An AI agent is a program that picks its own steps toward a given goal and calls your tools along the way: calendar, e-mail, CRM, stock system. A chatbot answers. An agent answers and then acts. Most firms with 10 to 100 people don't need an agent right away, though. Anthropic, the company behind the Claude models, advises looking for the simplest route and adding complexity only when it's needed. Often it's enough to ask a model one question over your documents. An agent pays off where the steps can't be written down in advance. What actually runs in firms in 2026 is covered in the article on AI agents for business.
- A chatbot answers, a workflow runs steps written in advance, an agent chooses its own steps.
- Anthropic advises starting with the simplest option; often one model query over your documents is enough.
- Give an agent the smallest possible rights and have a person approve high-impact actions.
- Run the pilot on 20 real cases and track how many came out right.
What an AI agent is and how it differs from a chatbot
In its guide of 19 December 2024, Anthropic separates two things. Workflows, it says, are systems where the model and the tools are run by code written in advance. With agents, the model directs its own process and its choice of tools. OpenAI puts it similarly: an agent can plan, complete tasks with tools and keep context across steps.
That gives firms three levels to choose from. The table puts them side by side.
| Chatbot | Workflow with AI | Agent | |
|---|---|---|---|
| Who sets the steps | Nobody, it only answers | You, in code, in advance | The model, at run time |
| Reaches into your systems | No | Yes, at prescribed points | Yes, as it sees fit |
| Example | An answer about your opening hours | Pulls the details out of every enquiry e-mail and creates a record | Checks the enquiry, finds what's missing, asks for it and proposes the next step |
| Main risk | A wrong answer | A misread value | A wrong action in your system |
The examples and the row on risks are our own reading, not quotes from the sources.
Lumenax describes an agent on its AI page like this: software that completes a task on its own and, unlike a chatbot, is connected to your systems and carries the action through. The last part matters. Whoever carries an action through changes something in your data, and that is where the risk sits.
When a firm needs an agent
Anthropic writes that agents often pay for better results with slower responses and higher costs, and that for many tasks one model query that first looks up material in your documents is enough. A simple test follows from that. Ask yourself three things:
- Can you write the procedure as a list of steps? If yes, a workflow is enough. It is cheaper, predictable and easy to check.
- Is an answer enough, or does something have to happen? If you only want to find a value in your documents, an agent is a needlessly heavy tool.
- Can you check the result? If nobody would catch a mistake, the agent doesn't belong in the system.
An agent makes sense when you answer no to the first question, "something has to happen" to the second and yes to the third. Typically these are tasks where the procedure differs from case to case: an enquiry with missing details, a case with several possible paths, material from several sources.
- Can you write the procedure as a list of steps?YesA workflow is enough
- Is an answer enough, or does something have to happen?An answer is enoughOne model query over documents is enough
- Can you check the result?NoThe agent doesn't belong in the system
- The procedure can't be written down, something has to happen and you can check the resultAn agent makes sense
What it looks like in practice
In our klientbox product for service firms, clients book online by themselves. The AI receptionist Klára works with the firm's data: she drafts a personal message for everyone who hasn't visited for a while, and the owner checks it and sends it. You can also ask her who hasn't been in for a long time or how much the firm earned this month, and she answers from the firm's data. According to the klientbox page, the last word stays with a person.
That is a good pattern to start with. The agent prepares, a person confirms. Only when the material proves reliable does it make sense to allow more.
What an agent needs to work
- Data it may use. The model doesn't know your firm. It has to see the documents and records it answers from.
- Tools with the smallest possible rights. OWASP compiles a ranking of risks for programs built on language models, and one of them it calls "excessive agency": the model gets more functions, rights or freedom than it needs, and then takes a harmful action in response to an unexpected or manipulated output. It recommends giving the agent as few tools and as few rights as possible.
- Human approval for high-impact actions. OWASP advises that a person approves them before they are carried out. For payments, contracts and deleting data, that always applies.
- A record of what the agent did. Without it you can't find out why something happened, even when it did happen.
What can go wrong
Agents are not yet reliable at everything. Researchers at Carnegie Mellon built a simulated software company and had agents carry out its work tasks. In the current version of the study, dated 10 September 2025, the best agent completed 30% of the tasks on its own. Newer models aren't in it, so today's result will differ.
Mistakes come in two forms. The agent gets it wrong and the firm notices from the result. Or the agent gets it wrong and the action goes through before anyone can stop it. The second case is worse, which is why an agent should wait for approval at sensitive steps.
One thing stays open: numbers from other people's tests don't transfer to your firm. You'll learn how reliable an agent is on your data only by trying it on your own cases.
How to start without risking operations
- Pick one process. Well bounded, repeated, and one where a mistake destroys nothing.
- Workflow first, then agent. If you can describe it step by step, write it as a workflow. Keep the agent for the part where the decision depends on the case.
- Take 20 real cases from the last few weeks and run the pilot on them. Write down how many came out right and what went wrong.
- Keep a person confirming. Loosen supervision only when the pilot numbers show the agent copes.
Lumenax works in four steps: scoping, prototype, operation with safeguards and a check of the benefit. According to its page on AI in operation, it delivers the first working prototype usually within two to four weeks and names a fixed price after a short scoping.
The European Commission states that the obligations around AI literacy have applied since 2 February 2025. Whoever uses AI tools in a firm should check what follows for them. The rest is summarised on the AI Act page.
Common questions
Is an AI agent the same as a chatbot?
No. A chatbot answers. An agent also calls tools and changes the state of your systems, for example by creating a record or writing a message. The risk differs accordingly: a chatbot's mistake is a wrong answer, an agent's mistake can be a wrong action.
Will an AI agent replace employees?
No source we opened claims that. An agent handles only part of the tasks on its own and needs a person's approval at sensitive steps. A well-bounded routine can be handed over to it, though.
Does a small firm need an agent, or is a chatbot or workflow enough?
Often a workflow or a single query to a model over your documents is enough. Anthropic recommends starting with the simplest option and adding complexity only when it's needed.
How much does an AI agent cost?
We can't give a general price, and no source does. With Lumenax you get a fixed amount after a short scoping. With cloud models you pay for usage according to the provider's price list.
How do I limit an agent's rights?
Give it only the tools and rights the task requires, and have a person approve high-impact actions. That is what OWASP recommends in its ranking of risks for language models.
Sources
- Anthropic, Building effective agents, 19 Dec 2024: https://www.anthropic.com/engineering/building-effective-agents
- OpenAI, Agents guide, viewed 25 Oct 2026: https://developers.openai.com/api/docs/guides/agents
- OWASP, Top 10 for LLM Applications, LLM06:2025 Excessive Agency, viewed 25 Oct 2026: https://genai.owasp.org/llmrisk/llm062025-excessive-agency/
- Carnegie Mellon and co-authors, TheAgentCompany, arXiv:2412.14161, v1 18 Dec 2024, v3 10 Sep 2025: https://arxiv.org/abs/2412.14161
- European Commission, Regulatory framework on AI, updated 19 Oct 2026: https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
- Lumenax, page on AI in operation, viewed 25 Oct 2026: https://www.lumenax.cz/ai-implementace
- Lumenax, klientbox, viewed 25 Oct 2026: https://www.lumenax.cz/klientbox
Ollama: how to run local AI over your own company data
AI agents for business: what actually works in firms in 2026
Risk analysis for NIS2 and ISO 27001: a step-by-step guide
