A chatbot talks. An agent works.
Chatbots answer questions. Agents get things done — they read what your business already knows, work out what needs doing, and then do it in your own systems. Here is the difference in plain English, with the technical names alongside for anyone who wants to check the working.
Why an off-the-shelf chatbot won't do much for you
A language model on its own knows a great deal about the world and nothing whatsoever about your business. It has never seen your prices, your stock, your contracts or the rules you work under. Ask it something it doesn't know and it won't hold back — it will answer anyway, sound perfectly certain, and now and then be wrong. In a business, that is worse than having nothing at all.
Everything that follows is there to fix those two things: to give it your operation, and to stop it making things up.
It reads your documents before it answers
Rather than trusting whatever the model happens to remember, the system searches your own material first — contracts, manuals, catalogues, policies, regulations — and hands it those exact passages to read. Only then does it write the answer. It is the difference between asking a colleague off the top of their head and asking them to check the file first.
Your documents are broken into passages and indexed by meaning rather than by wording, so someone can ask "what am I owed for my Christmas bonus?" and the system finds the clause about the annual premium, with not a single word in common. The technique is called RAG — retrieval-augmented generation.
- YOUR DOCUMENTSYour contracts, manuals, catalogues and policies
- INDEXED PASSAGESSearched by meaning, not by exact wording
- THE ANSWERWith the document and clause it came fromArt. 198 · staff_handbook.pdf
The agent searches your documents before answering, and shows you where it came from
The agent searches your documents before answering, and shows you where it came from
Two things follow from that: you update the agent by editing a document, with no retraining involved, and it will only ever answer on what you have given it.
Answering isn't the job. Doing is.
The agent is given a set of tools: look something up in your documents, save a customer, book an appointment, raise an invoice, put a report together. It works out which one to reach for and with what details, and your system carries it out for real. This is not text describing an action — it is the action. It is the same agent that handles the chat: it finds the answer in your own information first, then it acts on it.
Each tool is declared with a name, what it is for and what it needs to run. The model chooses; your code executes. That handover is called tool calling, and it is what separates an assistant that makes suggestions from one that gets on with the work.
- YOUR CUSTOMER“Do you service fleets? How much does it cost?”
- YOUR BUSINESSsearch_documentsLooks up the answer in your documents
- YOUR BUSINESSsave_leadSaves the lead, along with what they asked about
- YOUR BUSINESSbook_appointment · TUE 3:00 PMBooks against your genuine availability
- AVAILABLE 24/7LEAD RECORDED · APPOINTMENT BOOKEDNOBODY LIFTED A FINGER
The agent checks your documents, replies, records the customer and books them in — then carries on
The agent checks your documents, replies, records the customer and books them in — then carries on
You draw the line: the agent can only do the things you have handed it a tool for. Nothing else.
Every figure sourced, and no sums done by the model
When the answer matters, the agent shows you where it came from — the document, the clause, the article. Anyone can check it for themselves in about ten seconds.
And wherever there is money involved — a redundancy payment, a tax calculation, a total due — the model does not work it out. Tested, auditable code produces the figure and the agent simply writes it up. The model writes; the code counts.
- What are they owed in redundancy?What are they owed in redundancy?
- THE MODELWrites the answer, but does none of the sumsWRITES, NEVER COUNTS
- AUDITED CODETested code produces the figurecalculate_redundancy() · 12 tests passing
- THE FIGURE$1,240.50EXACT TO THE PENNY
- Art. 198 · Labour CodeThey are owed $1,240.50 for eight years' service.With the article of law behind it
The model writes; the code counts. Every figure with the law behind it
The model writes; the code counts. Every figure with the law behind it
That split is what makes an agent fit for serious work, rather than just good conversation.
Where your information lives is your call
There is no single right answer here. It depends on how sensitive your information is, what you are required to comply with, and what you can spend. We settle it together before a line of code is written, and we write it down. One thing never changes: your documents remain yours, and they are never used to train anybody's models.
Which job in your business would an agent take off your hands?
In our first session we take one real task from your day-to-day and tell you straight whether it can be automated, how, and what it would take. No obligation.