AI agents for business: what they do and cost

The managing director has heard the words AI agents at three conferences in a row this year, and the IT lead has been asked in four separate meetings whether the company should have one. Most people in those meetings nod along without a clear picture of what an agent does on a Tuesday afternoon, because everyone assumes the others already know.
Eurostat counted 20 per cent of EU enterprises with ten or more employees using AI in 2025, up from 13.5 per cent a year earlier.1 The question about AI agents is therefore arriving at every desk at once, often before anyone has defined the word. This guide explains what an agent is, the two kinds you will meet, what they cost in euros and when they are the wrong answer. A test you can run on your own today closes it.
What are AI agents?
An AI agent is AI that is handed a goal instead of a question, and it chooses for itself which steps it needs, which tools to use and when the job is done. Anthropic, the company behind Claude, separates two things: a workflow follows a path someone drew up in advance, while an agent directs its own process and tool use and decides how to solve the task.2 OpenAI puts it more briefly, describing agents as systems that carry out tasks on your behalf on their own.3
The difference shows when you ask a colleague to check one contract for its termination terms, compared with asking them to go through all fifty supplier contracts and flag the unusual clauses. A chatbot suits the first request, while an agent suits the second, and you only check in when the agent starts and when it finishes.
AI moves from answering to doing, which is why it needs rules. An agent that only reads is harmless, while an agent that sends emails or changes figures in the finance system needs a clear brief, exactly like a new member of staff.
Which kinds of AI agents are there?
Two kinds of AI agents turn up in practice, namely the one you work alongside and the one that works in the background while you do something else.
Claude Cowork: the agent you work alongside
Claude Cowork, meaning Claude working inside your own files and programs, does the job while you are there, and you approve whatever matters before it happens. Anthropic describes Cowork as Claude Code's agent abilities brought to knowledge work beyond coding, and it is included in every paid Claude plan.4 Since 16 September 2026 Cowork and the ordinary chat have been merging into one Claude, which means that for most teams the first agent is already inside the tool they pay for.5
Cowork suits tasks that vary from week to week, such as reviewing a folder of contracts, building a report from three Excel files or preparing a board meeting. Once Claude is connected to your own systems through MCP, meaning an open standard that lets Claude read your own tools, it can also read in the finance system or the project archive. A person still starts the task, steers it and approves the result, and our Claude Cowork setup guide covers the first hour step by step. A task that starts from a pile of PDFs gets easier with our guide to building an AI knowledge base from PDFs.
Niche agents that work in the background
The second kind is built for one task and runs with nobody watching, starting when an email arrives, a ticket is created or the clock strikes six. Facility services company Coor has an agent like that registering incoming service requests,6 and Spotify uses background coding agents that moved around 1,800 data pipelines between systems in one project.7
Niche agents are often built with the Claude Agent SDK, meaning a developer kit that gives the same tools and working loop that power Claude Code.8 They need connections to the systems they work in, and somebody who owns them, reads the log and handles the exceptions the agent cannot solve on its own.
| Claude Cowork | Niche agent in the background | |
|---|---|---|
| Who starts it | You | An email, a ticket or a schedule |
| Who watches | You, while it works | Nobody during the run, the owner afterwards |
| Tasks | Varied, one at a time | The same task, many times over |
| What it costs | The licence | Build, connection, running and usage |
| Risk | Lower, because you see what it does | Higher, because it needs clear rules |
Our experience is that most teams should start with Cowork, because that is where you learn what the task really consists of before anyone writes code. When the same brief is carried out the same way every week, it is time to consider moving it to an agent in the background.
How do chatbots, fixed workflows and AI agents differ?
| Chatbot | Fixed workflow | AI agent | |
|---|---|---|---|
| What it does | Answers a question | Follows a fixed rule | Chooses its own route to a goal |
| Good at | Quick answers and drafts | The same steps every time | Unstructured text, rules that are hard to write down |
| Example | "Summarise this meeting" | "Move every new invoice to the right folder" | "Go through this month's invoices and flag mismatches against the contracts" |
| Risk | Low | Low, it only does what it was given | Higher, it can pick the wrong route |
OpenAI names three situations where an agent earns its keep: decisions that need judgement, rules that are hard to maintain and work with a lot of unstructured text.3 Outside those three, a fixed workflow usually does the job. Anthropic goes further and advises finding the simplest setup that works, which can mean not building an agent at all.2
Examples: how European companies use AI agents
European examples are rarer than the conference stage suggests, so here are three you can check yourself, two niche agents working on their own and one Claude working beside the user.
Coor: two niche agents in daily operations
Coor, a Nordic facility services company, has two agents in production, where one named Geoffrey registers incoming service requests and another named Adde handles administration and invoice flows. Coor calls it "not a pilot project" in its Swedish press release, although the press release gives no figure for the time saved, so that number is still missing.6
Klarna: a customer service agent, and the way back
Klarna's AI assistant had handled 2.3 million conversations by February 2024, two thirds of all its customer service chats. The company counted that as the work of 700 full-time staff, with 25 per cent fewer repeat enquiries.9 A year later the chief executive said the savings in customer service had gone too far, and Klarna started hiring again so that customers can always talk to a real person.10
The lesson is not that the agent failed, since somebody simply has to decide which conversations a human should take.
Svenska Handelsfastigheter: Claude beside the user, read access only
Svenska Handelsfastigheter, a Swedish owner of retail properties and a satori. customer, has connected Claude to Vitec for its finance figures and to iBinder for its construction project archives. Claude pulls actual results against budget per property from Vitec and reads the text inside meeting minutes and inspection reports in iBinder.11 A question about what a site meeting decided three months ago used to mean opening one set of minutes after another, and now it is asked in the chat. Both connections can only read, which means no question can change a figure in the books, and the person who makes the decision still reads the source that Claude points to.
What do AI agents cost?
The answer depends on the kind, because the agent in Cowork costs little more than the licence while a niche agent in the background costs more to build but can take over a whole flow. Five parts make up the cost, all at fixed satori. prices excluding VAT.
- The licence. Claude Cowork is included in every paid Claude plan,4 and through satori. Claude Standard costs €22 per user per month, plus €14 for running it and a one-time €490 setup on satori-claude.
- The connection. An agent that needs your finance or ticket system needs a connection, built at a fixed price agreed after a scoping call and run for €110 per month per connection.
- The running. A niche agent in the background has to be monitored, fixed and updated, which costs from €230 per month per agent.
- The usage. Background agents usually run through the API and are billed per word, at $2 per million tokens in and $10 per million tokens out for Claude Sonnet 5.5, about €1.75 and €8.77.12 A run that reads fifty invoices comes to well under a euro at those prices.
- The person. Somebody has to own the agent, read the log and approve what goes out, and that time belongs in the sum.
At list prices, ten people on Claude Standard through satori. cost €360 per month plus the setup. Adding one connection to the finance system brings the running cost to €470 per month.
The licence is rarely what makes agent projects expensive, since Gartner expects over 40 per cent of agentic AI projects to be cancelled by the end of 2027.13 BCG found the same gap in Sweden, where 52 per cent of the companies surveyed expect agents to deliver more than a quarter of their AI value.14 Only about 7 per cent of them spend more than 15 per cent of their AI budget on agents, so belief runs far ahead of money.
How do you get started with AI agents?
Starting small, starting in Cowork and letting the agent earn its trust is the order we follow with our customers, step by step.
- Pick a dull task. It should come back every week, follow clear rules and take more than an hour, and invoice checks, contract reviews and summaries of meeting minutes are good candidates.
- Do it together with Cowork. Run the task with the agent beside you for a few weeks with read access only, because an agent that only reads cannot break anything. You learn where it is right and where it needs help along the way.
- Write the job description. Note what the agent should do, in which systems and where the limit lies, because anything you cannot write down you cannot check either, and that text becomes the brief for a niche agent later.
- Keep a person on anything that leaves the building. Emails, invoices, contracts and decisions about people go through someone with responsibility, which matches Anthropic's own advice to approve manually wherever a mistake is hard to undo.15
- Measure for four weeks. Record how long the task took before, how long it takes now and how many errors the reviewer found.
Steps two and four belong in your AI policy, and our guide on how to write an AI policy together with the AI policy template gives you a base to adapt. An agent policy adds rules for what the agent may do, not only for what people type into a chat.
Data location is the next question European teams raise. Anthropic states that inputs and outputs from its commercial plans are not used for training by default, and its data processing addendum with standard contractual clauses sits in the commercial terms.1617 Data from commercial products such as Claude Enterprise and the API is stored in the US, although EU cross-region inference on Amazon Bedrock keeps requests inside EU regions such as Frankfurt, Ireland, Paris and Stockholm.1819
Every company running an agent has one duty under the EU AI Act, since Article 4 has asked providers and deployers to take measures that support AI literacy among their staff from 2 February 2025.20 The Digital Omnibus on AI softened that wording when it came into force on 27 July 2026, without removing the duty.21
Automating sits as step five of the AI Ladder we work to, after the platform, getting started, training on real tasks and structuring your data. The order matters, because Deloitte found that only 5 per cent of Nordic companies have high expertise in agentic AI.22 Teams that skip the steps end up with an agent nobody understands.
When are AI agents the wrong answer?
A task that takes five minutes once a week does not need an agent, and a task whose rules fit into a simple if-then rule is better served by a fixed workflow. Neither does any team need an agent because a competitor talks about theirs on LinkedIn.
An agent is right when the job is large, comes back often, needs text to be read and judged, and has somebody who can check the result. Without that last part the team does not have an agent, only a risk.
The short version
An AI agent does the work instead of only answering, and it comes in two kinds: Claude Cowork that you work alongside, and niche agents that take one task in the background. Both do the most good on large, dull tasks with plenty of text, and both need what a new member of staff needs, namely a job description, the right permissions and somebody who follows up.
Take five supplier invoices as PDFs and the contract they belong to, put them in one folder and open Claude Cowork. Write this prompt: "Check whether these invoices match the contract and list every mismatch with invoice number and amount." Read the answer as a reviewer would, and within fifteen minutes you have a list of mismatches that shows whether an agent would earn its place in your team.
Sources
Footnotes
-
Eurostat (2025). 20% of EU enterprises use AI technologies. https://ec.europa.eu/eurostat/web/products-eurostat-news/w/ddn-20251211-2 ↩
-
Anthropic (2024). Building effective agents. https://www.anthropic.com/engineering/building-effective-agents ↩ ↩2
-
OpenAI (2025). A practical guide to building agents. https://cdn.openai.com/business-guides-and-resources/a-practical-guide-to-building-agents.pdf ↩ ↩2
-
Anthropic (2026). Getting started with Cowork. https://support.claude.com/en/articles/13345190-getting-started-with-cowork ↩ ↩2
-
Anthropic (2026). Claude Cowork and chat are now one Claude. https://claude.com/resources/articles/cowork-is-now-claude ↩
-
Coor (2026). AI tar över operativa kärnprocesser i FM: Coor visar vägen med AI-agenter (in Swedish). https://news.cision.com/se/coor-sverige/r/ai-tar-over-operativa-karnprocesser-i-fm---coor-visar-vagen-med-ai-agenter,c4361541 ↩ ↩2
-
Spotify Engineering (2026). Background coding agents: dataset migrations. https://engineering.atspotify.com/2026/4/background-coding-agents-dataset-migrations-honk-part-4 ↩
-
Anthropic (2026). Agent SDK overview. https://code.claude.com/docs/en/agent-sdk/overview ↩
-
Euronews (2024). Klarna's AI bot is doing the work of 700 employees. https://www.euronews.com/next/2024/02/28/klarnas-ai-bot-is-doing-the-work-of-700-employees-what-will-happen-to-their-jobs ↩
-
Realtid (2025). Klarnas vd erkänner: Vi gick för långt (in Swedish). https://www.realtid.se/bank-fintech/klarnas-vd-erkanner-vi-gick-for-langt/ ↩
-
satori. (2026). Kopplingar mellan Claude och era system (in Swedish). https://www.satoriml.se/kopplingar ↩
-
Anthropic (2026). Plans and Pricing. https://claude.com/pricing ↩
-
Gartner (2025). Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027 ↩
-
BCG (2026). The Nordic AI Inflection Point: Sweden. https://web-assets.bcg.com/77/ef/8b01d86c46d2af4ade98d575bae5/the-nordic-ai-inflection-point-value-creation-or-value-bubble-sweden.pdf ↩
-
Anthropic (2026). Use Cowork safely. https://support.claude.com/en/articles/13364135-use-cowork-safely ↩
-
Anthropic (2026). Is my data used for model training? Anthropic Privacy Center. https://privacy.claude.com/en/articles/7996868-is-my-data-used-for-model-training ↩
-
Anthropic (2026). How do I view and sign your Data Processing Addendum? Anthropic Privacy Center. https://privacy.claude.com/en/articles/7996862-how-do-i-view-and-sign-your-data-processing-addendum-dpa ↩
-
Anthropic (2026). Where are your servers located? Anthropic Privacy Center. https://privacy.claude.com/en/articles/7996890-where-are-your-servers-located-do-you-host-your-models-on-eu-servers ↩
-
Amazon Web Services (2026). Regional availability of models in Amazon Bedrock. https://docs.aws.amazon.com/bedrock/latest/userguide/models-region-compatibility.html ↩
-
European Union (2024). Regulation (EU) 2024/1689, Article 4 on AI literacy. https://eur-lex.europa.eu/eli/reg/2024/1689/oj ↩
-
European Commission (2026). AI Omnibus enters into force. Shaping Europe's digital future. https://digital-strategy.ec.europa.eu/en/news/ai-omnibus-enters-force ↩
-
Deloitte (2026). State of AI in the Nordics 2026. https://www.deloitte.com/dk/en/Industries/technology/research/state-of-ai-in-the-nordics.html ↩
Frequently asked questions
An AI agent is AI that is handed a goal instead of a question and chooses the steps and tools it needs to reach it. Anthropic describes agents as systems where the language model directs its own process and tool use. A chatbot answers and waits, while an agent reads, searches, compares and finishes the job inside the limits you set.
Yes. Claude Cowork is a general agent that works in your own files and programs while you are there: you give it a task, it does the work and you approve what matters. It is included in every paid Claude plan. The other kind is the niche agent, built for one task and running in the background with nobody watching while it works.
A chatbot answers one question at a time, and a fixed workflow follows a rule someone drew up in advance. An AI agent chooses its own route to the goal, which makes it good at unstructured text and rules that are hard to write down, but slower and dearer where a fixed rule would do.
It depends on the kind. The agent in Claude Cowork costs nothing beyond the licence, which through satori. is €22 per user per month for Claude Standard plus €14 for running it and a one-time €490 setup. A niche agent in the background adds a build at a fixed price agreed after a scoping call, €110 per month per connection, running from €230 per month and usage billed through the API.
Start in Claude Cowork with a dull task that comes back every week. Run it together with the agent for a few weeks with read access only, and write the steps down as a job description. Once the task is done the same way every time, it can move to a niche agent in the background, where a person still approves anything that leaves the building.
The AI Act regulates AI systems by risk, and an agent that checks invoices is usually not a high-risk system. Article 4 has asked providers and deployers to take measures that support AI literacy among their staff since 2 February 2025, in a wording the Digital Omnibus on AI softened from 27 July 2026. The staff who run an agent should learn what it does and where it fails.

About the author
Joachim helps European companies understand and use AI in everyday work. As founder of satori., he trains teams, coaches leadership groups, and shows how tools like Claude become a natural part of work.
