FAQ — What data does Self-Learning use?
A short answer to the most common question.
What gets used
In the regular Self-Learning feedback workflow, Self-Learning uses the written feedback your team reports on your tenant’s conversations:
- Reported AI responses from your conversations — reports submitted from the Conversations inbox, the Agent Stack test sandbox, and the Monitor.
- The comment text written on those reports, and the intent category when one is selected: wrong, missing information, too long, wrong tone, wrong phrasing, wrong specialist, knowledge gap. Manually selected categories are preserved when present; a report without a category may be classified later. (Other exists as the classifier’s own fallback and is never something a reader picks.)
- The message that was reported, and the specialist that produced it.
That’s it for the regular feedback workflow. No other conversation data feeds that proposal generator.
Answer ratings are not an input to it. The thumbs on an AI answer — including the ratings customers leave in the chat widget and the thumbs signed-in users leave in in-app Help — are counted as a signal, but a rating on its own never reaches the proposal generator. Chat-widget ratings are shown beside the answer in the Conversations inbox. In-app Help ratings are stored on the Atender help conversation and tied to the signed-in asker for traceability. They do not make the asker’s tenant learn from another tenant’s data. Only feedback carrying written text is classified and grouped, and these wordless ratings carry no text at all. Ratings are read in three places: that inbox display, cross-tenant superadmin analytics, and Atender Supervisor, where a thumbs-down is a detection signal in its own right — the heaviest-weighted one, so a rated conversation can be the reason a finding exists.
Self-Learning can also draft from handed-over conversations when the separate gated handover-mining workflow is enabled. This only runs when your tenant has the Draft from handovers toggle enabled and the required feature and environment gates are on. It looks for handed-over conversations where a human teammate provided a substantive new answer after handover, then creates reviewable drafts from that answer. These proposals are badged From a handover and do not go live automatically.
Owners and superadmins may also have access to Atender Supervisor, a separate gated workflow. Supervisor can analyze recent live AI conversations, create supervisor-sourced findings and proposals, and backtest those proposals before approval. Supervisor-sourced rows are kept separate from normal feedback groups and lists. Supervisor-generated results can also be read through tenant API keys or MCP connections that include the supervisor:read scope. Those reads stay tenant-scoped: the tenant is derived from the key, full Self-Learning opt-in is checked on each request, and the API can only read existing results — it cannot start or rerun analysis.
What does NOT get used
In the regular Self-Learning feedback workflow:
- Conversations from other tenants. Self-Learning is scoped to your tenant. Your data never shapes another tenant’s AI, and theirs never shapes yours.
- Unreported conversations. Conversations that nobody reports as needing improvement are never fed into the regular proposal pipeline. The AI doesn’t second-guess responses on its own in this workflow, except for explicitly enabled handover-mining workflows such as Draft from handovers.
- Customer PII for training. Proposals are generated from the substance of the issue, not raw customer identifiers. Anything matching your redaction rules on a capability is masked in that capability’s response, so it never enters the conversation in the first place. Note that this is where redaction happens — there is no separate masking pass inside Self-Learning itself.
Where do approved changes apply?
Only to your tenant’s stack. An approved prompt edit modifies your specialist agent. An approved KB-article proposal is published live in your Knowledge Base — it is not staged as a draft first. Nothing leaks across tenants.
Can I disable it?
Yes. Self-Learning access is set per tenant by an Atender superadmin, not from your own settings. Once it is off, no new feedback reports get processed and no new proposals are generated. Existing groups, proposals, and audit-log entries are preserved — you can re-enable later without losing history.
Who can see flagged content?
Anyone on your tenant with access to Self-Learning. By default, that’s owners and team leads. Review-only tenants can see Reports and History; full-access tenants can also see and act on Ready to publish. Ask your admin to adjust roles if you need to widen or narrow that.