16 October 2026 · 6 min read · by Bart Broothaerts
AI agents for business: what they can and cannot do
Short answer: An AI agent is software that watches for a situation, decides what to do and acts on it, using a language model for understanding and writing and your systems for data and actions. Agents handle repetitive work with clear rules well, such as following up invoices, routing leads or preparing documents. They still need people for judgement calls, money decisions beyond set limits, and anything sensitive.
What makes something an agent
A chatbot answers when spoken to. An agent works on its own: it notices that something happened (an invoice became overdue, a lead arrived, a document is about to expire), decides what that means, and does the next step, or escalates it.
Under the hood it combines three things: a language model to understand and write, access to your systems to read data and take actions, and rules that define what it may do without asking.
Tasks agents handle well today
Agents are strongest where work is frequent, follows rules and is costly when it slips:
- Following up overdue invoices with the right tone for each customer.
- Routing inbound leads and enquiries to the right person, with a short researched brief.
- Flagging permits, certificates and contracts before they expire.
- Preparing quotes, summaries and reports from data already in your systems.
- Spotting customers whose behaviour changes before they leave.
Where people stay in charge
Agents are not the right final authority on prices, discounts, payments or anything with legal or reputational weight. Good designs keep those decisions in hard rules or with a person, and let the agent prepare the decision rather than make it.
The same goes for unusual cases. An agent that knows when to escalate is worth more than one that tries to handle everything.
Keeping agents safe
Three principles keep agents reliable in a real company:
- Rules decide, AI words: limits and approvals live in code you control, not in the model's judgement.
- Correct beats plausible: agents act on checked data from your systems, not on what sounds right.
- Humans approve what matters: anything above a threshold goes to a person, with the context attached.
How to start
Pick one task that happens often, has clear rules and costs money when it slips. Build an agent for that, measure it against how the work was done before, and only then add the next one.