Tuning the AI's answers
The support AI is not one behaviour — it is several settings that decide when it answers, how much it trusts itself, and when it steps aside. This page covers the tuning that is not covered by response modes and handoff.
The controls that set its confidence
In Settings → Support → AI assistant → Setup:
Minimum confidence — the bar a potential answer must clear before the AI sends it. This is the single most useful dial you have. Raise it and the AI answers less but is more often right; lower it and it answers more, including where it should have handed off.
Response mode — whether it replies directly, or works in a more conservative mode. See AI response modes & follow-up limits for what each mode means in the thread.
Reply channels — which channels the AI may answer on. Restricting it to chat while email stays human is a common, sensible split: email tends to carry the complex or contractual conversations.
Support agent — which configured agent answers. Behaviour follows the agent's instructions and tools, so changing the agent changes the answers even when nothing else moves.
Check the code before replying by email — a guard for answers that involve code or configuration, checking it before the reply goes out on email.
Preferred answers
The Preferred answers tab is where you steer behaviour when several answers are defensible: which source or which style the AI should favour. Use it when the AI is technically correct but answers in the wrong register — too verbose, too terse, too eager to escalate.
Reading its behaviour as a signal
Treat the AI's numbers as a diagnosis, not a score:
High involvement with a low resolution rate — it is engaging and failing. Likely a knowledge gap, or a confidence bar set so low it answers things it should hand over.
Low involvement — the bar may be too high, the enabled channels are too narrow, or the assistant is off for the channel traffic actually arrives on.
Frequent handoffs on the same topic — a documentation problem, not a model problem. Fix the article before touching the confidence setting.
Answers that are right but disliked — a preferred-answers and tone problem.
Change one setting at a time and give it enough volume to show an effect. Adjusting confidence and preferred answers together tells you nothing about which one mattered.
The AI's own follow-ups
Separately from the follow-ups a teammate sets by hand, the AI can chase quiet conversations. On the Handoff & follow-up tab, Follow up on quiet conversations sets a first check-in, a final follow-up, and when to close the conversation, all in hours.
That means a conversation you leave alone is not necessarily abandoned — the AI may check in and eventually close it. Two consequences:
Know the cadence before you rely on it. A 24-hour check-in and a 1-hour close behave very differently for a customer who replies slowly.
A conversation closing itself is not a resolution. If the AI closes quiet conversations, watch the resolved counts and CSAT together, because "closed because nobody replied" and "closed because it was fixed" look identical in a raw count.
Related pages
Guidance, not just steering
The Preferred answers tab is the guidance layer: short instruction items with a title, the question patterns they apply to, and the answer or behaviour to use. Each can be scoped to a language and switched on or off individually.
Guidance shapes how the AI answers, not what it knows — it cannot make the AI answer from content that is not in your knowledge sources, and it does not override hard safety escalations. See Support AI assistant configuration for how to write items that actually fire.
Was this article helpful?