A chat assistant that answers questions using your own documentation, help articles and product pages is one of the most common AI features businesses ask for. Done well, it helps visitors find answers faster and takes pressure off support. Done badly, it confidently tells customers things that are not true, in your name.
This post explains how these assistants work, where the effort really goes, and what to decide before you build one.
How it works
Most assistants of this kind use an approach called retrieval-augmented generation. The name is more complicated than the idea:
- Your content is split into small passages, such as sections of help articles.
- Each passage is converted into an embedding, a list of numbers that represents its meaning, and stored in a search index.
- When a visitor asks a question, the system finds the passages most related to it.
- Those passages are sent to a language model along with the question and instructions, such as "answer only from the provided content and say so if the answer is not there".
- The model writes an answer, ideally with links to the pages it used.
The model is not trained on your content. It reads the relevant pieces each time a question is asked. That means updating an answer is as simple as updating the page, and it means the quality of your content directly limits the quality of the answers.
Where the effort actually goes
Connecting a model to a search index is the quick part. The work that decides whether the assistant is useful is elsewhere.
Content
If your help articles contradict each other, are out of date or never cover the questions customers actually ask, the assistant will reflect that. Most projects start with a content cleanup, and that cleanup alone often improves the site for every visitor.
Scope
Decide what the assistant is for and, just as important, what it should refuse. An assistant that answers product questions should probably not discuss pricing exceptions, give legal or medical advice, or comment on competitors. Clear limits make it more trustworthy.
Evaluation
Before launch, collect a set of real questions from support tickets, search logs and sales calls, along with the correct answers. Run the assistant against them and review the results. Repeat after every significant change to content, prompts or models. Without this, you are guessing whether changes made things better or worse.
Handoff
Some questions need a person. The assistant should make it easy to reach one, and pass along the conversation so the customer does not have to repeat themselves.
How it fails
Confident wrong answers
Language models can produce fluent answers that are not supported by your content, especially when the retrieved passages are close to the question but not quite on it. Requiring citations and instructing the model to say "I don't know" reduces this. It does not eliminate it.
Companies are generally held responsible for what their assistants tell customers. In a widely reported February 2024 decision in Canada, a tribunal ordered Air Canada to compensate a customer who had relied on its website chatbot's incorrect description of the airline's bereavement fare policy. The tribunal rejected the argument that the chatbot was responsible for its own statements. Plan as if every answer were a statement from your company, because it is.
Outdated information
If an old page is still in the index, the assistant will use it. Removing or updating content needs to update the index too, and that should be automatic.
Prompt injection
Visitors can type instructions designed to make the assistant ignore its rules, reveal its instructions or say something embarrassing. Content the assistant reads can contain hidden instructions too. Limit what the assistant can do, never give it access to private data or actions it does not need, and test it with adversarial questions before launch.
Cost and privacy
Running costs are usually driven by usage: each question involves a search and a model call. For most business sites these costs are modest, but they are not zero, so set budgets, rate limits and alerts. Bots and abusive users can drive usage up quickly.
On privacy, assume visitors will type personal information into the chat, even if you ask them not to. Choose providers whose terms fit your obligations, decide how long conversations are stored, mention the assistant in your privacy policy, and avoid sending conversation data anywhere it does not need to go.
The chat widget itself also needs to meet the same accessibility standards as the rest of your site. Many off-the-shelf widgets do not.
Should you build one?
An assistant is worth considering when:
- You have a substantial body of good content that visitors struggle to navigate.
- Support receives many questions that are already answered on your site.
- You can commit someone to reviewing conversations and maintaining content after launch.
It is probably not worth it when your site is small, your content is thin or outdated, or nobody will own it after launch. In those cases, better navigation, clearer pages and a well-tuned site search will help visitors more, for less money.
If you do build one, start narrow: one product, one audience, one well-maintained set of content. Measure whether it answers correctly and whether people use it. Expand from there.
Thinking about an AI assistant for your site? Contact us and we will help you decide whether it is worth building and how to scope it.