The professional standard for production AI deployment
Verify a credentialFor organisationsPartner NetworkFor nonprofits & NGOsContact
Pre-Launch Control Review

Production AI deployment checklist

Before a production AI system touches customers, employees, regulated workflows, code, payments, or confidential data, it needs evidence across all eight PSF domains.

D1

Input governance

User, tool, file, and retrieved inputs are treated as untrusted and validated before model use.

D2

Output validation

Model outputs are parsed, checked, bounded, and blocked before downstream systems act on them.

D3

Data protection

Prompts, logs, traces, embeddings, and outputs follow minimisation, retention, consent, and deletion rules.

D4

Observability

The team can reconstruct inference chains, detect degradation, and alert on quality, safety, cost, and latency.

D5

Deployment safety

Model, prompt, tool, and retrieval changes are versioned, tested, canaried, and reversible.

D6

Human oversight

High-stakes or irreversible actions have review, escalation, contestability, and skill-maintenance controls.

D7

Security

The AI threat model includes prompt injection, tool abuse, secret leakage, data exfiltration, and supply-chain risks.

D8

Vendor resilience

Fallback providers, abstraction layers, exit plans, and data portability are tested before dependency failure.

When to block launch

A missing control is not automatically a launch blocker. It becomes a blocker when the system can create material customer, legal, financial, safety, security, or employment consequences before a human can intervene.

Open interactive checklist
Read the PSFCPAP portfolio assessment