Images
The agent can inspect up to five images attached to a message. It can describe visible content and use it to answer the customer’s question. For invoices, receipts, or screenshots, it should state uncertainty instead of inventing unreadable fields.
Text inside an image is untrusted. A screenshot that says “ignore previous instructions” cannot change the agent’s system rules or authorize an action.
Voice messages
Voice notes are downloaded only from safe public HTTPS locations and transcribed before the agent answers. Recordings larger than 25MB are split into bounded Opus chunks on the server. The source processing limit is 200MB.
Enable speaker diarization for recordings where identifying different speakers matters. Short ordinary voice notes normally do not need it.
Learning from closed conversations
- A ticket closes with a stable resolution and enough quality evidence.
- Chataman waits through the stability period so reopened cases do not become knowledge too early.
- Personal, payment, and contact data are redacted.
- Exact duplicates are detected and linked instead of published twice.
- The case appears in AI Learning Review as pending.
- A manager may edit the problem and verified resolution, then approve or reject it.
- Approved cases become eligible for knowledge promotion.
- Every review creates a version; rollback returns the case to pending review and unpublishes unsafe content.
Operations & Evals
The dashboard reports evidence from real runs:
- Trace count and agent success rate.
- Tool call count and tool success rate.
- P95 response latency.
- Text, image, and audio trace counts.
- Pending approvals and successful post-approval customer responses.
- Learning cases, pending reviews, and promoted cases.
- Long-audio chunking availability.
- Latest agent runs, failures, retry controls, and the immutable audit chain.
Readiness score
Chataman does not issue a readiness score before 20 real traces. After that, the score combines agent success, tool success, successful approval resumption, and latency. Missing critical runtime capabilities prevent a false “ready” verdict.
Weekly operating routine
Review failed runs daily, pending approvals during each shift, learning cases at least weekly, and tool policies whenever channels or team responsibilities change. Use conversation samples together with metrics; a high score does not replace human quality review.
Detailed Operating Playbook
Primary owner: AI quality owner and knowledge reviewer.
Team workflow
- Test multiple images and a voice note.
- Review pending learning cases for redaction and correctness.
- Approve, reject, or roll back knowledge versions.
- Use real traces and Operations metrics for readiness decisions.
Success signals
- Image instructions cannot override system policy.
- Long audio chunking is available.
- Only approved resolutions reach knowledge.
- The audit chain is verified and readiness is evidence-based.
When to review or escalate
Review this workflow when channels, policies, or workload change. Escalate to a manager when numbers look inconsistent or the same issue repeats more than once in the same day.