How an AI assistant on your site actually answers a customer

Most AI assistants on store and service websites today work the same way, whatever the logo on the chat bubble. Intercom Fin, Zendesk AI agents, Tidio Lyro, Chatbase and the assistants built on them all follow one pattern, called retrieval-augmented generation. The name is clumsy, the idea is simple: before the model writes an answer, the system looks up the parts of your own content that seem relevant, and hands them to the model with the question.

If you know the steps, you know where to look when an answer is wrong. This note walks through them with a made-up example: a car rental company with three branches, a website, and an assistant on it.

Step 1. Your pages become a pile of snippets

When you connect the assistant, it reads the sources you point it to: the website, a help centre, PDFs, sometimes past support tickets. It cuts them into short pieces, a paragraph or a few sentences each, and stores each piece with a numeric fingerprint of its meaning (an embedding).

Our rental company points the assistant at its whole site. The page “Young drivers” says drivers aged 21 to 24 pay a surcharge of €12 a day. The FAQ says the minimum age is 21. An old blog post from 2023, still online, says “drivers under 25 can’t rent our premium cars”. All three go into the pile.

What the assistant knows is exactly that pile. Not your intentions, not the policy you changed last week, not what your staff say on the phone.

Step 2. The question is matched against the pile

A customer asks: “I’m 23, can I rent a car with you?”

The system turns the question into the same kind of fingerprint and pulls out the snippets that look closest. Usually that is three to ten pieces. It does not read your whole site for every question; it reads what the search returned.

This is the first place answers go wrong. If the search returns the 2023 blog post and not the “Young drivers” page, the model never sees the surcharge. If it returns both, it has to choose.

Step 3. The model writes an answer from those snippets

The model gets an instruction from the platform (“answer only from the provided content, be friendly, be brief”), your own instructions if you wrote any, the snippets, and the question. It writes a reply.

A good reply here: “Yes. Drivers aged 21 to 24 can rent with us and pay a young driver surcharge of €12 a day. Some premium models may not be available.”

A bad but plausible reply: “Unfortunately, drivers under 25 can’t rent our cars.” The model summarised the old post too broadly. Nothing in the logs will flag this. The customer leaves, and the conversation looks finished.

Step 4. The platform decides whether to answer at all

Most platforms add a last check. If the snippets don’t seem to answer the question, the assistant says it doesn’t know and offers a hand-over to a person or an email form. Intercom, for example, describes how Fin only answers from the content you give it and hands over when it can’t find an answer.¹

This check is why a well-set assistant says “I don’t know” more than you expect. It is also why a badly fed one says “I don’t know” to questions your FAQ answers word for word: the right snippet was there, but the search didn’t find it.

Where it goes wrong, in practice

Where	What you see	Usual cause
The pile	Old prices, an old policy, a fact you removed	Outdated pages, PDFs or blog posts still connected
The search	“I don’t know” on a question your FAQ answers	The answer sits in a table, an image, a tab or a long page the search misses
The writing	A confident answer that mixes two rules	Two snippets disagree, or one is about a different product
The hand-over	A customer asked for an email address before getting any answer	The platform’s rules, not the model

None of these show up as errors on a dashboard. The platforms report how many conversations the assistant “resolved”, meaning the customer didn’t ask for a person. A customer who got a wrong answer and left counts as resolved too. That is why we check answers against your pages, not conversation counts.

What this means for you

You don’t need to understand embeddings to run a good assistant. You need three habits.

1. Know what is in the pile. List every source connected to the assistant. Remove what is old. If a blog post contradicts a policy page, the post goes or gets a note.
2. Put rules where search finds them. One rule, one short paragraph, in plain text. Not inside an image, not in a table split across tabs, not in a PDF from 2022.
3. Ask it what customers ask, and compare with your pages. Not once at launch: every time a page changes. A wrong answer only counts when you can show the page that says otherwise, which also makes it easy to fix.

The car rental company in our example fixes its problem in ten minutes: it deletes the 2023 post, and the “Young drivers” page becomes the only source on age. The hard part was noticing.

Further reading

• Intercom’s own explanation of how Fin finds and writes answers.¹
• Zendesk’s overview of how its AI agents use help centre content.²
• The original research paper on retrieval-augmented generation, if you want the mechanism in full.³

If you’d rather someone else set the assistant up on your pages and keep checking it, that is what Assistant Setup and Assistant Care are.

Footnotes

1. Intercom, “How Fin works”, intercom.com/help. ↩ ↩²

2. Zendesk, “About AI agents”, support.zendesk.com. ↩

3. Lewis et al., “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks”, 2020, arxiv.org/abs/2005.11401. ↩