What an AI agent is and what you can delegate to it
An AI agent is a program you delegate a task to: it has a goal, some tools and some rules, decides which steps are needed, carries them out and checks the result. Anything that can’t be undone, it checks with you first. It’s more like delegating to a new hire than using a piece of software: you tell it what it has to achieve, what it has to work with, what it must never do and what it has to ask you about. Here we explain how it works under the hood, what it does with a real case, what you need to give it to delegate well and how to tell a real agent from a chatbot with a new name.
How an AI agent works under the hood
A chatbot receives a question and returns an answer. An agent receives a goal and works until it achieves it, in a loop that repeats as many times as needed:
- Goal. What it has to achieve: “log this service request and suggest who goes,” not “reply to this message.”
- Plan. What it’s missing to get there. If the customer hasn’t given the address, ask for it; if they don’t know what equipment it is, look it up.
- Act. It uses the tools you’ve given it: searching your management software, writing an email, opening a work order, checking a calendar. Without tools, an agent is just a chatbot that reasons.
- Check. It looks at whether what it did adds up: that the customer exists, that the work order was created, that no required field is missing. If something doesn’t add up, it goes back to the plan.
- Close or ask permission. If the next step can’t be undone, like sending a quote or assigning a technician, it leaves it ready for a person to approve.
An example from start to finish
At ten o’clock on a Tuesday night, the manager of a residential building messages their maintenance company on WhatsApp. On the left, what they see; on the right, what the agent does in the meantime.
| Time | In the conversation | In your systems |
|---|---|---|
| 10:04 p.m. | Customer: “The boiler in building 3 won’t turn on” | Looks up the building and the boiler in the management software: finds them, with a maintenance contract |
| 10:05 p.m. | Agent: “Is there a code on the boiler’s display?” | It’s missing one detail to judge the urgency, so it asks |
| 10:06 p.m. | Customer: “It says E04” | Checks the equipment documentation for what that code means |
| 10:07 p.m. | Agent: “It’s logged. First thing tomorrow we’ll confirm who’s coming” | Opens the work order with the fault, the code and the contact, pending approval |
| 8:00 a.m. | — | Suggests the technician who covers the area and knows that boiler; the office approves with one click |
| 8:02 a.m. | Agent: “Marc will be there between 9 and 11” | Lets the customer know once the office has approved, not before |
Chatbot, assistant, automation and agent
The four words get used as if they meant the same thing, and they don’t. The question that separates them is simple: who does the work, and what happens when something doesn’t go as planned?
| What it is | What it does | Where it falls short |
|---|---|---|
| Chatbot | Answers with a script or with your documents | It doesn’t act: a person does the work afterward |
| Assistant or copilot | Helps a person while they work: summarizes, drafts, searches | Without the person there, nothing happens |
| Traditional automation | Repeats fixed steps: “if an email arrives with this word, move it here” | If the case changes a little, it breaks or gets it wrong |
| Agent | Understands the case, decides the steps, acts in your systems and checks the result | It needs clear rules and testing with real cases before it’s let loose |
At a field service company, the three types sold most often are the chatbot, the AI receptionist and the operations agent. We compare them in AI agents for maintenance companies.
What agentic AI is
Agentic AI is the name given to artificial intelligence systems that act as agents: instead of generating a piece of text or an image and stopping, they chain steps together, use tools and pursue a goal. It’s not a different technology, but a different way of using the same language models: with a goal, with access to tools and with the ability to check what they do.
What makes it useful now is that these models make far fewer mistakes than they did a couple of years ago when choosing the next step and using a tool. That makes it possible to delegate multi-step tasks to them that used to have to be broken into pieces and supervised one by one. It doesn’t make them infallible: that’s why the check step and permission for anything irreversible aren’t optional.
How to delegate well to an agent
An agent performs only as well as the brief it gets. Just as you wouldn’t tell a new hire “take care of the service requests” and leave them on their own, you need to give an agent five things:
| What to give it | The question it answers | Example |
|---|---|---|
| A goal | When is it done? | The work order open, with every detail, and the customer notified |
| What the company knows | What does it need to know so it doesn’t make things up? | Service areas, hours, services, what counts as urgent, each customer’s equipment |
| Tools | What can it read and what can it write? | It reads customers and contracts; it writes work orders, but not invoices |
| Rules | What does it always do and what does it never do? | It always says it’s an AI; it never promises a time without approval |
| When to check in | What has to go to a person? | Risk to people, a customer it doesn’t recognize or someone asking to talk to a person |
How much to let it do: the levels of autonomy
Delegating isn’t all or nothing. An agent, like a person who has just joined, gets autonomy little by little, as it shows it gets things right. We think of it in four levels:
The way to move up a level is to measure. While it’s at level 2, every case a person approves without changes counts in its favor; every serious error counts against it. When, for weeks, almost everything comes out right the first time and there are no serious errors, that type of task can move up to level 3.
Which tasks make sense to delegate to an agent
A good candidate if it meets almost all of these
- It happens a lot and almost always the same way: logging service requests, answering the same questions, following up on quotes or invoices.
- It follows rules that already exist, even if they’re in someone’s head: what’s urgent, who gets called, what information is needed.
- The data is in your systems: the management software, email or a spreadsheet, and the agent can read it.
- You can tell when it worked: the work order is open and properly filled in, the customer has replied, the invoice has been paid.
- It can be reviewed before it takes effect, or undone if it goes wrong.
Better not to delegate it, or not completely
- Anything that puts people at risk. An agent can spot that a service request is serious and hand it off immediately, but safety instructions come from a person.
- Anything without clear rules or a way to check whether it worked.
- Judging people. Suggesting who’s a good fit for a specific job, yes; scoring employees, no.
- Anything irreversible without approval, until the agent has shown for weeks that it gets that type of case right.
Five mistakes when delegating to an agent
- Handing it everything at once. Start with one task, with approval, and expand once it works.
- Not writing down what “everyone knows.” That the care home customer is always urgent, or that there are no afternoon appointments on Fridays. If it isn’t written down, the agent doesn’t know it.
- Testing it with made-up examples. Tests are done with your real cases from the last few weeks, typos and half-finished messages included.
- Measuring whether it sounds good instead of whether it gets it right. What matters is how many cases come out right without changes, not how natural it sounds.
- Not deciding up front what counts as a serious error. A work order opened for the wrong customer isn’t the same as an extra comma. The list of serious errors is written before you start.
How to spot a chatbot with a new name
With “agent” the word of the moment, plenty of products use the name without being one. Five questions settle it during a demo:
| Question | If it’s an agent |
|---|---|
| When the conversation ends, what’s been done? | Something in your system: a work order, an email sent, a record |
| Does it write to your software or to its own? | To yours, without typing anything twice |
| If it’s missing a detail, what does it do? | It asks or looks it up; it doesn’t make it up |
| How does it know it did it right? | It checks the result and flags anything that doesn’t add up |
| What can’t it do without asking? | A specific list, decided by you |
What the law says
- 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.
- The data is still yours. If the agent processes your customers’ data, the provider does so on your behalf and you need a data processing agreement (Article 28 of the GDPR).
Frequently asked questions
Does an AI agent replace a person?
It replaces part of their work: the repetitive part, done the same way every time and at any hour. The person keeps what requires judgment, what carries risk and approving what the agent leaves ready. In practice, what changes is that the person stops typing and starts reviewing.
Does it make mistakes?
Yes, like anyone. That’s why a good agent checks what it does, says when it doesn’t know something instead of making it up, and leaves anything irreversible for a person to approve. And that’s why it’s tested with real cases before it goes live.
Do you need to know how to code to use one?
To use it, no: you talk to it in plain language and review what it leaves done on a screen. Setting it up properly does take technical work, especially connecting it to your management software, and that’s usually done by the vendor.
How long does it take to get it running?
It depends mainly on what has to be connected. What you shouldn’t skip is the test: we let the agent work for four weeks with the company’s real cases, touching nothing, before switching it on.
In short
An AI agent doesn’t just answer: it does an entire task with your tools, checks the result and asks you before anything that can’t be undone.
You delegate to it like you would to a new hire: with a clear goal, what the company knows in writing, a set of rules, and autonomy that grows as it shows it gets things right.
If you work in maintenance, HVAC or installation, in AI agents for maintenance companies we show what one does with a real service request, who’s already using them and eight questions for any vendor. And in what Delegic is, where we’re starting.
Sources
About the author
Joaquín TenésCo-founder of Delegic
