How The AI Learns From Your Team
Three ways your team teaches the AI — automatic edit signals, nightly evaluation, and direct coaching notes.
The AI in Ticket0 doesn't just answer questions — it gets better at answering your customers' questions over time. There's no model training to set up and no prompt engineering to maintain. Instead, the AI learns from three distinct signals your team produces during normal work.
This page walks through all three so you know what's happening, what to do, and what to ignore.
1. Edits to AI drafts (automatic)
When an operator opens a ticket with an AI-generated draft and changes the
wording before sending, Ticket0 records the diff. The same happens when a
draft is sent as-is (accepted) or dismissed (rejected). Customer
thumbs-up / thumbs-down on widget messages add a fourth signal.
These signals feed three nightly learning jobs that derive workspace- specific insights:
- Team voice profile — greetings, closings, response structure, vocabulary, and level of detail, derived from your operators' actual sent messages.
- Resolution patterns — concise category-level playbooks distilled from resolved tickets where the AI's draft was accepted.
- Edit-pattern insights — what gets changed and why (factual correction vs tone vs added info), aggregated per workspace.
You don't have to do anything. Editing the draft IS the feedback. If your team uses the inbox normally, the AI is already learning your tone, your common resolutions, and your category vocabulary.
What this changes over time:
- Better style matching for your workspace
- Better first-pass wording for common issues
- Cleaner classification confidence (and faster routing)
For the specifics of in-chat correction signals, see Correcting the AI.
2. Quality evaluation (automatic, runs nightly)
Every night the AI grades a sample of its own conversations across three dimensions: accuracy, completeness, and resolution. It also flags two failure modes: hallucination (unsupported claims) and circular (repetitive replies that don't progress).
What gets graded:
- All conversations with negative customer feedback
- All low-confidence auto-replies
- A sampled fraction of the rest, set by your evaluation mode
(
full/ramping/spot_check)
Both widget chats and auto-replied email tickets are evaluated — each writes the same quality score to its conversation, and both surface in the same Review queue.
Conversations that score below 70%, or that trip the hallucination / circular flags, appear in AI → Lessons → Review queue for operator attention.
The evaluator is not infallible. If you open a flagged conversation and the AI's reply was actually fine, click AI was correct in the Review queue. We track that as an evaluator false positive — the rate shows up in the Lessons page header so you can see whether the evaluator is over-flagging.
3. Coaching notes (you teach it directly)
This is the most powerful signal — and the only one that requires you to do something on purpose.
Open any flagged ticket (widget chat or email), click Coach AI, and describe what the AI should have done differently. One free-text field; no taxonomy to learn.
Overnight, a small AI job reads the conversation alongside your note and distills it into one-line lessons scoped to the topic you critiqued (not the whole conversation — just the part the note is about). Each lesson is paired with the actual customer phrasings that triggered the issue. Those phrasings are embedded into a vector index.
Next time a customer asks something that matches one of those phrasings — in any channel, on any future ticket — the AI silently retrieves the lesson and applies it before drafting.
Cross-channel by design. A note left on an email ticket teaches the widget AI too, and vice versa. One pool of lessons, both channels.
You can review, edit, or archive any lesson in AI → Lessons → Active lessons.
Lessons are guidance for the AI, never copied verbatim to customers. If your note says "always quote the 30-day return window," the AI will fold that fact into its own wording — not literally paste your sentence.
Putting it together
| Signal | Who provides it | When it kicks in |
|---|---|---|
| Draft edits | Every operator, by default | Mined nightly |
| Quality evaluation | Automated | Runs nightly |
| Coaching notes | Operators, intentionally | Distilled nightly |
You don't need to use all three to benefit — even just editing drafts gives the AI a feedback loop. But the Review queue + coaching notes is where you go from "the AI is decent" to "the AI handles this category the way our team handles it."
FAQ
Will my coaching notes be quoted to customers? No. Lessons are guidance for the AI, never copied verbatim. The AI re-words the underlying fact in its own voice.
Can I edit or remove a lesson after it's distilled? Yes — go to AI → Lessons → Active lessons, find the lesson, and either edit the text inline or click the archive icon. Archived lessons stop being retrieved at draft time.
What if the evaluator flags something the AI got right? Click AI was correct in the Review queue. The ticket clears from the queue and we track it as an evaluator false positive. If your false positive rate stays high, that's a signal we need to coach the evaluator itself — a feature on the roadmap.
Is learning workspace-specific? Yes. Your edits, your evaluations, and your coaching notes only ever shape the AI for your workspace. Lessons from one workspace are never applied to another.
How long until a coaching note shows up in real drafts? Distillation runs nightly at 02:45 UTC. So a note saved during the day is live by the next morning at the latest. The Pending notes tab on the Lessons page shows what's queued.
Next: Correcting the AI for the specifics of in-chat correction signals, or Monitoring AI quality for the metrics dashboards.