Wovekit

Putting an AI assistant on your website without it making things up.

How to add a Claude- or GPT-based assistant to a small-business website safely: approved content only, honest identity, human hand-over and logging. With this site's own assistant as the worked example.

By the Wovekit team · · 7 min read

The orange dot in the corner of this page opens an AI assistant. It can explain what we build, suggest a sensible first step for a project and find a time for a call. It will also tell you plainly that it is an AI, refuse to invent a price, and hand you to a person the moment a question is outside what it has been told. That last paragraph is the whole design, and it is what most website chatbots get wrong.

The failure everyone has seen

A chatbot confidently quotes a discount that does not exist, promises a delivery date nobody agreed, or argues with a customer about a refund. The model was not malicious. It was asked an open question with no boundaries and did what language models do: produce a plausible answer. The fix is not a better model. It is a better set of rules.

Five rules we build in

  1. 1.Approved content only. The assistant answers from a knowledge base the business has written and signed off: services, products, pricing language, process. Our knowledge base is the same copy you read on this site, so the two can never disagree.
  2. 2.Say what you are. It introduces itself as an AI and does not pretend otherwise when asked. Visitors deserve to know, and in practice they relax once they do.
  3. 3.Never invent what matters. Prices, timelines, guarantees and legal positions are out of bounds unless they appear verbatim in approved content. Ours knows the published budget bands and the shape of each service, and offers a call for anything more specific.
  4. 4.Hand over early. Complaints, legal or contractual questions, security concerns and anything the model is unsure about go to a person, with a clear sentence saying so. A graceful 'I would rather the team answered that' is worth more than a confident wrong answer.
  5. 5.Log everything and read it. Every conversation is recorded so the business can see what people ask, where the assistant struggled and what to add to the approved content next.

Letting it do things, carefully

An assistant that only talks is a glossary. Ours has two tools: it can read live availability from the booking calendar and offer times, and it can raise an enquiry so the team follows up. The important design choice is that before it creates anything it reads back the name, email and summary and waits for a yes. Actions that cannot be undone need a human confirmation, however small.

How it is built

Technically it is less exotic than it sounds. A system prompt carries the rules and the approved content, cached so the cost is dominated by what visitors actually type. The model streams its reply to the page. Tool calls are checked against strict schemas and clamped server-side before anything runs. If the AI provider is unreachable, the assistant says so and points to the form and the calendar rather than guessing. We build these on Anthropic's Claude; the pattern applies to any capable model.

What to ask before adding one to your site

  • What is the approved content, and who owns keeping it current?
  • Which questions must always go to a human?
  • What is the assistant allowed to do, and which of those actions can be undone?
  • Who reads the logs, and how often?
  • What does the visitor see when the AI is unavailable?

Done this way, an assistant is a well-behaved member of staff who has read everything you have published and knows when to fetch a colleague. If you want to see one working, click the dot and ask it something awkward.

Questions people ask

Short answers.

Which AI model should we use?

It matters less than the rules around it. We build on Anthropic's Claude for assistants that must stay inside approved content and hand over gracefully, but the design (grounding, identity, hand-over, logging) applies to any capable model.

What does an assistant cost to run?

Usage is metered per conversation and, for a small-business site, is usually a modest monthly figure. Caching the approved content keeps the cost dominated by what visitors actually type.

Will it replace our enquiry form?

No. It should make the form and the call easier to reach. Ours can book a slot or raise an enquiry, and it reads back the details before it sends anything.

Discovery call · 30 minutes · free

Pick a time. We’ll arrive prepared.

Times are shown in your time zone. You get a confirmation and calendar invite straight away, and a link to move it if life happens.