Teach the business with the Learning Inbox

A correction becomes useful beyond one conversation when the team agrees it should guide future work. The Learning Inbox lets you propose and review reusable guidance with its scope and source, so future work can use an explicit decision rather than an informal assumption.

Review a proposal

1. Open the Learning Inbox and read the rule, its rationale, and its confidence.
2. Check conflicts and affected workers. A proposal that contradicts an approved policy needs scrutiny, not speed.
3. Choose an outcome. Send to approvals routes it through the normal approval queue; This conversation keeps it local to the current chat; Reject or Ignore ends it with a recorded receipt. Nothing becomes shared truth without approval.
4. After approval, the learning persists as shared policy with its authority tier and provenance. Revisit it any time; superseded entries link forward to their replacement.

Conversation-derived content always starts at the lowest authority tier. Official policies outrank approved learnings, which outrank working notes. Retrieval shows the model what to believe, so a casual chat message can never silently override an approved rule.

Propose from everyday work

Use Propose learning for anything you find yourself correcting twice: positioning, workflow preferences, offerings, messaging, or process rules. Give it a confidence and the workers it affects. The assistant can also propose learnings conversationally through the same governed path; those proposals wait in the inbox like any other.

Approved rules are supplied to interactive assistants, scheduled coworkers, reply drafting, proposal drafting and content briefs when their scope matches. Independent rules remain active together, while an explicit replacement preserves the older rule in history.

Add the reference behind the rule

Open Business references in Learning Inbox, choose a PDF, DOCX, text or Markdown file, and confirm that workspace members and agents may use it. Uploads accept up to 4 MB; PDFs support up to 100 pages, and extracted text is limited to 500,000 characters. The file stays in private workspace storage while the work engine extracts its text. Once it shows Searchable, ask the assistant a specific question about its contents and inspect the returned source.

For example, Northline Roofing could upload its appointment policy, then ask which information a customer should provide before a site visit. Search also draws from customer records, notes, conversations and authorized indexed Drive documents. This baseline uses text search, so it needs no embedding service or additional AI key. Your database, storage and any model calls still use your installation's resources.

Capture a correction while replying

Apply a suggested reply in Conversations, then edit its wording. The correction control identifies changed sentences and lets you describe the general rule worth keeping. Choose Save correction for review to create a proposal, then use the normal Learning Inbox and approval steps. Keep customer-specific facts in the conversation so a one-off instruction does not become guidance for everyone.

Learning evidence separates corrections, rejected actions, missing sources and execution failures. The work engine records each new receipt, creates a review task for execution failures, and records the proposed remedy. A correction that cannot be reviewed appears first under Learning evidence, with its original rule and recovery instructions, while other evidence continues through review. Successful actions remain useful evidence; compare the original failing example with later results before judging whether an approved correction improved the work.

Review what the AI remembers

Coworkers and the assistant keep short notes between runs: prior work, research, scheduled checks and open questions. Open What your AI remembers in the Learning Inbox to see each note with its coworker, how long it is kept and a link to the run that created it. Filter by type or search the text. Choose Correct to fix wording that is wrong; the previous text stays in the audit log. Choose Remove and confirm to stop future runs from seeing a note.

Each run records the memories it was given. Open the run from AI runs and look for Memory used and Context used in its trace to see which notes, rules and evidence shaped the answer.

What to check

If a document stays queued, inspect the work engine and scheduler. A failed extraction offers Retry; scanned PDFs need OCR before upload. Archive removes a reference from later searches while retaining its history. Drive results are withheld when current access or the indexed revision cannot be verified, including a workspace disconnect or folder removal during the search; reconnect or sync Drive, then repeat the query. If the assistant reports incomplete source coverage, retry after restoring the connection before treating an empty result as proof that the information is missing.

When evidence says recovery required, check its source and plugin access. The work engine retains the original input and retries after a 15-minute cooldown; keep the scheduler running, then refresh evidence to inspect the result. If an item instead needs manual review because its original input is incomplete, inspect the source and submit the correction again through its original workflow.

After approving a rule, run a relevant task and inspect its context receipt and actual result. The receipt proves which guidance was supplied, while reviewing the result establishes whether it helped. Rules with unknown older scopes stay out of context until reviewed. If Earlier rules need review appears, review each suggestion before proposing it again.

The fictional demo supports proposal and approval practice. Use your own workspace to verify private uploads, scheduled indexing, model execution and real outcomes. When upgrading the inbound responder, review and approve its current policy version in Setup before automated replies resume.