AI agents for maintenance companies: what they do, what they don’t and how to test them

An AI agent is a program that doesn’t just answer: it does an entire task, from start to finish, inside your systems. At a maintenance, HVAC or installation company, that task is usually the office work around each breakdown: taking the service request, asking for what’s missing, opening the work order in the ERP, suggesting the technician and keeping the customer posted. Here we explain what one really does, where it has to stop, who’s already using one and how to test it without putting a single customer at risk.

What an AI agent is, and how it differs from a chatbot

A chatbot answers questions: it has a script or a document and replies with what it finds. When the conversation ends, someone has to take what the customer said and do something with it.

An AI agent does that something. It understands what the person is asking for, decides which steps are needed and carries them out using the tools you’ve given it: your ERP, the technicians’ calendar, email or WhatsApp. The practical difference shows at the end of the conversation. With a chatbot, you get a transcript; with an agent, an open work order, with the customer recognized and the technician suggested.

That’s why the terms you hear in the market, AI receptionist, AI voice agent or operations agent, don’t mean the same thing. The table sums up the difference.

Where work gets lost today

Before talking about agents, it’s worth looking at what goes wrong. We read 445 one- and two-star Google reviews of 111 maintenance and installation companies in Madrid by hand, and classified each complaint by its cause.

Most complaints aren’t about the technical work but about what happens around it: nobody answers the phone, nobody says when the technician will arrive and nobody warns about delays. When customers give numbers, they say they called 5 times on average and waited 27 days on average. One in four waited a month or more.

That’s exactly the work that fits an agent best. It’s repetitive, it follows rules the company already knows and it has to be done at any hour, including when the office is closed.

What an AI agent does at a maintenance company

Take a real breakdown: a property manager writes at ten o’clock on a Tuesday night because the chiller in a residential building is showing an alarm. This is what the agent does, and who does each step.

• It takes the service request wherever it comes from. On the phone it acts as a voice agent; on WhatsApp or email, it reads the message. If something’s missing, it asks on the same channel.
• It recognizes the customer and the equipment. It looks up the building, the address and the chiller in your ERP, and if the customer doesn’t exist, it doesn’t make one up: it asks and flags it.
• It opens the work order with the breakdown described, the urgency and who will open the machine room. At first, a person reviews it before it goes into the ERP.
• It suggests who goes based on your rules: on-call schedules, zones and who knows that installation.
• It keeps the customer posted that the request is logged and when someone will come, and writes again if the time changes.

The same applies to the work that falls through the cracks between visits: quotes nobody follows up on, maintenance contracts that expire and unpaid invoices. The agent drafts each message and a person approves it before it goes out.

Who’s already using them

AI agents are already in production at field service companies, though almost always at large companies or outside Spain. Here are four public cases:

The pattern is the same in all of them: the agent doesn’t replace field service, but the part of the office that consists of collecting information, opening work orders and keeping people posted. And all of them leave a way out to a person.

What a person should keep doing

A good agent knows where to stop. In installations with gas, electricity or water, a mistake isn’t just a badly opened work order, and people’s safety is still the company’s responsibility. The rule we follow is simple: anything irreversible or risky is decided by a person.

There’s also something it should never do: score technicians. An agent can recommend who’s a good fit for a specific job, but keeping a success rate next to each name is something else, and it doesn’t help fix breakdowns.

What the law requires since August 2026

Two regulations apply to any AI agent that talks to your customers:

• It has to say it’s an AI. Since August 2, 2026, Article 50 of the EU Artificial Intelligence Act applies: anyone talking to an AI system has to know it, unless it’s obvious. On a call, that means saying so at the start, not hiding it in the terms and conditions.
• Your customers’ data is still yours. The provider processes personal data on your behalf, so you need a data processing agreement, as Article 28 of the GDPR requires. Also ask where it’s stored: the sensible answer is that it doesn’t leave the European Union.

How to test an AI agent without risk

The safe way to start isn’t to switch it on and see what happens. It’s to test it with your own service requests before it touches anything, and to switch it on in stages. We do it in four phases:

1. Discover. How service requests come in today, what gets asked, who’s on call and what must never be done.
2. Set up. The questions, the rules and the connection to your ERP, using what you already have in writing: nobody starts from scratch.
3. Test. The agent receives the real service requests but doesn’t write anywhere. Each work order it would have opened is compared with the one your office opened.
4. Go live. First with approval of every work order; after that, only what the company decides.

During the test, two numbers are enough to decide:

• How many work orders the office accepts without changes. That’s the measure of whether the agent saves work or creates it.
• How many serious errors there are. For example, a work order opened for the wrong customer or site, or for something that wasn’t a breakdown. With even one in the final week, it doesn’t go live.

Eight questions for any vendor

If you’re looking at AI agents, these questions quickly separate what works from what just demos well:

In short

A useful AI agent in maintenance isn’t just another chatbot: it takes the service request at any hour, opens the work order in your ERP and suggests the technician. Anything risky or irreversible is still decided by a person.

Before it goes live, it’s tested with your real service requests, measuring how many work orders would come out right without changes.

If you’re curious how the loop closes after the visit, we cover it in what a service report should include. And if you want to know how we’re starting, it’s in what Delegic is and where we’re starting.