An AI agent is software that doesn’t just answer questions: it carries out a task from start to finish, using your tools and your data. This guide explains what AI agents can do for a business today, where they are worth the money, what they cost and how to start with a small pilot.
A lot of companies have tried AI by now. Somebody on the team uses a chatbot to rewrite emails or summarise a document, and that’s where it stops. It saves a few minutes, but the work itself hasn’t changed: people still copy data from one system to another, still prepare the same report every Monday, still sort the same tickets by hand.
Agents are the step after that. We build them for clients and we use them ourselves every day, so this is what we have learned in practice.
AI agent or chatbot: what is the difference?
A chatbot replies to a message. You ask, it answers, and anything that needs doing is still up to you.
An agent receives a goal, plans the steps and performs them. It can read from your CRM, query a database, call an API, write a file or open a ticket, and it checks its own result before it moves on. The model does the reasoning, while the tools it is allowed to use decide what it can actually do.
That last part matters. An agent only touches the systems you connect, with the permissions you give it, and where a mistake would be expensive it stops and waits for a person to approve.
Where AI agents pay off
The best candidates are tasks that are repetitive, follow clear rules and already cost somebody a few hours every week. Some examples:
- Recurring reports. The agent collects figures from sales, e-commerce and analytics, writes the weekly summary and flags anything unusual. Nobody has to export spreadsheets any more.
- Support ticket triage. Incoming requests are read, classified, routed to the right person and answered with a draft reply when the case is a known one.
- Keeping systems in sync. Orders, customers and stock stay aligned between the CRM, the shop and the accounting software, and the agent reports the records that don’t match.
- Documents. Invoices, contracts and forms are read, the relevant data is extracted and entered where it belongs. We covered a simple case in our post on using AI to read and write Excel files.
- Development and operations. Code review, deployment checks and error monitoring. Patcherly, our own product, is an agent of this kind: it detects a bug in production, drafts the fix and applies it only after you approve.
Agents are a poor fit for work that needs judgement, negotiation or responsibility. Those stay with people, and a good project says so from the start.
Do agents work for a small business?
Yes, and often better than in a large one. A small company has fewer systems, fewer approvals and a clearer view of where time is lost, so a single well-chosen workflow can free up hours every week.
What a small business should avoid is the big project. You don’t need an “AI strategy” to begin. You need one process that annoys everybody, and a way to measure whether it got better.
What about our data?
It depends on what the agent handles. For public or low-risk information, a commercial model through its API is fine and quick to set up. For client data, contracts or anything covered by GDPR, the agent can run on a private model inside your own infrastructure, so nothing leaves your network. We explained the options in how to safely use AI at work.
What drives the cost of an AI agent
There is no price list, because an agent is built around your process. Four things decide the budget:
- How many workflows you want to automate. One is a pilot, five is a programme.
- How many systems need to be connected, and whether they have a usable API.
- How much control you need: approvals, logs and tests take time to build, and they are what makes an agent safe to leave running.
- Where the model runs. A cloud API has a small cost per use. A private model has a higher initial cost and lower running costs.
There are two kinds of cost to plan for: building the agent, and running it afterwards (model usage, hosting, maintenance). We quote the first as a fixed or phased price before work starts, and we estimate the second with you based on expected volumes.
As a reference, a first pilot on a single workflow starts at around €15,000. Larger projects, with several workflows and systems to connect, are quoted in phases.
How a first pilot works
- Pick one workflow. Something frequent, well defined and measurable, for example “prepare the weekly sales report”.
- Define what success means. Hours saved, errors avoided, response time. Without a number you can’t tell whether it worked.
- Connect the tools. We give the agent controlled access to the systems it needs, usually through the Model Context Protocol (MCP), with read-only permissions wherever possible.
- Run it with a person in the loop. For the first weeks every output is reviewed. This is where the agent gets tuned.
- Measure and decide. If the numbers are good, the agent goes into production and you choose the next workflow. If they aren’t, you have spent little and learned a lot.
A focused pilot usually takes a few weeks.
Frequently asked questions
What are AI agents, in simple terms?
Software that uses an AI model to complete a task on its own: it plans the steps, uses the tools it has been given and checks the result, asking a person for approval on the delicate parts.
Do AI agents replace employees?
No. They take over the repetitive part of a job, so people can spend their time on decisions, exceptions and clients.
Do we need to change the software we already use?
Usually not. Agents are connected to what you have, such as your CRM, your e-commerce and your shared folders, through their APIs.
What if the agent makes a mistake?
Every action is logged, risky steps need human approval, and the agent only has the permissions it needs. A well-designed agent fails in a visible and reversible way.
Have a workflow in mind? See our AI agents and automation services, or tell us about it and we’ll suggest a realistic pilot.