AI PROOF PACKET TEMPLATE (ENTERPRISE-READY)

Purpose
This packet is what you hand to a security reviewer, procurement, or a regulated customer to answer: what the system does, what data it touches, how it’s controlled, and how you respond when it fails.

1) System Overview (1 page)
- Product description: what tasks the AI system performs in production.
- Deployment model: SaaS / single-tenant / on-prem / customer cloud.
- Primary users: roles and permission levels.
- Critical actions: list actions that can change data, send messages, approve transactions, or impact users.

2) Architecture + Data Flow (diagram + bullets)
- Inputs: user prompts, files, APIs, event streams.
- Processing: orchestration layer, retrieval layer, tool calls.
- Model providers: list (e.g., OpenAI, Anthropic, Google) and where requests are sent.
- Storage: logs, embeddings/vector DB, uploaded files, caches.
- Egress: where outputs go (apps, emails, tickets, CRM writeback).

3) Data Handling Policy (clear and specific)
- Data classification: what types you accept (PII, PHI, financial data) and what you prohibit.
- Retention: how long you store prompts, outputs, and traces; deletion workflow.
- Tenant isolation: how you prevent cross-customer data exposure.
- Access controls: who inside your company can access production data and under what approval.

4) Model & Prompt Change Control
- Versioning: how you version model selection, system prompts, tools, and retrieval settings.
- Release process: what must pass before deployment (tests/evals, approvals).
- Rollback plan: how you revert quickly and what triggers a rollback.

5) Evaluation Summary (what you test, not vague claims)
- Task list: 10–30 representative tasks tied to customer workflows.
- Acceptance criteria: what “good” means per task (accuracy, citations, refusal behavior, action correctness).
- Known failure modes: top risks (prompt injection, data leakage, unsafe actions, wrong citations).
- Regression checks: how you detect quality drops after changes.

6) Logging & Auditability
- What you log per request: request ID, user ID, model ID, prompt hashes, tools invoked, retrieval sources, policy checks.
- How logs are protected: encryption, access control, tamper resistance.
- Export process: how you produce an audit extract for a customer.

7) Incident Response Runbook (operational, not aspirational)
- Severity levels: define what counts as a critical AI incident.
- Triage steps: how you confirm and reproduce the issue.
- Containment: disabling tools, blocking actions, model switch, feature flag off.
- Customer communication: who notifies, what timeline, what artifacts you provide.
- Post-incident: root cause, corrective actions, eval additions.

8) Customer Controls (what buyers can configure)
- SSO/SAML availability.
- Role-based access control.
- Approval workflows for high-impact actions.
- Data source connectors and permission enforcement.
- Policy settings (content filters, action scope limits, retention choices).

How to use this template
- Fill it once, then treat it as a living document tied to releases.
- Make every claim testable: if you say “we log X,” show an example redacted trace.
- Store the packet in the same repo/process as code changes so it stays current.