[SAMPLE] The kill switch problem: who can stop an agent once it has your data?

SAMPLE ITEM for template preview. Replace or delete before launch.

The question boards are asking

"If an agent is acting on our data right now, who can stop it, and how fast?" In our conversations with [N] enterprises, the honest answers fell into three buckets.

• The vendor can. The stop button belongs to someone else.
• Someone can, eventually. There is a runbook.
• The access layer can. One enforcement point, closed in milliseconds.

The first group is larger than most vendors would like to admit. The second group believes it has a kill switch, and in a sense it does; it just takes forty minutes and four people. Only the third group can answer the board's question in one sentence.

What the data shows

Two things stand out. First, the "don't know" group is not small companies. Second, the enterprises with a single enforcement point were running the most agents in production.

That second finding runs against the intuition that control slows adoption. In the data it is the reverse: the teams with the clearest stop button were the ones comfortable letting agents do more, because the downside of a mistake was bounded in time.

"We stopped asking whether the model was safe and started asking whether the plumbing was." - [Name], [Title], [Company]

Why the model is the wrong place to stop

A model has no conscience, just code. It keeps producing actions for as long as it is given context and permission, and both of those are properties of the data layer.

Prompt-level guardrails are worth having, but they are a request, not an enforcement. Enforcement lives where the agent's actions become real: at the gateway between the agent and the systems it touches.

What "correct by design" looks like in practice

The teams that answered fastest shared three traits: every agent action is a message that can be audited, every data source is reached through one gateway, and autonomy is granted in levels and earned, not assumed.

None of this requires a new model or a new vendor relationship. It requires deciding, once, where the stop button lives, and building the plumbing so that decision is true for every agent, not just the ones a particular team happens to run.

What to do this quarter

Inventory the agents already acting on production data. For each one, write down who can stop it and how long that takes. If the answer is a runbook, that is the gap. Then pick the gateway you will route them through and make the kill switch a property of the gateway rather than of each agent.

The board's question is not about model quality; it is about control. Answer it in one sentence and the rest of the AI conversation gets easier.