Model · M-002 · Responsible AI by design
The Responsible AI Lifecycle
The eight stages by which every AI use case is governed, from idea to retirement, each passed on a named artefact rather than an assertion. How it is proven.
- Structure
- 8 stages, 8 artefacts
- Programme
- Responsible AI by design
- Companion
- M-001 Domain Map
- Reviewed
- 2026-11-05
The Lifecycle is the "how" of Responsible AI by design. A use case moves through eight stages, and no stage is passed on assertion: each is passed on an artefact, the evidence named against it, retained for supervisory inspection.
Three stages are the ones commonly skipped: screening (1), approval (5) and change or retirement (8). Each has a note below on why.
Five activities are not stages because they never stop: board and committee oversight, independent assurance, incident response, horizon scanning, and skills and capability. They run alongside every stage.
The loop matters. A material change re-enters at stage 2, assessment, not at build. Continuously learning and agentic systems have no stable post-deployment state, so they are re-tiered on a defined cadence, not only on change. Paired with the Domain Map: the domains say what must be true; the lifecycle says when it is checked, by whom, and on what evidence.
The eight stages
- Stage 1 Ideate & screen
Is it AI? Is it prohibited? Which domains apply?
Passed on Screening record
Commonly skipped Skipped because everyone assumes they know what AI is. The vendor spreadsheet with a model inside it is the one that is missed; screening is what stops shadow AI.
- Stage 2 Assess & tier
Impact assessment, tiering, lawful basis, obligations mapped.
Passed on Risk tier + DPIA
- Stage 3 Design & build
Controls designed in: oversight, explainability, logging, security.
Passed on Design review
- Stage 4 Validate & assure
Independent challenge, bias testing, robustness, red-team.
Passed on Validation report
- Stage 5 Approve
Named accountable owner accepts residual risk. The moment of record.
Passed on Approval minute
Commonly skipped Skipped because approval gets folded into the design review. It must not be: approval is the moment a named senior manager accepts residual risk in writing, and under SM&CR that record is the difference between a defensible decision and an indefensible one.
- Stage 6 Deploy
Staged rollout, customer communications, rollback tested, staff trained.
Passed on Go-live readiness
- Stage 7 Operate & monitor
Performance, drift, outcomes by segment, complaints, overrides.
Passed on Outcome MI
- Stage 8 Change or retire
Material change re-enters at stage 2. Retirement is a decision, not a lapse.
Passed on Change decision
Commonly skipped Skipped because nothing forces it. Models decay silently; retirement should be a decision someone makes, not something that happens by neglect.
What never stops
A material change at stage 8 re-enters at stage 2. Continuously learning and agentic systems are re-tiered on a cadence. Stage 5 is the SM&CR moment of record.
Generalises to. The eight stages and their artefacts are sector-neutral. The SM&CR moment of record at stage 5 is the UK banking instance; elsewhere it is whichever regime makes a named person accountable for accepting residual risk.
Part of Responsible AI by design. Companion entry: M-001 The Responsible AI Domain Map.
Source: Responsible AI by design, portfolio version 1.0 (27 August 2026): the Lifecycle and its talk track.
How to cite
Axiom Verity (2026). The Responsible AI Lifecycle. Reference Library, M-002. https://axiom-verity.com/library/models/responsible-ai-lifecycle. Reviewed 2026-11-05.
Found an error or a better source? Report a correction.