← AI at Invent / How we govern AI
Capgemini Invent AI Design Authority
Lifecycle review gates

Five checkpoints from idea to operation.

Each gate is a short, structured review with clear entry and exit criteria. Pass the gate for your risk class and you carry evidence into the next phase. The gates are proportionate: a Risk Class 3 product moves through a lighter version of the same path. They align to the Group Cyber Approval Process and the ASLM asset lifecycle, so a single set of evidence serves both.

The path

Five gates make up one lifecycle.

Gates map to the natural decision points of an AI build, and to the five steps of the Group Cyber Approval Process. You do not stop work to pass them; you bring the work you have already done.

Every gate ends in one of three decisions, the Group Cyber Approval vocabulary. Approved means proceed. Approved with conditions means proceed while named actions close. Not approved means a material standard is unmet; resolve it or raise an exception. Conditional approvals record the condition, an accountable owner, a due date, and a re-review trigger.

Gate 0

Discovery & Scoping

At opportunity or discovery, Cyber Step I
Entry: bring this
  • A documented business purpose, intended use, and ROI justification
  • Model and technology selection from the Group-approved list, with rationale
  • Data identification and classification against the Capgemini scheme
Exit: you leave with
  • A confirmed risk class (1–3) and EU AI Act classification
  • An agreed review plan and named reviewer for the build
  • Early flags on regulatory, IP, or Responsible AI exposure
Gate 1

Design & Definition

Before significant build, Cyber Step II
Entry: bring this
  • Solution architecture mapped to the Gen AI Risk Reference Architecture
  • Completed AI solution risk assessment questionnaire (CIA needs)
  • Resource documentation: data, tooling, and system resources
  • Draft roles & responsibilities, incident response, and continuity plan
Exit: decision on
  • Architecture is fit, secure, and evaluable against the 16 risks
  • Responsible AI risks are understood and owned
  • The UX approach is sound for the risk of being wrong
  • Evaluation approach and success thresholds agreed
Gate 2

Build & Data Readiness

Mid-build, before scale-up; Cyber Step III
Entry: bring this
  • Working evaluation harness with a baseline result
  • Bias and fairness testing against agreed thresholds
  • Data provenance and masking; live/sensitive data cleared by DPO & CISO
  • Grounding (RAG), guardrails, and degraded-mode behaviour in place
Exit: decision on
  • Model meets the quality and safety bar it was set
  • Failure modes are known, bounded, and handled
  • Cost and latency are within an agreed budget
  • Clear to invest in productionising
Gate 3

Security & Verification

Before go-live; Cyber Steps IV–V
Entry: bring this
  • Evidence against all mandatory standards for the risk class
  • Red-team activity and vulnerability assessment completed
  • SAST, SCA, secret and IaC scanning clean in CI
  • Usability validation with representative users
  • Monitoring, alerting, incident response, and continuity plan
  • Documentation, model card, and named owner
Exit: decision on
  • The product is safe, sound, usable, and supportable
  • CISO sign-off; independent Responsible AI sign-off (Risk Class 1)
  • Decision: Approved, Approved with conditions, or Not approved
  • Residual risks logged; open items carry an exception with owner and date
Gate 4

Deployment & Service

30–90 days after launch, then periodic; Security in service
Entry: bring this
  • Production quality, drift, cost, and incident metrics
  • Threat and compliance monitoring; alerting tested with the SOC
  • User feedback and any fairness signals from real use
  • Status of conditions or exceptions carried from Gate 3
Exit: decision on
  • The product behaves as designed under real load
  • Open conditions and exceptions are closed or renewed
  • Re-classify or re-review triggered by material change
  • Lessons captured back into the standards
Proportionate

How the gates scale by risk class.

The path is the same for every risk class; only the weight changes. The risk class is set at Gate 0 and drives how much review each gate carries.

Risk ClassGate 0–2Gate 3Gate 4
Risk Class 3: Foundational
Internal or low stakes		Self-certify against a checklist; async reviewLight review with one reviewerMetrics submitted; review by exception
Risk Class 2: Elevated
Client-facing, contained		Live gate reviews with the AuthorityFull sign-off; Responsible AI assessmentScheduled operate review at 30–90 days
Risk Class 1: Critical
High-stakes or regulated		Named reviewer across all gatesIndependent Responsible AI & security sign-offPeriodic review; re-classify on any material change
One gate serves three frameworks

How our gates map to Group processes.

The Authority's gates are a single front end to the Group Cyber Approval Process and the ASLM asset lifecycle. Pass a gate once; satisfy all three.

AI Design Authority gateCyber Approval ProcessASLM lifecycle
Gate 0: Discovery & ScopingStep I: Discovery & ScopingInitialize, then Ready to Plan
Gate 1: Design & DefinitionStep II: DesignPlan & Design
Gate 2: Build & Data ReadinessStep III: Development or IntegrationIncubation, then Ready to Industrialize
Gate 3: Security & VerificationSteps IV–V: Test/QA & Deployment (CISO decision)Industrialization, then Ready to Run
Gate 4: Deployment & ServiceStep V: Security in serviceRun (then Ready to Retire)

Investments over €1M also dock with the Group Investment Committee (GIC) per ASLM guidelines.

Ready for a gate, or not sure you are?

Book the review, or raise an advisory request first if the work is still in progress. Both start in the same place.

Book a gate review Check the standards