Knowledge
What are AI agents, and where are they worth it for mid-sized companies?
AI agents connect a language model with tools: they don't just answer questions, they carry out tasks inside your systems — transferring data, triggering processes, drafting emails. For mid-sized companies they pay off on clearly bounded, recurring tasks where a person signs off on the result.
Chatbot, assistant, agent — the difference
- Chatbot: answers questions in a dialogue. It reads, but it doesn't act — the classic example is the FAQ help on a website.
- Assistant: supports you with a task, such as drafting an email or summarizing a document. Carrying it out stays with the human.
- Agent: is given a goal, plans the necessary steps itself, and carries them out through interfaces inside your systems — it queries data, transfers it, and triggers processes.
What agents can do reliably today
Agents work reliably where the task is clearly defined and the systems involved offer clean interfaces. Typical examples:
- Transfer data: read information out of emails or documents and enter it into the ERP or CRM.
- Trigger processes: create an order, open a ticket, set a follow-up reminder.
- Prepare emails: pre-sort incoming messages and draft replies with the right data from your systems.
- Gather information: pull data from several systems together for a single case, for example as groundwork for a quote.
Where the limits are
- Error tolerance: language models make mistakes. That's why an agent belongs only in processes where a single error is noticeable and correctable — not in irreversible actions without oversight.
- Auditability: every action an agent takes should be logged. What it did must be verifiable if there's ever any doubt.
- Permissions: an agent needs its own technical account with minimal rights. An agent working with broad user permissions is a security risk.
Sensible first use cases
- Pre-sort incoming email in sales or service and prepare replies.
- Transfer data from orders or invoices into the ERP — with sign-off from an employee.
- Groundwork for quotes: pull master data, prices, and history from several systems together.
- Take in internal requests (for example to IT or administration), categorize them, and prepare the standard cases.
What to watch for when you introduce one
- Start small: one process, one team, one measurable result — expand only after that.
- Human in the loop: at the start, a person always signs off on critical actions. Automation without sign-off is the final stage of expansion, not the first.
- Limit rights: its own technical account, only the access it needs, logging from day one.
- Set expectations: an agent doesn't replace case handling, it takes the routine cases off its plate.
FAQ
Frequently asked questions
What's the difference between a chatbot and an AI agent?
A chatbot answers questions; an agent carries out tasks: it queries systems, transfers data, and triggers processes. The chatbot talks, the agent acts — sensibly, with human sign-off.
Are AI agents reliable enough for production use?
For clearly bounded tasks, yes, provided the surrounding process catches errors: actions are logged, a person signs off on critical steps, and the agent only has the rights it genuinely needs.
Which tasks are a good place to start?
Recurring tasks with clear rules and manageable damage when something goes wrong: pre-sorting incoming email and preparing replies, transferring data from documents into the ERP or CRM, compiling information from several systems.
What does introducing an AI agent cost?
A bounded use case can be delivered as a fixed-price pilot from €39,000, in production within 6–8 weeks. To clarify up front whether an agent is worth it for your task, there's the Discovery from €1,900.