PSF v1.0 — April 2026

Pre-deployment checklist

A practical checklist based on PSF criteria for teams preparing a production AI deployment. Use this to identify gaps before go-live, not as a substitute for full PSF assessment.

Progress0 / 44 items (0%)

All user inputs are sanitised before reaching the model

Input length and schema are validated before the API call

Rate limiting is active on all AI-facing endpoints

Logs capture sanitised inputs (no raw PII) with request IDs

Prompt injection detection is applied to user-supplied content

The system prompt is structurally isolated from user content

All model outputs pass through a validation layer before delivery

Output schema is enforced — type, length, and required fields are checked

A content filter is applied to detect harmful, biased, or out-of-scope responses

Confidence scores (where available) are logged and low-confidence outputs are flagged

Structured outputs are parsed and validated before downstream use

Edge-case and refusal outputs are handled gracefully, not silently

PII is identified and handled according to a documented policy

Training data provenance is documented

Retention periods for inference logs are defined and enforced

Right-to-erasure requests can be fulfilled without model retraining

Data residency requirements are understood and met

Third-party data used in context is licensed for the intended use

The model selection is documented with a rationale against the use case

Benchmark results relevant to the production use case have been reviewed

Model updates and provider changes are subject to a regression test gate

Fallback models are identified for provider outage scenarios

Vendor terms permit the intended use (commercial, data processing, etc.)

Human review is implemented for high-stakes or irreversible AI actions

Escalation paths are defined for out-of-scope or uncertain model outputs

Users are informed when they are interacting with an AI system

Autonomy level is defined and documented for each AI decision point

Override mechanisms exist for all AI-driven actions

Human oversight design is reviewed at deployment and after significant model updates

Latency, error rate, and throughput are monitored in production

Model output quality is monitored via sampling or automated evaluation

Anomaly detection alerts are configured for unusual output patterns

Dashboards exist for the operations team to monitor AI health

Logs are retained for at least 30 days for incident investigation

A documented incident response plan covers AI-specific failure modes

The team can disable or degrade the AI system within 15 minutes

Post-incident reviews include root cause analysis and remediation tracking

Customer-facing communication templates exist for AI incidents

Incident response has been rehearsed (tabletop exercise or drills)

A named accountable owner is documented for each production AI system

Bias and fairness evaluation has been conducted for the use case

The system complies with applicable AI regulation (EU AI Act, sector-specific rules)

Accountability documentation is accessible to internal audit and regulators

Ethical review has been conducted for sensitive or high-impact use cases

Using this checklist for CPAP submissions

The CPAP portfolio assessment uses the PSF framework criteria, not this checklist directly. However, completing this checklist against a real deployment produces evidence that maps well to CPAP assessment dimensions. See the CPAP rubric for full assessment criteria.