AI Copilot

The copilot has three modes, from least to most autonomous, all available from the same ticket detail view.

Summarize

The Summarize button generates a short summary of the whole ticket — useful for picking up an old case or handing it to another agent without re-reading the entire thread. It shows up in a highlighted panel above the messages, and never replaces the real history.

Suggest reply

Suggest reply drafts a response based on the ticket's content and, when it finds something relevant, on published knowledge base articles (semantic similarity search, not exact keyword matching). The draft appears directly in the reply box — the agent edits it or sends it as-is, it's never sent on its own. If it used an article, it's listed as "Based on:" below the box.

Triage of new tickets

Settings → AI Settings → AI triage of new tickets has the AI read each new ticket's subject and first message (or an alert's description) and pick a priority and one of your teams:

• Off (default): nothing happens.
• Suggest: the ticket shows "AI suggests priority …, team …" with the reason and an Apply button.
• Apply automatically: the priority and team are set right away, but only if nobody set them already (a priority chosen on a catalog form or mapped from an alert's severity is kept). An internal note on the ticket says what changed and why.

Triage runs in the background after the ticket is created, so it never slows down intake. It works for every channel, email included. Each new ticket costs one small model call, shown under AI cost transparency as triage.

Autonomous mode ("Let AI try")

Here the copilot can investigate the ticket and, if confident, act on it — not just draft text. The key difference from the two modes above is that it can actually execute tools (look up data, change status, assign, apply a macro, add a reply), always subject to the tenant's Autonomy Policy.

Autonomy Policy

From Settings → AI Agent Activity, you configure:

• Auto-execute tool allow-list — only the tools explicitly allowed there execute immediately. Any other call lands in a "pending approval" state for a human to review before anything happens.
• Daily action limit — a hard cap, independent of how many tools are approved for auto-execution.

Every call — auto-executed or pending — lands in the same audit log, with the tool used, its arguments, the result, and where the call came from (the internal copilot, or an external agent via the MCP server — they share the exact same catalog and the exact same rules, there's no separate path with fewer controls for external integrations).

AI cost transparency

Settings → AI Usage shows the real dollar cost of every copilot call, aggregated per tenant. Since the AI provider key can be the tenant's own (bring your own key, from Settings → AI), Seredina never adds a markup — the cost shown is exactly what the provider charges.

No AI provider configured

If neither the tenant nor the deployment has an AI key configured, these three buttons return a clear error instead of failing silently or blocking the rest of the app — the rest of Seredina works exactly the same without AI configured. See Environment Variables to configure a provider at the deployment level.