AI assistants
An assistant that answers from your material. Not from the internet’s.
A general chatbot knows everything and nothing about your business. A custom assistant knows your product, your methodology, your catalogue and your policies — and says so with a link to where it got each answer.
Custom AI assistant development means building an assistant on your own material, wearing your name, doing one job you actually need doing. It answers from documents you own and edit, and it refuses questions outside them.
The three jobs it usually does: answering support questions your documentation already covers, guiding people to the right thing inside a library too big to browse, or delivering a version of your methodology to people who cannot afford your time.
Trained on your content — and what that really means
"Trained" is the word everybody uses and it is usually the wrong one. For most businesses nothing gets retrained. Your material is supplied to the model as the source it must answer from, which is faster to build, cheaper to run, and correctable in seconds — change the document and the answer changes with it.
The practical consequence matters: there is no retraining cycle to wait for and no black box holding a stale version of a policy you updated last month.
- Your help articles, so support questions get answered at three in the morning
- Your course library or catalogue, so people find the right thing instead of the search box
- Your methodology, so the framework you deliver by hand can also be delivered by software
- Your product documentation, inside the product, on the screen the question is about
Bounded, so it cannot embarrass you
The risk everybody names first is the assistant inventing something. It is a real risk and it is a solvable one, and the solution is boundaries rather than optimism.
An assistant built here answers only from the material it has been given and says it does not know when the answer is not in there — with a route to a human attached. It links its source on every answer, so it is auditable. And it is scoped to read and talk rather than to act, so the worst thing available to it is an unhelpful answer, not a confident wrong action.
Where it should have more reach than that — seeing account data, taking actions — that is a deliberate decision made with you, with its own safeguards, rather than a default.
What it costs and what it runs on
Usually the cheapest impressive thing you can add.
Most of the work is deciding the boundary, not writing the code.
- Built into something you already run
- From around £4,000.The usual case. An assistant inside an existing product, site or membership, using furniture that is already there.
- A standalone assistant
- From around £6,000.Its own small application with its own branded world, embedded wherever your traffic lands — and portable if you change platform.
- Running cost
- Usage-based.You pay the AI provider per conversation. It is typically pennies each, modelled against your real traffic before you commit — and rate-limited so a bad day cannot produce a surprising bill.
- Who owns it
- You do.The code, the knowledge base and the prompts. Including the prompt, which is usually the most valuable single file in the project.
The evidence
The claim, with a working example attached.
The boring answers
Asked and answered, without the discovery call.
Can AI be trained on my own content?
Yes. Your articles, courses, product documentation or methodology become the material the assistant answers from, and it is instructed to work from nothing else.
For most bodies of content nothing is retrained in the technical sense — the material is given to the model as its source. That is a feature: edit the document, and the assistant’s answer changes immediately.
How do you stop an AI assistant making things up?
Three things together. It is instructed to answer only from your material and to say it does not know otherwise. It links the source of every answer, so anyone can check it. And it is scoped so it can read and talk but not act.
MAVIS, the assistant inside ECHO, is built exactly this way — she cannot see a user’s account, cannot change anything, and hands over to a human in one tap when she is out of her depth.
Can I control what the AI is allowed to say?
Yes — its knowledge, its tone, and its refusals are all specified. You decide the topics it will not touch and the answers it must give verbatim.
This is ordinary and expected. Every assistant built here ships with a written list of what it will not discuss.
Can an AI assistant be added to the software or site I already have?
Usually, and that is the cheaper route. MAVIS was built into a help panel ECHO was already shipping. Feral Bot embeds into pages that already existed.
You rarely need to rebuild anything to add one.
What does an AI assistant cost to run each month?
You pay the AI provider for what gets used, typically pennies per conversation, plus hosting in the tens of pounds a month.
It is modelled against your actual traffic before you commit, and rate-limited per user so nothing runs away. You are never told to find out in arrears.
Tell me what you are trying to build.
Describe the problem rather than the specification. Working out what it actually needs to be is the first part of the job.