TractionScout
← Back to the blog
Lifecycle messaging

AI support chat: what it should answer, and when to hand over to a person

Let it answer what your pages already answer, cite its sources, and pass everything else to your team with the context they need

7 Nov 2026 · 5 min read

An AI support chat should answer questions your own help pages and site already answer, show where each answer came from, and hand everything else to a person: billing disputes, account changes, suspected bugs, anything it could not answer after a couple of tries, and topics you decide always need a human. A good handover keeps the conversation, a draft reply and a way to reach the visitor.

What to let it answer, and what to hand over

Let it answerHand to a personKeep out of the chat entirely
How-to questions your help pages coverBilling disputes and refund requestsPasswords, card numbers, secrets
What a plan includes, as your pricing page saysAccount changes, such as ownership or deletionLegal or medical advice
Where a setting lives, with a linkSuspected bugs and outages affecting this visitorPromises your pages do not make
The visitor's own billing status, if signed in and verifiedAnything it could not answer in two turns
Topics you mark as "always a person"

The left column is where AI saves your team real time: the same ten questions, asked a hundred ways, all answered in your docs. The middle column is where a confident but wrong answer costs a customer.

Five rules for answers people can trust

  1. Answer only from your own sources. Help pages, your site and question-and-answer pairs you wrote. If the answer is not there, the chat should say so, not improvise.
  2. Cite where the answer came from. A link to the help page lets the visitor check, and lets you see which page did the work.
  3. Check before showing. A second pass that compares the draft with its sources catches the confident sentence nobody wrote.
  4. Act only on the visitor's own data. Looking up a signed-in visitor's own subscription is helpful; anything that touches someone else's account is not the chat's job.
  5. Fail toward a person. When the model is unavailable or keeps failing, route visitors to your team rather than showing errors or guesses.

What a good handover contains

A handover is not "please email support". It should carry the full conversation, the visitor's email address so you can reply, a short summary of what they need, and a draft reply your team can edit. The visitor should not have to repeat themselves, and your team should be able to answer in one step.

Unanswered questions are your next help pages

Every question the chat could not answer tells you something your pages do not say. Review them weekly. Some become a new answer in the chat; others deserve a help page or a blog post, which also helps search engines and AI assistants find you. A chat that learns only what you approve keeps mistakes from spreading.

Two numbers tell you whether it works: how many questions it answered without help, and how many it sent to your team. A rising share sent to the team after a release is often the first sign that something changed for customers.

How TractionScout's support chat works

TractionScout Work queue on sample data: items that need you, are ready to ship, in progress or blocked, with the selected item's details and an Approve button

The Work queue in the TractionScout demo workspace on sample data. Handed-over support cases land here beside other work to review; the sample has none.

TractionScout's support chat is set up in Configure › Data Sources › Support chat: Scout reads your site and help pages, drafts what it should know and puts the chat live, and you can change everything afterwards. Sites that already run TractionScout tracking show the chat automatically; other sites paste one snippet.

Scout answers from your own answers, help pages and site text, cites its sources and gets one correction pass. For signed-in visitors it can look up their own billing through Stripe and check recent crashes. Topics you list always go to a person; after two unanswered turns the chat asks for an email and opens a case; if the model fails repeatedly it hands visitors to your team for a few minutes. Cases land in Work with the transcript, a case draft and suggested steps, and you reply with Send reply or Send and resolve.

Questions it could not answer and answers from resolved cases become answers to review; nothing is used until you add it. Conversations are kept for 90 days. See AI support chat.

Questions

Will an AI support chat make up answers?

It can, if it answers from general knowledge. Limit it to your own sources, require citations and a check before showing, and have it say "I don't know" and hand over when the sources are silent.

Should the chat handle refunds?

It can explain your refund policy as your pages state it, but the decision and the refund itself belong to a person. Mark refunds as a topic that always goes to your team.

How quickly should a person reply after a handover?

Say so in the chat. Promising a reply time you keep, such as "within one business day", builds more trust than an instant answer that is wrong.

Empfohlen

Put a support chat on your site that answers from your own pages

Cited answers from your help pages, and a clean handover to your team for everything else.