AI agents with Claude: what they do for EU firms

The managing director has heard the words AI agents at three conferences in a row this spring, 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. 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, so the question is arriving at every desk at once.¹

AI agents differ from the chat window most teams already know, and the gap is larger than the name suggests. This guide explains what an agent is without developer jargon, shows three cases a European office can copy, and ends with the honest list of when an agent is the wrong answer.

What are AI agents, in plain words?

Anthropic separates two kinds of systems, where a workflow follows a fixed path written in code and an agent decides its own next step. In Anthropic's words, agents are systems that "dynamically direct their own processes and tool usage".² The agent plans, acts, checks the result and adjusts, in a loop, until the task is done or it needs your input.

Picture handing an assistant one goal in the morning, for example going through last week's customer emails and flagging the ones that need a reply today. A good assistant reads the emails, judges what is urgent and leaves a sorted list on your desk by lunch. An agent does the same job, although somebody still reads the list before any reply goes out.

Chatbot or AI agent: what is the difference?

A chatbot waits, since you ask, it answers and then it waits for the next question, which leaves every copy and paste between steps to you. An agent keeps going, because it takes the task to the end through several steps and often several tools, such as reading files, searching systems and writing documents.

The difference shows when you give Claude a folder instead of a paragraph. Ask Claude in chat to summarise one contract and you get one summary. Hand it a folder of fifty supplier contracts with the instruction to flag unusual termination clauses, and you get one report covering all fifty. The model is the same in both cases, while the way of working is not.

Agent Skills make the second kind of request repeatable. Anthropic describes skills as folders of instructions, scripts and resources that an agent finds and loads when a task calls for them.³ You write the working method for a monthly Excel analysis once, and the agent picks it up every time the analysis comes round, instead of you explaining it again.

Three ways European teams use AI agents with Claude

Anthropic folded its Cowork mode into the main Claude app on 16 September 2026, so the agent that works through folders now starts from an ordinary conversation.⁴ The change reaches Pro and Max first and Team soon after. Each of the three cases below runs on that setup, and each still has a step a person does by hand.

Invoice checks that run while you do something else

Picture the finance team at a Rotterdam freight forwarder that matches 140 quarterly invoices against contracts, a job that takes one person a full afternoon. With Claude pointed at the invoice folder and given the instruction once, the agent works through every file and returns a table of mismatches in about twenty minutes. The controller still opens every flagged invoice before anything is paid, because a badly scanned PDF can still be misread and the agent does not always say so.

Contract review in bulk

Legal and sales teams that read contracts of a hundred pages or more get the largest gain from an agent, because it reads every page the same way every time. The job changes from one lawyer reading one contract to the agent flagging risk clauses across the whole folder while a lawyer reviews only what it found. A human still signs off, since the agent cannot know which clauses your company agreed to accept last year.

Internal support that answers before IT opens the inbox

Most internal support tickets repeat the same five questions, such as how to reset a password or who approves purchases over €5,000. An agent connected to your internal documentation through MCP, meaning an open standard that lets Claude read and write in your own tools, can answer those directly and pass the rest to a person. The IT team still owns the answers in the documentation, and an agent reading an outdated policy gives an outdated answer. Our Anthropic MCP guide covers that connection step by step.

Who may the agent act as under GDPR and the AI Act?

An agent that only suggests carries little risk, while an agent that sends emails, approves invoices or updates customer records needs three limits written down before it goes live. Decide who owns the agent, which systems it may touch and which actions need a human approval, the same questions you would ask about a new hire.

• Give the agent the smallest set of permissions the task needs, with nothing added just in case.
• Keep a human approval on payments, customer data and contracts until the agent has run for a month without a wrong action.
• Log every action, so that you can show what happened and why when your data protection officer asks.

Data location is the second question European teams raise. Anthropic states that inputs and outputs from its commercial plans are not used to train its models by default, and its data processing addendum with standard contractual clauses is built into the commercial terms.⁵⁶ Data from Claude Team and Enterprise is stored in the US, although Amazon Bedrock offers Claude through an EU cross-region profile that covers Frankfurt, Ireland, Paris and Stockholm.⁷⁸

The EU AI Act adds one duty that applies to every company running an agent. Article 4 has required AI literacy for the staff who operate AI systems since 2 February 2025, meaning they know what the system does and where it fails.⁹

What does it cost to get started with AI agents?

Nobody pays extra for the word agent, since the cost sits in setting up the base properly and connecting the agent to the task that eats the hours. A Claude Team Standard seat costs $20 per user per month billed annually, which is about €17, and every agent feature above runs inside that seat.¹⁰

satori-launch gives the whole team that base, with a webinar, a full-day workshop on your own documents and up to five finished skills rolled out, for €560 per user from eight users. After that, satori-claude runs licences and support for €14 per user per month plus the licence, after a one-time €490 setup. Connecting the agent to your CRM or ticket system is a separate build through satori-mcp or satori-automation, at a fixed price agreed after a scoping call. All satori. prices are fixed and exclude VAT.

When are AI agents the wrong answer?

Gartner predicts that over 40 per cent of agentic AI projects will be cancelled by the end of 2027, because of rising costs, unclear business value or weak risk controls.¹¹ The technology is rarely the cause, since most of those projects never had one clear task to solve in the first place.

A task that takes five minutes once a week does not need an agent, and neither does a task whose rules change every month and that nobody can write down. Agents pay off when the task repeats, follows rules and takes at least two hours a week today, which is why we test narrowly for four weeks before we build anything larger around it.

The short version

AI agents take a goal and finish it over several steps, which makes them useful for invoice checks, contract review and internal support. Set the permissions first and log everything, since the governance decides whether the project survives.

Write down the three tasks from last week that took you longest, with the hours each one took, and mark the ones that follow the same rules every time. In fifteen minutes you have a short list, and the task at the top of it is your first agent pilot.

---

Sources

Footnotes

1. Eurostat (2025). 20% of EU enterprises use AI technologies. https://ec.europa.eu/eurostat/web/products-eurostat-news/w/ddn-20251211-2 ↩

2. Anthropic (2024). Building effective agents. https://www.anthropic.com/research/building-effective-agents ↩

3. Anthropic (2025). Equipping agents for the real world with Agent Skills. https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills ↩

4. Anthropic (2026). Claude Cowork and chat are now one Claude. https://claude.com/blog/cowork-is-now-claude ↩

5. 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 ↩

6. 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 ↩

7. 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 ↩

8. Amazon Web Services (2026). Regional availability of models in Amazon Bedrock. https://docs.aws.amazon.com/bedrock/latest/userguide/models-region-compatibility.html ↩

9. European Union (2024). Regulation (EU) 2024/1689, Article 4 on AI literacy. https://eur-lex.europa.eu/eli/reg/2024/1689/oj ↩

10. Anthropic (2026). Plans and Pricing. https://claude.com/pricing ↩

11. 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 ↩