The word "agent" is doing a lot of work right now, most of it marketing. Here is what it means when we say it.
An agent is a program that reads state, decides what to do, and does it through a fixed set of tools. The model supplies judgement. The tools supply reach. Everything else is engineering.
Chat is not the job
A chat box next to your ERP can explain your data to you. It cannot reconcile a payment, update an order or send a confirmation, because it has no hands.
Giving it hands is the whole problem. An agent that can act inside a business system needs:
- Tools with narrow contracts. Not "access the database" but "mark order 4821 as paid" and "send this template to this number". Every tool does one thing and validates its input.
- A log of every action. What the agent saw, what it decided, what it did, and what came back. If you cannot replay it, you cannot trust it.
- A place to stop. Anything that costs money or reaches a customer can require approval. The agent proposes; a person confirms. As the record of good decisions grows, the approvals can shrink.
Deterministic where it matters
Billing, access control and inventory follow rules. They should keep following rules. We do not put a model in the path of "did this customer pay" or "does this member have access". Those are lookups, not judgements.
The agent sits beside those systems and acts through them. It notices that a payment does not match an order and raises it. It sees that a booking changed and sends the update. It reads an enquiry and routes it. Judgement at the edges, rules at the core.
What it looks like in practice
A payment arrives that does not match any open invoice. Today, someone notices at month end. With an agent: the mismatch is spotted within minutes, the closest invoice is identified, and a person gets one message with a proposed match and a button to confirm.
That is not dramatic. It is also not a chat box. It is a piece of the operation that used to depend on someone remembering, now running on its own with a person in the loop.
What we are careful about
We measure. Every agent we build ships with an evaluation: what it got right, what it got wrong, and how that changes as the system learns. If the numbers are not good, it does not go live, and if they slip, we find out before the client does.
That is the difference between an AI feature and an AI system. We build the second kind.