An AI agent for a small company
An ordinary filter looks at the sender and the subject line. An agent reads the content and understands what it is about. That one difference decides what you can do with it.
What it is good for
Questions that come up over and over. Is there a slot free, what does it cost, how do I get there. The agent recognises the question and puts an answer together from what you give it: the price list, the calendar, the booking rules.
Pulling data out of documents. The amount, the due date, the number and the counterparty from an invoice land in a summary without retyping. It works on scans and on photographs taken with a phone.
Long correspondence. A thirty-message thread boiled down to three sentences and a list of what was agreed. Useful when you come back to a matter a month later.
Emails written any old how. It pulls the facts out of a message where the customer described everything at once and named none of it directly.
Sorting what comes in. It recognises what a message is about and routes it where it belongs, instead of guessing from a word in the subject line.
What it is not
It is not a bot that talks to your customers without your knowledge.
By default a human has the last word, and I think it should stay that way at least for the first few weeks. A model can write something foolish with complete confidence, and it is you, not me, who will be explaining it to your customer.
Once you have watched it for a month and seen the answers are good, we can let some of them go out automatically. We start with the simplest ones, like confirming a slot. But that is a decision made on evidence, not at the start.
There is no sense in it for difficult matters either. A complaint from an annoyed customer looks like any other message in the system, and costs the most. I mark those as urgent and leave them to a person.
When it pays
When a large share of what comes in is variations on the same few things. Then the agent takes the repetitive part off your hands and you deal with what actually needs thinking about.
If every matter is different and needs its own decision, the agent will not help much. Then it is better to automate what happens after the decision: confirmations, reminders, calendar entries.
What it costs
Setup from 1 500 PLN. That covers preparing the knowledge base, connecting it to your inbox or your form, and a week of watching it, during which I fix whatever the agent reads wrongly.
At a small company’s volume the model itself usually costs a few tens of złoty a month. I give you that figure up front, together with the quote.
Where the agent gets it wrong
Worth knowing in advance, because it decides where to let it in at all.
It copies numbers well and loses context. It will pull the amount off an invoice without a mistake. But “same as last time, only twice as much” needs a memory of what happened last time — and the agent does not have one unless you build it.
It does not know what it does not know. Asked about something outside its scope, it answers in a confident tone instead of admitting it has no idea. That is why everything that goes out passes through a person.
A format change on the other side is not an error to it. If a supplier rearranges a document, the agent will still pull something out — just not the right thing. Hence the alerts when a reading is uncertain.
These are not reasons not to do it. They are reasons for the agent to propose, and to leave the decision and the sending to you.
Where to start
Write down what comes into your business most often and what you then do with it by hand. I reply within a day, and the first half hour of talking is free.