Skip to content
Guides16 Nov 2026

What is an AI agent

Felix Hellström

Felix Hellström

Founder, Klarhimmel

A square warm-white robot head with two navy eyes and a thin antenna topped by a pale-blue ball, on a navy base on a round sky-blue pedestal.

An AI agent is software built on an AI language model, the kind behind ChatGPT or Claude, that you give a goal instead of a script. It decides its own next step, uses tools to act in other systems (a booking system, an inbox, a calendar, a website), looks at what happened, and keeps going until the job is done or it needs a person.

What an agent does

A small square warm-white robot head with an antenna at the start of three stepping blocks that lead to a navy flag, a pale-blue ribbon tracing the path.

It starts from a goal. You say what you want done, such as "answer this guest" or "fix this page", not every step on the way, and it works out the steps itself.

The same robot head with three pale-blue ribbons running to a calendar block, an envelope and a small cabinet of drawers.

It uses tools. To act, it reaches into the systems you already have: it reads a booking, writes in a calendar, drafts an email, opens a file. Each tool is a door you choose to open for it.

The same robot head beside a navy front-desk bell, one pale-blue ribbon running from the robot to the bell.

It knows when to stop. When the answer isn't there, or the next step would cost money or can't be undone, it rings for a person instead of guessing.

The last part, knowing when to stop and ask, matters more than any definition. We run Klarhimmel with agents: they write and ship our code, keep our project board and draft our emails, and our agent platform answers guests for a hotel group. What follows is what an agent is, in plain words, and what we have seen running them every day.

Agent or chatbot: one question tells them apart

Ask: what can it do without asking you?

A chatbot answers. You type, it replies, and the conversation is the whole job. An agent also acts: it looks up a booking, writes a draft, changes a file, sends a message, and chooses which of those to do next. Putting a smarter AI model behind a chatbot does not make it an agent. Giving it tools and letting it decide the steps does.

The big labs draw the line in slightly different places. OpenAI's guide to building agents says agents are "systems that independently accomplish tasks on your behalf", and that "simple chatbots" are not agents. Anthropic separates workflows, where a model follows steps a person wrote, from agents, where models "dynamically direct their own processes and tool usage". Google Cloud puts it as software that uses AI "to pursue goals and complete tasks on behalf of users". We use the strict version: if a model chooses the steps, it is an agent.

What it looks like when agents run a company

Klarhimmel is one founder and a team of AI agents. As I write this, seven of them are at work, each on its own part of the company. In the two weeks from 3 to 16 November they wrote down more than a thousand decisions in our decision log, each with its reason, so a person can read why something changed.

They also got things wrong, and we make them write that down too. In the same two weeks the agents logged 950 times that a tool misled them or a safety check blocked them. Three ordinary examples:

  • A list of email drafts came back empty right after a refresh, when 36 drafts had been there a moment before. An agent that trusted it could have reported "nothing to send".
  • A code-review helper sat waiting for typed input that never came, and produced no review.
  • An agent could not close a finished task because the tool wanted a field nobody had told it existed.

None of these is dramatic. That is the point: an agent's mistakes are usually small, confident and easy to miss. So our rules require every agent to check its own work before it says done (a test, a screenshot, the live page), and a second agent to re-check any number that changes a plan.

Our hotel chat sits at the narrow end, on purpose

Ett Smart runs two hotels in the Stockholm area, and their guests ask questions in a chat on their phone. The chat runs on Klar Agent, our agent platform. From 29 September to 14 November guests asked it 130 questions. It answered 93 of them straight from the hotel's own text. The other 37 it passed on, because the answer wasn't in the text or the question needed a person.

It is built not to make an answer up: when the text doesn't cover a question, it asks the guest to leave it for the staff. Replies took about six seconds on average in October and about three in November. These counts include our own test chats, which we can't separate out, so read them as the shape of the work rather than a precise result.

By the strict definition, a chat that only answers from approved text is closer to a chatbot than an agent. We built it that way deliberately. A hotel's front desk is the wrong place to start with an AI that improvises. Its freedom grows one tool at a time, each with a person able to see and undo what it did.

What an AI agent can do for a small business

Good first jobs are the repeated ones with a clear answer, and the agent is only as good as the information you give it.

At a hotel, an agent answers guests' questions about check-in, parking or breakfast from the hotel's own information, and takes messages for the staff. Booking changes, complaints and anything unclear go to a person. It works beside the front desk, not instead of it. Before you pick one, check which channels it covers (a chat, the phone, email) and which languages it answers in.

At a shop, online or on a street, it answers questions about products, delivery and returns from your own product and policy pages. It can tell a customer where their order is only if it is connected to your shop system. Refunds, disputes and anything unusual go to you.

Anywhere, it can draft replies to email for you to send. And it should hand to a person anything that costs money or can't be undone.

Do you need one?

Only if you have a job that repeats. If you get the same twenty questions every week, or messages arrive at night when nobody is there, an agent can take that load. If you get a few questions a week and every one is different, a good FAQ page or a voicemail is enough. Start with the simplest thing that works, and add an agent only when that stops being enough. Anthropic gives the same advice to people who build agents: the simplest solution "might mean not building agentic systems at all".

What it costs

Ready-made agents for small businesses are typically priced from a few tens to a few hundred dollars a month. When we priced Klar Agent in October, the hotel and chat tools we compared ran from about 30 to 400 dollars or euros a month. A custom-built agent costs far more, as a one-off project plus upkeep. Either way the bill is made of the same parts: the subscription or build fee, the usage (the AI model's work, plus phone minutes or message volume) and any connections to your other systems. Most products include usage up to a limit. Before you sign, check what the price includes and what happens when you go past the limit. Klar Agent is 30 days free, then 990 kr a month excluding VAT, or 490 kr for a hotel with under 20 rooms.

Is it safe?

Safe enough when its permissions are narrow. The risks are real: it can give a wrong answer with confidence, it can leak information it was given, and it can take an action that can't be taken back. So:

  • Let it read only what it needs.
  • Never let it spend money or change records without a person's approval.
  • Make it hand over to a person when it is unsure.
  • Read its conversations.

Can it send email? Yes, if you give it access to an email account. Start with drafts you approve and send yourself, and let it send on its own only routine, low-risk messages. In our own company the agents draft the emails we write to customers, and a person presses send. Remember that marketing email and personal data are covered by law (in the EU, the GDPR) whoever writes the message.

How to know it gives the right answers

Test it before your customers do. Ask it the twenty questions you get most, then the awkward ones: the question it shouldn't answer, the one in another language, the angry one. Afterwards, read its conversations every week and fix each wrong answer at its source, the text it answers from, not in the chat. Keep a visible way for customers to reach a person. In Klar Agent, correcting an answer shows the old sentence and the new one side by side, and nothing reaches guests until you press Publish.

In one sentence

An AI agent is software that takes a goal, chooses its own steps and acts with tools, and the useful question for an owner is not how clever it is but what it may do without asking you.

See the narrow kind on your own house rules

Klar Agent answers your customers from the information you have approved, in their language, and hands the rest to you. 30 days free.