes

Two points emerge from the available information:

• Kill switches. The technical and organizational ability to quickly stop a system if it behaves dangerously.
• A duty to the victim. If an AI agent breaches someone else's system, the injured party should have a legal right to demand protection or compensation.
• A general principle. AI safety requirements should be comparable to those for ordinary consumer goods.

The "kill switches" deserve a separate note. The idea is not new: in 2023, Smith already spoke about "safety brakes" for AI that controls critical infrastructure. Now the emphasis is shifting to agents that act autonomously and can harm third parties.

The question without an answer

The main difficulty is the chain of responsibility. The developer of the base model, the platform provider, the company that configured the agent, or the end user could each be held responsible for an agent's actions. Until lawmakers define how to divide that responsibility, every party will try to shift the risk onto another. License terms already try to relieve vendors of most claims.

There is also a commercial context. Microsoft sells agentic products, including Copilot and tools for building enterprise agents. Clear rules favor large players: they have lawyers, auditors and compliance resources, while the cost of entry rises for small teams. This does not make Smith wrong that the regulatory vacuum is real, but it explains why such appeals should be read alongside the speaker's interests.

What teams can do right now

There is no need to wait for legislation. Companies that deploy agents would do well to take a few steps in advance:

• limit the agent's permissions under the principle of least privilege and do not give it access to everything;
• plan for a manual shutdown: who will disable the agent in an incident, how, and within how many minutes;
• keep an action log so the chain of events can be reconstructed;
• review contracts with AI vendors: who is liable for damage to third parties;
• ask your insurer whether the policy covers incidents involving autonomous agents.

If Smith's proposals become rules, companies with such processes in place will adapt faster than the rest.