What is Self-Learning?

Self-Learning is how your Agent Stack gets better over time without you rewriting prompts by hand. When something the AI said wasn’t quite right, anyone on your team flags the response and adds a note. Atender groups similar flags together, generates concrete proposed fixes, and lets you approve and publish those fixes from one place.

The result: a feedback loop that turns “the AI keeps getting this wrong” into a system-prompt edit, a routing tweak, or a new knowledge article — approved by you, then published to your stack in one deliberate step.

How the loop works

1. Report — anyone reviewing a conversation reports an AI response that missed the mark. There are two ways in, both in the conversation itself rather than in Settings → Self-Learning. Answer thumbs — Good answer / Bad answer — sit directly on Agent Stack answers in the Conversations inbox and on the Live transcript card of the Agent Stack testing panel; a thumbs-down opens a What was wrong? popover with reason chips and an optional detail box. The heavier path is Review this conversation, which annotates the thread as a whole. Reason chips are: Wrong, Missing information, Too long, Wrong tone, Wrong phrasing, Wrong specialist, Knowledge gap. Atender picks the feedback mode itself from who was involved in the conversation, and only asks you to choose when both the AI and a human were involved in it.
2. Group — Atender clusters semantically similar reports into groups. If five different customers all asked about returns and the AI gave a fuzzy answer five times, those five reports collapse into one group with a clear summary.
3. Propose — for each group, Atender generates one or more proposals — concrete changes you could apply. Possible proposal types are summarized below.
4. Approve, then publish or stage — you review each proposal under Reports, and approve or reject it. What happens on approval depends on the surface, proposal type, and your permission. Eligible one-card fixes can show Approve and publish, which applies the live fix immediately. Other approvals stage the proposal in the Ready to publish queue, where deployment happens when someone presses Publish all. Published changes are recorded under History. Live prompt, routing and model-setting fixes can be rolled back from the proposal’s own card; when you roll one back, the fix reappears in review so you can edit it, approve it again, reject it, or dismiss it.

Related: Atender Supervisor adds an owner- and superadmin-only workspace for deeper review. The standard Self-Learning tabs remain flag-driven, while Supervisor can run Analyze now, cluster detected failures, draft fixes, backtest them, and send supervisor-sourced proposals through the same approval queue. Supervisor checks both knowledge base articles and Handbook procedures when deciding whether an answer already exists. If a matching Handbook procedure is written but hidden, Supervisor reports it as written-but-hidden because the Agent Stack only reads visible Handbook entries. Supervisor-generated KB proposals only state grounded facts; if facts are missing, Supervisor shows a Waiting for your answers state. The reviewer answers one field per missing fact, then clicks Rewrite the article. Supervisor rewrites and rechecks the article before Apply or Publish becomes available, and any unanswered facts are carried forward rather than silently cleared.

What can be proposed

• Prompt edit — Updates a specialist agent’s system prompt to handle the recurring case better.
• KB article — Suggests a grounded new or improved knowledge base article that fills a recurring knowledge gap; if key facts are missing, it creates a reviewable incomplete draft with a missing-facts list instead of a publishable article.
• Handbook article — Suggests a new internal-handbook entry or revisions to an existing procedure to guide the AI’s behavior on the topic.
• Routing update — Adjusts the router so similar future questions reach the right specialist.
• Temperature adjustment — Tweaks the AI’s creativity / variance for the specialist when its style is off.

Each proposal includes a diff summary showing exactly what would change before you approve it — no black-box updates.

Three tabs in Settings → Self-Learning

• Reports — Browse the clustered feedback. Each group shows what’s been reported and how often. Drill in to see individual annotations and act on each proposal. Depending on the surface, proposal type, and your permission, approval may either publish an eligible one-card fix immediately with Approve and publish, or stage the proposal for later publishing.
• Ready to publish — Proposals you have already approved but not yet deployed. One Publish all button ships the whole queue.
• History — A log of what happened — timestamp, actor, action, target. Rolling back a live fix happens on the proposal’s own card, not here. A rolled-back fix returns to review so you can edit it, approve it again, reject it, or dismiss it.

The whole section sits behind the settings.ai permission. Access is one of three tenant levels — off, feedback-only, or full — and there is no “limited” tier. If your tenant is on full and you hold supervisor.manage, the Reports tab disappears here and review moves to the Supervisor page instead.

How it relates to manual flagging

Self-Learning shares the same flagging mechanism across live conversations and the Agent Stack Live transcript; the test sandbox’s own Flag this reply for tuning runs the separate Agent Stack tuning workflow instead. A flag from a real customer conversation, a rating on a test call’s Live transcript, and a flag from a Monitor review all flow into the same Feedback → Group → Proposal pipeline. The more flags you collect, the better the groups and proposals get.

Can I disable it?

Self-Learning access does not gate the answer thumbs. Anyone who can open the conversation, or the Agent Stack, can rate an answer whatever the tenant’s access level — the thumbs answer to the surface they sit on, not to the tier. A rating that carries written text is classified and grouped like any other report; a wordless one is only counted.

Everything heavier is gated. Writing a full Review this conversation annotation needs at least feedback-only access, as does browsing Reports; approving proposals or pressing Publish all needs full access. Setting the tenant to off stops all of it.

Supervisor needs two things, not one: full Self-Learning access and the supervisor.manage permission, which by default only an owner holds.

What it’s not

• Not automatic — proposals don’t auto-deploy. Every change goes through your approval, and staged changes also go through an explicit publish.
• Not cross-tenant — the AI never learns from other tenants’ data. Improvements made in your stack apply only to your stack.
• Not retroactive — past conversations aren’t re-answered for the customer under new prompts. Changes apply going forward.

Where to start

• See it in action: open Settings → Self-Learning and browse the Reports tab — even on a fresh stack, the empty state explains the workflow.
• Flag your first response: Flag and tune bad responses
• Approve a proposal: Review and approve a proposal