Add AI Ticket Summaries and Reply Drafts to Our Customer Support SaaS

The feature we are building

Our customer support SaaS needs an AI-assisted ticket view. When a support agent opens a long conversation, they should be able to generate a concise summary and request a suggested reply.

The reply must remain a draft. Sending it continues to require an explicit action by the support agent.

Our application uses React, FastAPI, and PostgreSQL. Existing authentication, organisation membership, and ticket permissions must be reused.

User flow

1. The user opens a ticket they are authorised to access.

2. They select “Summarise conversation.”

3. The application returns the issue, actions already taken, open questions, and suggested next steps.

4. The user can generate and edit a reply draft.

5. The user accepts or discards the draft through the existing interface.

Implementation requirements

Integrate one LLM provider selected during discovery. Keep credentials server-side, validate structured summary responses, and provide clear loading, timeout, and retry states.

Requests must use only the authorised ticket context. Add configurable input and output limits, per-organisation usage limits, and operational logging without storing conversation content in logs by default.

Acceptance checks

- Cross-organisation ticket access is blocked before any model request.

- Summaries follow the agreed response schema.

- Drafts are never sent automatically.

- Empty, oversized, and unavailable-provider cases produce usable error states.

- Quality is reviewed against 40 anonymised conversations using an agreed rubric.

We expect source code, automated integration tests, staging deployment, one production release, and a setup guide. Include 14 days of bug support.

This scope excludes knowledge-base retrieval, autonomous actions, and a redesign of the ticket interface. Provider usage fees are paid separately.

Please include a relevant shipped feature, your proposed implementation plan, and a milestone breakdown.