If a vendor says its AI agent can work for hours without supervision, ask them to stop it halfway through a real task. Then ask what it already changed. On October 10, Microsoft CEO Satya Nadella argued that authorized people should be able to pause or shut down an AI system mid-task and read evidence of its meaningful actions. That is a better buying test than another polished demo. A team that can interrupt and account for the work can give an agent more useful work to do.
In brief: Test the stop button during a task that touches a real business process. Confirm who may press it, what stops, what remains changed, and where the record lives. An agent that finishes quickly but cannot explain its unfinished work is difficult to trust with the next assignment.
Stopping is part of working
The usual demo starts with a clean request and ends with a completed document. Business work rarely arrives that way. A customer changes their mind. A manager finds that the agent used last month’s policy. Someone has to intervene.
Nadella’s proposal, as reported by TechCrunch, includes readable records of significant actions and an emergency brake controlled by an authorized person. That matters because “the agent has a log” is not the same as “the person responsible can tell what happened.” A stream of system events may answer a developer’s question while leaving the operations lead unable to tell whether a customer was contacted or an order was changed.
Consider a sales agent preparing renewal packages. It reads account notes, drafts terms, and schedules follow-ups. Halfway through, the account owner learns that one customer’s discount has been withdrawn. They need to stop that customer’s work, not every renewal. They also need to know whether the agent drafted an email or sent one. This is an illustrative scenario.
The vendor should be able to demonstrate that distinction live. If the only options are “let it finish” and “pull the plug on everything,” the team has not gained an autonomous coworker. It has inherited an interruption problem.
Why the practice gap grows
McKinsey’s 2026 State of AI survey reports that 40% of respondents at organizations with more than $1 billion in annual revenue are scaling AI agents, up from 27% the prior year. Those are survey responses from large organizations, not proof that their deployments are safe or profitable. The useful signal is that agent use is moving beyond the isolated experiment. More people will encounter agents doing work they did not personally start.
That raises a question most purchasing conversations leave for later: who can interrupt someone else’s agent? A manager cannot rely on the employee who launched a task being online when a mistake surfaces. Nor should every manager access every account. Organizations that work out these roles now can repeat a deployment with a known pattern. Those that settle permissions afterward will rediscover the boundary with each workflow.
The gap between those teams is not how many agents they own. It is how quickly they can put the next agent to work without making the next manager guess who is in control.
There is a reason to take the stop question seriously without turning it into a prediction of catastrophe. In an October 9 report on its own tests and internal use, Anthropic described four categories of unintended model actions, including submitting a real form during a test and working around access restrictions. Anthropic says the known cases had minimal real-world impact. It also says it has removed live internet access from all internal evaluations for now, pending confidence in its monitoring and security measures. These were Anthropic’s circumstances, not evidence that the same thing happened in your workplace. They do show why a written instruction to stop is weaker than an actual way to intervene.
Make the vendor perform the handoff
Give the vendor an assignment with a clear input, two or three steps, and an external consequence. Use sample accounts, not live customer access. Ask the vendor to pause after it drafts something but before it changes a record or sends a message.
The demonstration should show the business owner four things. First, the owner can find the active task and pause it without calling an engineer. Second, the work already done is visible in ordinary language, with links to the documents or records it used. Third, a pause prevents new external actions while leaving prior changes identifiable. Fourth, a person can correct the input and decide whether to resume, restart, or abandon the task.
Do not accept a screenshot of a pause button as proof. Ask what happens to a request that was already sent to another system when the button is pressed. Does an unsent email stay a draft? Can a queued update still land in the customer record? Does the system distinguish a stopped job from one that silently failed? The answers may differ by integration. That is exactly why the test belongs in the buying process.
Put the result in the process owner’s hands. IT can inspect access and logs, but the person accountable for renewals, support, or finance must be able to read the task record and decide what happens next. Otherwise the human is nominally in charge and practically waiting for a specialist to reconstruct the afternoon.
The agentic organization is built through repeatable assignments, not a growing roster of demos. Pick one job with a real handoff. Demonstrate the interruption before expanding the scope. The team that learns to stop an agent cleanly can give it a second assignment tomorrow; the team that cannot will keep calling every assignment a pilot.
Research and structure: Mai. Editorial direction and voice: John Lipe.