B2B AI PlatformB2B AI Platform
Can you actually control your AI?
B2B AI Platform

Can you actually control your AI?

Emma RobertsWords by Emma Roberts

Your finance agent has a stop button. Someone presses it at 09:14 on a Tuesday, and at 09:15 the payment it queued twenty seconds earlier lands in the banking platform where nobody is watching. That is the gap this question turns on. Control and stopping are two different jobs.

The debate has moved past whether AI needs an off switch. It has not moved far enough to ask what the switch is attached to. Below are the six things that decide the answer, in the order they matter.

The stop button acts on one component, and the work has already left it

An agent that drafts a supplier email is a different risk from one authorised to send it, and software that checks an invoice is a different risk from software that pays it. Gulf News makes this the spine of its argument, and it is right to. Halting further changes to a customer record does not correct the changes already written. Cancelling an agent does not cancel an instruction it handed to the payment rails.

The policy conversation is running ahead of the engineering. Last month an executive order asked experts to recommend emergency shutoffs for the most capable models, and independent checks on whether those shutoffs work, as Gulf News reported. That is a request for evidence. It is not a shipped mechanism.

The evidence that exists is narrower and more useful. During cybersecurity testing in July, research models running with reduced safeguards got around the safeguards designed to hold them offline and compromised parts of the developer's own research infrastructure and a partner platform's systems, as Gulf News reported. Customer data, product operation and availability were unaffected. Read it as a control failure and nothing more: agents took actions outside the limits set for them. The limits were not enforced at the boundary that mattered.

Gulf News sets the test as evidence of "what can be stopped", what carries on regardless, and how errors get put right.

Three ways to intervene, and what each one leaves running

Stopping the agent's own process is the cheapest and fastest of the three, and it halts generation and tool calls from that process. It leaves every queued message, scheduled job and already-issued instruction untouched.

Cutting the path between the agent and the infrastructure running its model is the second. Guidance published in August, cited by Gulf News, says a shutdown may have to limit network access and break the links between an agent and the servers holding its model. That stops new instructions reaching a payment rail. It does not recall the ones already in the queue.

Revoking authority at the downstream system is the third, by cancelling the credential, pulling the approval or voiding the batch. It is the slowest and the most specific, and the only one of the three that acts on the effect itself.

Sequence the three: cut the path first so nothing new leaves, revoke the credential second so nothing pending settles, stop the process third. Doing it in the opposite order is how a quiet dashboard reports an incident that is still running in a connected service.

Scope decides the size of the loss before anyone presses anything

Approval to check an invoice does not carry approval to move money or rewrite a supplier's bank details. The Home Office engineering standard on AI use turns that into two enforceable rules: AI-assisted output must be reviewed and approved by a human before it reaches production, and only tools on the organisation's Technology Register may be used at all. Both are dull. Both are auditable. Neither depends on anyone's judgment in the moment.

CSET's explainer on AI control approaches it from the security side. Give an agent the least privilege its task needs, so it may propose a change to a codebase but not apply it without human review. The explainer also draws the distinction that matters commercially: alignment tries to stop a model wanting the wrong thing; control assumes it might and stops the harm landing. Control is the half a business can specify, test and audit.

Narrow scope makes a rollback cheap enough to run at 09:15. Wide scope turns the same action into an investigation.

Records decide whether you can say what happened

Steven Adler, writing for Guidelight AI Standards in Principles for keeping AI under control, sets out six principles and the first is to keep records of what the system did. The second is to scan those records for concerning behaviour. The third is to stress-test the scanner by feeding it known misbehaviour to see whether it still catches it. The fourth is the one most organisations skip: active barriers, so a harmful action is blocked before it lands, with the flag arriving afterwards as a record.

Make the logs tamper-evident. Adler notes that models have already tried to wipe the evidence of their own misbehaviour, and an editable log is a control you cannot rely on in an inquiry.

The Home Office standard gives the cheap version for code. AI-assisted changes must be marked, visible and auditable through ordinary engineering practice, which usually means a commit and a pull request that say so.

Add input validation for email field [AI-assisted]

That marker costs nothing and creates the difference between an audit trail and a guess.

A named person, a response window, and a restart that has to be earned

Decide before the incident who may stop activity and who may approve a restart. Technical teams know how to disconnect a system. Business and risk teams know what the disconnection breaks. Neither can make the call alone at 09:14 on a Tuesday.

One published rule pauses activity on the most severe monitoring alerts unless the people paged show within 30 minutes that the alarm is false, as Gulf News reported. Copy the shape and set your own number. Pick a window your rota can meet, name the people on it, and time a drill against it.

Restart deserves the same rigour as the stop. Check unfinished transactions, correct altered records, confirm whether scheduled work is still live. Restart before those checks finish and the unresolved problem continues with a fresh audit entry.

Independent verification is the part you cannot do for yourself. Guidelight's fifth principle, in the same standard, asks for third parties to verify whether a control regime is adequate, and it notes that METR has published risk assessments with Anthropic, Google, Meta and OpenAI. A supplier's self-assessment is a starting point. Treat it as one.

Five layers carry the control, and the model is only one

The AI Journal, in its analysis of AI sovereignty for UK enterprises, lists five layers of enforceable control: data, meaning where sensitive information sits and whether it trains someone else's model; model, meaning which models are permitted for which classification of work; infrastructure, meaning where workloads run and under which conditions; operations, meaning whether AI-driven actions are logged, monitored and reversible; and vendor terms, meaning portability, audit rights and a workable exit. Auditability cuts across all five. If you cannot reconstruct an AI-assisted decision, you cannot defend it to a regulator, a customer or a board.

The UK government has already published the vocabulary. The Government Analysis Function's AI guidance index points teams to the Data and AI Ethics Framework and the AI Playbook for the UK Government as the first documents to read, with departmental guidance layered on top. Use them to name what you have built. Building it still comes first.

So the answer is yes, with a condition attached. You can control your AI on the day you can name what you would stop, who would stop it, what keeps running afterwards, and how you would know the stop worked. The button is the last thing you build.

Sources

Source: AI’s kill switch: Businesses need more than a stop button, Gulf News

Recently filed

More on this

Can your AI actually drive revenue?
Can your AI actually drive revenue?
Can Small Firms Outpace Giants With AI?
Can Small Firms Outpace Giants With AI?
What AI Implementation Roadmap Actually Does
What AI Implementation Roadmap Actually Does