AI production assurance
Describe the actual approved state of the service so every future variation can be interpreted, assigned and decided.
A production baseline connects an intended purpose, executed version, population, observation period, conditions of use and accountable people. It includes the components that turn an input into a decision or action: data, rules, prompts, interfaces, human controls, third-party services and recovery mechanisms.
The inventory reconciles what was approved, deployed and observed. Every critical execution must be traceable to the relevant artefacts.
A rate is interpretable only when its denominator, population and window are known. The baseline states included uses, exclusions, critical segments, small-volume rules and what to do when data is missing.
Measures connect business performance to system quality, human factors, security and effects. We retain the protocol, sample, evaluator, uncertainty and limits of generalisation.
| Field | Question to resolve | Expected record |
|---|---|---|
| Purpose | Which service does the measure represent? | Objective and supported decision |
| Population | For which cases is the value valid? | Segments, exclusions and volume |
| Method | How is the value produced? | Calculation, sample or review protocol |
| Limit | What does the value not establish? | Uncertainty and validity conditions |
The baseline establishes when the service remains acceptable, becomes restricted or must enter a safe state. The safe state may be manual operation, a previous version, reduced capability or suspension. Its feasibility must be tested; a theoretical option is not a capability.
The business owner validates the service and limits; technical owners attest the version and capacity to act; risk, security, data or legal functions contribute as appropriate. Approval also names who may alter the baseline.
A change in intended purpose, population, model, data, interface, supplier, rule or oversight may require a new baseline. The previous one is retained to reconstruct the transition, tests and decision.
Primary sources
Last documentary review: 6 September 2026.
Continuous risk management, deployer duties, post-market monitoring and serious-incident reporting.
Monitoring categories and measurement challenges for AI systems in real-world operating conditions.
Govern, map, measure and manage the risks of an AI system.
Architecture, lifecycle, dependencies, pre-production tests and logging.
Before you write to us
No. It covers the socio-technical service: model, data, rules, interfaces, dependencies, population, human controls and safe state.
When intended purpose, population, version, data, interface, supplier or oversight makes the previous comparison invalid.
The functions carrying the service and its effects: business owner, technical owners and risk, security, data or legal functions as appropriate.
Production baseline
Present the version and expected service. We assess whether the baseline already supports comparison, decision and recovery.
Review the baseline