AI Copilot
AI Copilot adds provider-backed support and CX assistance without changing the ticket workflow. Suggestions are drafts only: the app never sends a customer reply automatically and never closes a case automatically.
Provider Setup
| Setting | Meaning |
|---|---|
| Enable AI Copilot | Shows AI actions on ticket intelligence screens when the provider is ready. |
| Provider | The adapter to use. The current adapter supports OpenAI-compatible chat completion APIs. |
| Base URL | The provider API root, for example https://api.openai.com/v1. |
| Model | The model name sent to the provider. |
| API Key | The credential used for provider requests. |
| Timeout Seconds | HTTP timeout for provider responses. |
Ticket Actions
- Ticket Summary: concise current state, impact, and next action.
- Suggested Reply Draft: editable customer-safe reply text.
- Category, Priority & Tags: classification recommendations with reasons.
- Duplicate / Similar Ticket Hints: signals the agent should verify.
- Sentiment & Risk: requester tone, customer risk, and escalation posture.
- SLA Escalation Note: internal escalation draft for risky cases.
AI Capability Matrix
The AI Capability Matrix documents which modules can use assistive AI and which guardrails apply. Current action families include reply, rewrite, summarize, search, translate, classify, route, ticket generation, knowledge writer, incident summary, release notes, report summary, executive dashboard, customer summary, sentiment, priority, SLA prediction, escalation, and copilot.
Platform Guardrails
- No auto-send: generated customer messages stay editable until a human sends them.
- No auto-close: AI cannot close tickets or incidents by itself.
- Human approval: higher-risk actions are logged as pending or disabled until reviewed.
- Source-bound knowledge: customer-facing knowledge answers require approved sources.
- Audit trail: platform AI runs record module, action, prompt payload, source references, confidence, guardrail flags, user, and status.
Workflow Actions
| Action | Purpose | Safety behavior |
|---|---|---|
triage | Predict category, priority, tags, sentiment, duplicate risk, SLA risk, and suggested assignee. | Suggestions are auditable and do not overwrite critical fields unless an admin explicitly enables low-risk auto-apply. |
reply_compose / reply_rewrite | Draft, improve tone, shorten, expand, translate, or make a reply more professional. | Generated text stays editable in the reply composer before sending. |
reply_qa | Warn about unsupported claims, private details, weak tone, missing timeline, or mismatch with ticket facts. | Warnings do not block by default. Strict blocking is controlled by admin setting. |
resolution_summary | Draft internal resolution summary, customer closure note, root cause, and prevention steps. | Only accepted summaries are saved to closure fields. |
knowledge_draft | Create draft FAQ or knowledge article from resolved tickets. | AI drafts require human review before publishing. |
AI Trust Report
/reports/ai-trust summarizes AI suggestion runs by action, status, applied rate, accepted/rejected review status, provider failures, model, and provider. It also includes source-bound knowledge-search signals from knowledge_search_events. This report is audit-only; it does not perform new AI generation.
Source-Bound Knowledge Search
AI knowledge search is restricted to approved knowledge-base articles and resolved-ticket summaries. Every answer must include citations, and customer-facing answers are blocked when no approved source supports the response.
Audit and Safety
- Every request writes an
ai_suggestion_runsrecord with ticket, user, action, provider, model, status, prompt payload, response payload, suggestion text, metadata, review status, confidence, and error message when available. - Completed suggestions can be marked as applied, but applying only updates the audit record. It does not post a ticket reply.
- If AI is disabled or credentials are missing, the run is skipped and the ticket shows a clear setup message.
- Provider failures are recorded as failed runs so admins can troubleshoot credentials, endpoint, timeout, or model names.