AI and models

AI and models

Most support tools rent AI back to you: their model, their prompts, their bill. TicketMan AI treats AI as infrastructure you own. A single IChatClient abstraction fronts five providers, and every function picks its own connection, model and prompt.

Bring any model, or no cloud at all

Because each function binds to its own connection and model, you get a per-task cost and quality dial. Cheap, high-volume work such as summarizing and beautifying can target a small or local model, while heavier reasoning such as root-cause triage or duplicate judgment can target a larger one.

What the AI does

• Summarize a ticket and its conversation into a glanceable brief.
• Draft replies, affirmative or declining, grounded in the conversation and delivered into the composer. Never auto-sent.
• Beautify rough agent notes into a clear, professional message.
• Suggest assignment of the right team or agent.
• Detect duplicates with a heuristic prefilter feeding an LLM judge.
• Analyze diagnostics crash reports from the Diagnostics SDK to identify root causes.
• Classify confidentiality of new tickets (optional safeguard run once after auto-assignment).

Editable prompts

Each function is bound to its own Scriban prompt template, editable by admins under Settings → Prompts. Sensible defaults seed automatically and are never overwritten by an update.

Built for trust and operations

• Human-in-the-loop by design. Drafts and suggestions are surfaced to an agent; nothing sends itself.
• Full audit log. Every call records tokens, cost and latency.
• Encrypted, write-only keys. Provider keys are stored encrypted at rest and never returned to the client.
• Resilient. Each connection has its own timeout, retry and circuit-breaker pipeline.

See also Security and data ownership.