An AI knowledge base is the set of approved documents a support chatbot searches before it replies, so every answer comes from your own policies with a link to its source, and anything it does not cover goes to a person. With Praxen, building one and connecting a chatbot on one channel is a 2–4 week pilot costing £2,900 (£2,030 with the launch offer), plus AI usage billed at cost on your own accounts. Prices exclude VAT.
What an AI knowledge base does
A support chatbot built this way works in two steps. First it searches your knowledge base for the passages that match the customer’s question. Then it writes a reply using only those passages and shows where the answer came from. This is called retrieval-augmented generation, and a chatbot that works like this is often called a RAG chatbot.
The practical consequence is simple. The quality of the answers depends mostly on the documents. A clear, current knowledge base gives clear, current answers. Old or contradictory documents give wrong ones, however good the model is.
What goes in and what stays out
Start from what customers actually ask. Export last month’s tickets, chats and emails and group them by question. The largest groups show what the knowledge base must cover first. Then collect the content that answers them.
| Content | In the knowledge base? | Notes |
|---|---|---|
| Help centre articles | Yes | Check prices, dates and screenshots first |
| Returns, delivery and warranty policies | Yes | The current version only, with the date it took effect |
| Price lists and product details | Yes | One source file, updated whenever prices change |
| Opening hours and contact routes | Yes | Including holiday hours and which team handles what |
| Good past replies from your team | Yes, as examples | Names, order numbers and other personal data removed |
| Internal notes and draft policies | No | They often contradict the published version |
| Customer and order records | No | Looked up live through a separate, read-only connection |
How answers cite their sources
- Each document is split into short sections, each with a title and a stable link. The chatbot quotes from a section and links to it.
- The customer sees the source under the answer, for example “Returns policy, section 3”, and can open it.
- Prices, dates and time limits are given exactly as written in the source. The chatbot does not round, estimate or combine them.
- Answers about a specific order combine a passage with live data, for example the returns policy and the delivery date of that order. Both are shown, so the customer can see why the answer applies to them.
- If two documents disagree, the chatbot does not pick one. It hands over and logs the conflict, so the owner can fix the documents.
- If the passages it finds do not answer the question, it says so and passes the conversation to a person.
- In the weekly review, your team sees the passage behind every answer, which makes a wrong answer quick to trace.
Keeping the knowledge base up to date
A common cause of wrong answers after launch is a document that was not updated. These habits prevent it.
- Give each document a named owner and a review date. When the review date passes, the owner gets a reminder.
- Add “update the knowledge base” to the checklist for every price, policy or product change, on the same day the change takes effect.
- Give seasonal content, such as holiday delivery cut-offs, a start and an end date, so it switches off on its own.
- Write articles in the words customers use. If customers ask about “sending something back”, the returns article should use that phrase too, so the search finds it.
- Keep a log of questions the chatbot could not answer. Each week the owner decides which ones need a new or corrected article.
- Archive old versions instead of deleting them. You can then see what the chatbot said on a given date if a customer disputes it.
When the chatbot hands over to a person
Handover rules sit next to the knowledge base and are tested with it. The conversation moves to your helpdesk, such as Zendesk, Intercom or Freshdesk, with a summary, the transcript and the sources the chatbot used, so the agent can continue without asking the customer to repeat anything. These are the usual triggers:
- The customer asks for a person, in any words.
- The answer is not in the knowledge base, or two documents conflict.
- Two replies in a row have not resolved the question.
- The message mentions a complaint, legal action or a chargeback, shows strong frustration or suggests the customer may be vulnerable.
- The request involves money: a refund outside the policy, a discount or a change to payment details. People approve all of these.
How to test it before launch
The test shows whether the knowledge base and the handover rules work together. It runs in the second half of the pilot, before any customer sees the chatbot.
- Build a test set from last month’s real questions, with the answer your team gave to each.
- Add awkward cases: questions the knowledge base does not cover, two questions in one message, an angry message, a request for a discount and a question about someone else’s order.
- Run the full set and mark each reply as correct, incomplete, wrong or correctly handed over.
- When a reply is wrong, find the passage it used. Usually the fix belongs in the document.
- Agree in advance what result is good enough to launch, and let the team that handles support today make the call.
- Run the full set again after every large update. In the first month after launch, review a sample of real conversations each week.
Cost, timeline and data
A pilot that builds the knowledge base and connects a chatbot on one channel, usually website chat or email, costs £2,900 (£2,030 with the launch offer) and takes 2–4 weeks, including the test set. Adding more channels and agents that look up orders or start returns starts at £6,500 (£4,550). Ongoing care is £350 a month (£245) and covers document updates and the weekly review. AI usage is billed by the providers at cost on your own accounts. Prices exclude VAT.
The knowledge base and the conversations stay in your own accounts. The chatbot runs on enterprise APIs that do not train on your data, customer records are read only when needed, and we sign a data processing agreement before launch.