How do I control and undo what an AI agent does in my field service software?

Say a commercial maintenance company with 25 technicians runs quarterly filter changes on 60 sites. A dispatcher types: "Move all next week's filter changes for the north region to Thursday." The AI matches 38 visits. Twelve of them were already moved by the customer to a fixed date, and should not change.

With no controls, the AI moves all 38 and sends 38 customer texts. The office finds the problem when the first site manager calls. Someone rebuilds 12 visits by hand from memory, and sends 12 apology texts.

With approval only, the dispatcher sees one card saying "Move 38 visits to Thursday" and taps approve, because 38 looks about right. The texts go out. Approval did not help, because the card did not show which visits were affected.

With approval that shows before and after, plus undo, the card lists all 38 visits with the old and new dates. The dispatcher spots two locked dates and approves anyway. A day later the other 10 turn up. One undo puts all 38 back as they were, and the dispatcher reruns the request with "except fixed-date visits". Because customer texts were set to need approval separately, none had been sent yet.

The lesson: approval is only as good as what the approver can see, and undo covers what they miss. Keep sends to customers behind their own approval, so a reversed schedule change does not leave a trail of wrong texts.

The EU AI Act's human oversight article (Article 14) applies to high-risk AI systems, not to every AI feature. Where it applies, people overseeing the system must be able to disregard, override or reverse its output and to stop it with a "stop" button or a similar procedure. Most job scheduling and customer texting will not be high risk, but uses that touch workers' management or safety components can be, so check with your adviser. The page lists application dates of 2 December 2027 and 2 August 2028 for the two groups of high-risk systems.

Outside the EU there is no equivalent rule for field service AI, but the same ideas appear in voluntary guidance. NIST's Generative AI Profile suggests comparing outputs with known ground truth through human oversight and automated checks, and recording who holds human oversight roles.