Why didn't the AI answer my visitor's chat?
In live chat, the AI answers first and hands over to a person when it can't help. It's built to hand over rather than guess, so a chat that goes to your team is usually the AI doing its job. Here's how it decides, and how to get it to answer more.
First, read the hand-off note
When the AI hands a chat to your team because it couldn't answer, the ticket gets an internal note starting AI hand-off:. Open the chat's ticket and read it first. It shows:
- the visitor's question,
- what the AI searched for,
- which help articles it considered, and
- why none of them answered (for example they don't cover it or it needs a person).
If none of the articles it considered should have answered the question, that's a gap: write a help article for it, or add a line to the desk's Notes for the AI assistant (see below).
How the AI decides, every message
Each time the visitor writes, IntelliDesk:
Checks you've connected your own AI key under Settings → AI. No key, no AI answers.
Checks the limits: up to 20 AI answers per conversation, and 500 per workspace per day.
Finds the help articles that might answer it (see How it finds articles below), from your published articles only.
Asks Jev two things about the articles it found:
- Do these articles actually answer what the visitor asked?
- Is this something that needs a person anyway: a complaint, a refund or billing dispute, a cancellation, a legal or security matter, or an action on the customer's own account?
Jev also checks the message against your workspace's Always hand to a person topics (Settings → AI), and scores how relevant each article is: only the relevant ones (at most 3, plus your notes) are used, best first.
Writes the answer from those articles (and your notes) only. If the writer finds they don't cover it after all, it declines.
Checks the answer before the visitor sees it. Jev checks that every factual claim (steps, setting names, prices, times, policies) is backed by the articles or your notes, and that it answers what was asked. If not, the answer isn't sent.
If any step says no, what happens next depends on the desk's Who answers chats setting (in the desk's Live chat section):
- AI, then your team: the visitor is told the AI will get a person, and the chat moves to Waiting in the Live Chat console. (If nobody is online, the visitor is told so instead and can leave an email address for a reply.)
- AI only: the visitor is told their question is being passed to the support team, and the chat becomes a normal support ticket in your inbox, which your team answers like any other ticket.
How it finds articles
The AI first rewrites the visitor's question, using your own AI key: it fixes spelling, splits or expands words (such as helpdesk → help desk), and adds the words an article would likely use. It then searches article titles and text for each version, plus a meaning-based search over article sections, and merges the results.
So typos and different wording no longer stop it finding the right article. After you add or edit an article, the meaning-based search picks it up within about 10 minutes.
The common reasons, and what to do
No AI key connected. The chat AI only runs on your own AI key. IntelliDesk's platform AI isn't used to answer your customers. Connect one under Settings → AI. See Setting up AI: your own key, the platform AI and what each feature needs.
No help article covers it. This is the most common reason. Write an article that answers the question, and publish it: drafts are never used.
The article is brand new. An article you've just published or edited can take about 10 minutes to be picked up by the meaning-based search.
The article exists but doesn't really answer it. The AI only answers when an article actually covers the question. An article that mentions the topic in passing, or answers a slightly different question, isn't enough. The hand-off note shows which articles it considered and why it rejected them.
The answer check stopped it. The AI wrote an answer, but the check found it wasn't fully backed by your articles, so it wasn't sent. The hand-off note says An answer was drafted but not sent: the check found it wasn't fully backed by the articles, followed by the reason: (it included details the articles don't state) or (it answered a different question). For example, an article said proof of delivery "is available" but not where to get it; the AI guessed a location, and the check stopped it. What to do: write or clarify the article so it states the detail the visitor needs (here, where to get proof of delivery).
It's one of your "Always hand to a person" topics. If the message matches a topic listed under Always hand to a person in Settings → AI, the chat goes to your team, and the hand-off note says it's a topic your team always handles personally. This is by design for that topic. If you'd rather the AI answered it, remove the topic from the list.
It's a "needs a person" topic. Refunds, disputes, cancellations, complaints, legal and security questions, and anything about a customer's own account always go to a person, even if an article mentions them. This is deliberate.
The visitor asked for a person. They can do this at any time. From then on, a person owns the chat (on an AI only desk, it becomes a support ticket).
The chat hit a limit. After 20 AI answers in one conversation, or 500 across your workspace in a day, chats go to your team.
The AI provider had a problem. If your key is rejected, out of credit, or the provider is down, the chat goes to a person. Click Test in Settings → AI to check your key.
Once a person has taken over
The AI stops answering as soon as a chat is escalated. To hand it back, click Return to AI at the top of the chat. The AI answers the visitor's next message.
Notes for the AI assistant
Some things don't deserve a whole article: a term your customers use differently, or a rule about what not to promise. Put those in the desk's Notes for the AI assistant: under Settings → Help Desks → your desk, in the Live chat section (visible when live chat is on). You have up to 2,000 characters; click Save notes. For example:
- A desk is the same as a help desk. Each desk can have its own mailbox.
- Never quote delivery times for Zimbabwe; say a person will confirm.
The AI treats these as trusted material alongside your help articles, for live chat answers and AI-written acknowledgement emails. It still answers only from the notes and the articles. Only the desk's managers can edit the notes, and saved changes are recorded in the Audit Log.
For things that apply to every desk (your tone, what your business does, the words you use, and topics that always go to a person), use the workspace AI profile under Settings → AI, card How the AI talks about your business. Desk notes add to it. See Setting up AI: your own key, the platform AI and what each feature needs.
Getting the AI to answer more
- Write for the questions people actually ask. Read the hand-off notes on chats your team picked up and turn the common answers into articles.
- Title articles as questions, in the customer's words: How do I reset my password? rather than Password policy.
- One topic per article, answer first. The AI uses at most the 3 most relevant articles, so short, focused articles work best.
- Say exactly where and how. The answer check stops answers with details the article doesn't state, so spell out the steps, setting names and places a customer needs.
- Keep them accurate and current. The AI repeats what the article says, including out-of-date prices or hours: a wrong article gives a confidently wrong answer.
All published articles are public on your help centre, and all of them are available to the AI. There's no way to publish an article for the AI only; use the desk's Notes for the AI assistant for short standing facts instead.
Seeing how it's going
Reports → AI & automation shows Chats resolved by AI alone and AI replies sent, and Reports → Live chat covers chat volume. See The AI & automation report: what Jev did and how much time it saved.
Related: Live chat: setting it up and answering chats · Your knowledge base: help articles for customers, agents and the AI