Every field service vendor now says it has AI. For an operations head running hundreds of sites, the useful question isn't whether software has AI in it. It's which parts of the day's work a machine can take over today without someone checking behind it, and which parts still need a person.
This post answers that plainly. It's written for people who run ATM networks, branch estates, guarded sites and facility teams, where most of the effort between a fault and a closed ticket goes on follow-up rather than repair.
What an AI agent is, in field service terms
A chatbot answers questions. An agent takes steps towards a goal, inside rules you set: it notices something, decides what to do next, does it, and records what it did. In field service that usually means the small, repetitive actions that sit between systems and people. Asking an engineer for an ETA. Escalating when nobody replies. Filing a vendor's email against the right ticket. Checking that a closure has the photos it needs.
None of these is hard. Together they eat a coordinator's day, and they are exactly the steps that get skipped when the team is busy, which is when they matter most.
Where agents already work well
Agents are good at work that is frequent, rule-based and easy to check afterwards. In field operations that covers a lot:
- Chasing ETAs and escalating. As soon as a job is assigned, ask the engineer or vendor for an arrival time on the channel they actually use, such as WhatsApp or SMS. If they go quiet, retry, then escalate up the ladder you defined: a call, then the supervisor, then the zone head.
- Turning events into clean tickets. Device alerts, QR reports from site staff and calls from a client's CRM arrive in different shapes. An agent can merge duplicates, classify the call type and start the right SLA clock, so coordinators start from a clean queue.
- Keeping the record complete. Vendor replies that come back by email can be logged against the right ticket automatically, so the history doesn't depend on someone copying messages across.
- Drafting tomorrow's plan. Weighing SLA, fault type, site importance, spares and travel time to draft each engineer's day is a good fit. A coordinator reviews and adjusts the draft instead of building it from nothing.
- Guiding the repair. Showing the engineer the steps for that fault on that model helps first-time fixes, especially when the nearest engineer isn't the most experienced one.
- Checking proof before closure. Are the required photos there, taken in the app at the site? Is the checklist complete? Was the spare scanned? An agent can send a gap back to the engineer now, rather than an auditor finding it next month.
Where people must stay in charge
Agents are weak where the decision is a judgement call, the cost of a mistake is high, or the situation is new. Keep a person in charge of:
- Money. Approving a quote outside the contract, authorising overtime or replacing an expensive unit. An agent can prepare the request; a person approves it.
- Safety and security. A fire panel fault, a tamper alarm or a guard missing from a sensitive post needs someone to decide, not a script.
- Client relationships. A disputed invoice or an unhappy bank should hear from a person, with the evidence the system has gathered.
- Unfamiliar problems. A fault nobody has seen before, or a site with unusual access rules, is outside what any agent has learned from.
- Accountability. Someone has to own each outcome. An agent can do the chasing, but a named person is responsible for the result.
The repair itself, of course, still needs a technician on site. What changes is how much of their day goes on phone calls, and how much on the work.
The rules that make agents safe to use
The difference between a helpful agent and a risky one is rarely the AI model. It's the controls around it:
- Per-step settings. For each kind of action, choose whether it runs automatically, asks first or stays off, and set this per client, zone and fault type. A bank client with strict rules can have more "ask first" steps than an internal facility.
- Escalation ladders people define. The agent follows the ladder; it doesn't invent one.
- A full audit trail. Every action is written to the ticket: what was sent, to whom, when, and what came back. If you can't see it, you can't trust it.
- A clear handover. When the agent is unsure or a rule says so, it hands the ticket to a named person with the context attached.
How to start without risk
Don't switch on everything at once. Start with one step that is frequent, low risk and easy to measure. ETA chasing is a good first choice almost everywhere.
- Note how it works today: how many calls a coordinator makes per ticket to get an ETA, and how long it takes.
- Turn the agent on for that one step, in one zone or for one client, with escalation to the people who handle it now.
- Read the agent's log every week with the coordinators. Look for cases where it escalated too early, too late or to the wrong person, and adjust the rules.
- When the team trusts it, add the next step, such as logging vendor replies or checking proof before closure.
Coordinators usually become the strongest supporters once the chasing leaves their day. Their job moves from asking for updates to deciding what to do with them.
Questions to ask any vendor
- Can I see every action the agent took on a ticket, and why?
- Can I set automatic, ask-first or off per step, per client and per zone?
- What happens when the agent isn't sure?
- Does it work on the channels my engineers and vendors actually use?
- Can I start with one step in one zone, and turn it off instantly?
Where Fieldor fits
Fieldor's agents are built for this follow-up work. ChasePilot asks for ETAs and climbs the escalation ladder you configure. Planner drafts tomorrow's plan for coordinators to review. FixMate guides the repair on site. EmailPilot logs vendor and partner replies against the right ticket. You choose which steps run automatically, which need approval and who gets escalations, per client, zone and fault, and every action is recorded on the ticket. See how the agents work, or book a walkthrough on your own sites.