Almost every software supplier now calls its product an AI agent. For a business of five or twenty people it comes down to something practical: software that reads an incoming question, looks up what is needed and does something with it.
What is an AI agent?
An AI agent is software that handles a task by itself using a language model, the kind of AI that reads and writes text. It is given a goal, a few tools and limits. Which steps it takes to reach that goal, it decides case by case.
The tools are the things it is allowed to do, such as checking the schedule, opening the calendar or preparing an email. If a customer asks whether next week's booking can still be changed, the agent finds the booking, checks whether the change fits and writes a reply. The next message asks something else, and then it takes a different route.
How does it differ from a chatbot and from an automation?
An old-style chatbot follows a script with buttons and gets stuck on a question that is not in it. An automation carries out fixed steps in a fixed order. An agent understands a question in ordinary language and chooses which step is needed.
That freedom comes at a cost. An automation does exactly the same thing every time and costs next to nothing to run. A paid order that has to become an invoice therefore belongs in a fixed workflow, which is what we build under business process automation. An agent earns its place where the same question comes in ten different ways. Often it ends up as a combination, with a workflow for the fixed work and an agent at the point where something has to be read.
An agent does not have to be a chat window either. It can just as well work only in your inbox.
What can an AI agent do for a small business?
It can answer customers from your own documents, assess enquiries and book them in, or prepare a draft quote. Reading an incoming PDF and putting the details into the right system is work for an agent too. Each time it is work where something has to be read and understood first.
There are more examples on the page about AI agents for business.
A worked example: airport transfers
This is a worked example, not a client of ours. A company that runs transfers between the airport and the hotels in the south of the island gets 40 messages a day. Most of them ask the same things: what the trip to a given hotel costs, whether there is a child seat, what happens if the flight is late. They come in Spanish, English and German, seven days a week. In a 30-day month that makes 1,200 messages. At four minutes a message, someone spends 80 hours a month on them.
The agent reads each message, checks the fares per zone and the drivers' schedule and writes the reply in the customer's language. Suppose it deals with 60 per cent by itself. That is 720 messages. The other 480 go to a person with the draft already written, who spends two minutes on each on average. That is 16 hours. Someone also reads through the log for ten minutes every day, 5 hours a month in all. That leaves 21 hours of work, so the agent saves 59 hours a month.
If an hour of office work costs the company €15, which is an assumption and every business has its own figure, that is €885 a month, about €10,600 a year. From that you take off the use of the AI model, which in this example stays far below that figure, and the cost of building the agent, which depends on what it has to do and which you know upfront. If the same company gets 20 messages a day, the agent saves 27 hours a month, about €405 on the same assumption, and the sum is a good deal tighter.
What does an AI agent cost?
An AI agent has two costs. Building it is paid once, and use of the AI model is paid to the provider every month, depending on volume. At CanarySites you get a fixed price upfront once we know exactly what needs to be built, and we estimate the model use before we start.
| Item | How much | When you pay |
|---|---|---|
| Building the agent | a fixed price upfront, once we know what needs to be built | once |
| Use of the AI model | depends on volume; in the example, $60 with Claude Sonnet 5.5 | every month, to the model provider |
| Checking by a member of staff | in the example, 21 hours | every month |
The model provider charges per token, a fragment of a word, and counts what goes in separately from what the model writes back. Anthropic, one of the large providers, charges $2 per million input tokens and $10 per million output tokens for its Claude Sonnet 5.5 model. Those are the prices it published on 16 November 2026.
Take the transfer company again and suppose one message, with the instructions, the data looked up and a few rounds back and forth, uses 20,000 input tokens and 1,000 output tokens. With 1,200 messages that is 24 million input tokens, which cost $48, and 1.2 million output tokens, which cost $12. Together that is $60 a month. With the cheapest model on the same price list, Claude Haiku 5.5, at $0.10 and $0.50 per million tokens for prompts of up to 100,000 tokens, the same volume costs $3. Whether such a light model handles the awkward questions well enough is something testing has to show.
When is an AI agent not worth it?
An agent does not pay off with low volume, with work that always follows the same steps or with knowledge that is not written down anywhere. Four messages a day you answer faster yourself. If the answer only exists in one colleague's head, the agent has nothing to draw on.
| Fits | Does not fit |
|---|---|
| The same kind of question comes in dozens of times a week | It is a handful of cases a week |
| The question is worded differently every time | The work always follows the same steps |
| The answer is in a document or a system | The answer is in someone's head |
| A mistake can be put right with an apology by email | A mistake costs money straight away or has legal consequences |
If the work always follows the same steps, an automation is cheaper and more dependable. And if the task takes you less than two hours a week, we would not start on it.
What can go wrong, and how do you guard against it?
A language model can give an answer that sounds convincing and is still wrong. You guard against that by having the agent answer only with what is in your data, and by having a person look at anything involving money or a sensitive conversation. A log shows afterwards what it did and which source it relied on.
In practice that is a fixed division of work. The agent drafts the quote and you send it. A refund, a discount or a complaint goes to a member of staff, with the draft reply alongside. Before the agent goes live it is given real messages from your inbox, the awkward ones included.
Two more things go wrong. The first is out-of-date information. An agent working from last year's price list gives last year's price with complete confidence. So let it look in the system itself and do not give it a copy that someone has to keep up to date. The second is a customer who tries to talk the agent round. Set the limits in what the agent is able to do, and do not count on an instruction being enough. An agent that cannot apply a discount cannot give one away either.
If you want to know whether an agent suits your work, tell us through the contact form which questions or documents come in, how many a week and which systems hold the answer. You will get a reply within one business day. Once it is clear what needs to be built you get a fixed price, or the answer that in your case it does not pay.

