Four pillars hold every AI product we build to the same bar. Each carries a small set of principles and a set of standards — marked mandatory or recommended. They operationalise the Group Gen AI Guidelines, Cyber AI Policy (ISO/IEC 42001) and Risk Reference Architecture for Invent teams — see Foundations. Build to these from day one and the gates take care of themselves.
A mandatory standard must be met or formally waived before the relevant gate is passed. A recommended standard is expected good practice; deviate with judgement, not permission. Mandatory standards scale with the product's risk class — a Risk Class 1 product answers to all of them.
Must be met or waived. No sign-off without it.
Expected practice. Deviate with reasoning recorded.
An AI product earns the right to ship by being safe, fair, transparent, and accountable — not by being accurate alone. This pillar is non-negotiable at every risk class.
Every AI decision has a named human owner. The model advises; a person remains answerable.
The higher the stakes, the more human control. Automation is earned, not assumed.
Users know when they are interacting with AI and can understand, at their level, why it did what it did.
| Standard | What it requires | Level |
|---|---|---|
| Risk & impact assessment | A documented assessment of harms, affected groups, and failure modes completed before build. Sets the product's risk class. | Mandatory |
| Regulatory classification | Product classified against the EU AI Act and any sector regulation. Prohibited and high-risk use cases identified and controlled. | Mandatory |
| Bias & fairness testing | Representative evaluation across relevant groups, with thresholds agreed and results recorded. Re-run on material data or model change. | Mandatory |
| Human oversight design | A defined point of human review or intervention appropriate to the risk class, with escalation paths. | Mandatory |
| Transparency & disclosure | Users are told they are interacting with AI. AI-generated content is labelled where it could be mistaken for human or fact. | Mandatory |
| Explainability | Decisions can be explained at a depth matched to the audience — end user, operator, and auditor. | Recommended |
| Data provenance & consent | Training and prompt data has known origin, lawful basis, and licence. No unlicensed or unconsented personal data. Sensitive or live personal data requires a DPO and CISO assessment. | Mandatory |
| Approved technology only | Only Group-approved Gen AI technologies and models, accessed via the Talent page — never through private accounts. Prohibited use cases (e.g. most creative image generation) are declined. | Mandatory |
| Code of Ethics for AI | Aligned to the Capgemini Code of Ethics for AI. On a client environment, the client's guidelines apply first. | Mandatory |
| Sustainability | Model and infrastructure choices weigh compute cost and carbon against value delivered — a frugal approach, smaller models where they suffice. | Recommended |
Where we draw the line. Any use case the EU AI Act prohibits, or that we judge to create unmanageable harm, is declined regardless of client demand. This is not a commercial negotiation. Route to Responsible AI escalation.
Architecture is where good intentions meet production reality. This pillar keeps AI systems secure, scalable, evaluable, and cheap enough to run.
The simplest architecture that meets the need. Reach for a large model only when a smaller, cheaper, or deterministic option cannot do the job.
No AI system moves toward production without a repeatable evaluation harness and a baseline to beat.
Threat modelling and data protection are designed in, not bolted on. Prompt injection and data exfiltration are treated as first-class risks.
| Standard | What it requires | Level |
|---|---|---|
| Risk Reference Architecture | Design assessed against the GCS Gen AI Risk Reference Architecture (the 16 risks) using the AI solution risk assessment questionnaire, and mapped to an approved pattern (RAG, fine-tune, agentic, classical ML) or a justified departure. | Mandatory |
| Approved models & selection | Only Group-approved models. The chosen model, and the alternatives rejected, are recorded against cost, capability, latency, and data-residency criteria. | Mandatory |
| Evaluation harness | An automated, versioned eval suite covering accuracy, safety, and regression. Ground truth defined; results tracked over time. | Mandatory |
| Security & threat model | Threat model covering the 16 risks — prompt injection, data leakage, excessive permission, supply chain, and more. Layer isolation and least privilege; secrets managed; inputs and outputs guarded. | Mandatory |
| LLMOps & observability | Versioned prompts, models, and datasets. Logging, tracing, cost and quality monitoring in production, with alerting on drift. | Mandatory |
| Data architecture | Pipelines are reproducible and governed. Lineage is traceable; PII is minimised, masked, or tokenised. | Recommended |
| Guardrails & fallbacks | Input/output filtering, grounding checks, and a defined degraded-mode behaviour when the model is unavailable or low-confidence. | Mandatory |
| Cost & performance budget | A unit-economics model (cost per request, latency target) agreed before scale-up and monitored against it. | Recommended |
An AI product is only as good as a user's willingness to trust and use it. This pillar keeps us solving the right problem, in an interface people understand.
We validate the user need and the value case before committing to an AI solution. AI is a means, never the brief.
The interface assumes the model will sometimes be wrong, and makes that safe, visible, and correctable.
We help users trust the system exactly as much as it deserves — no more, no less. Confidence is communicated honestly.
| Standard | What it requires | Level |
|---|---|---|
| Discovery & value case | Evidenced user need, target outcome, and success metrics before build. AI is shown to be the right tool, not the default. | Mandatory |
| Human–AI interaction patterns | Uses approved patterns for suggestion, confidence, correction, and hand-off. Users can always challenge or override. | Mandatory |
| Error & uncertainty design | Model uncertainty, hallucination risk, and failure are handled gracefully in the interface, not hidden. | Mandatory |
| Accessibility | Meets WCAG 2.2 AA. Tested with assistive technology; not degraded by AI-driven or dynamic content. | Mandatory |
| Design system alignment | Built on the approved design system and Capgemini Invent brand. Consistent components, tone, and accessibility baked in. | Recommended |
| Feedback loop | Users can rate, flag, or correct AI output, and that signal reaches the team and, where useful, the model. | Recommended |
| Usability validation | Tested with representative users before launch. Trust, comprehension, and task success measured, not assumed. | Mandatory |
The product does not end at launch — someone runs it long after we leave. This pillar makes what we build testable, maintainable, and safe to hand over.
Assume a different team runs this in a year. Documentation, tests, and ownership make that a non-event.
Quality that depends on someone remembering will fail. Testing, checks, and deployment are automated.
Small, reversible releases behind gates and flags beat big-bang launches every time.
| Standard | What it requires | Level |
|---|---|---|
| Lifecycle gate compliance | The product passes the review gates for its risk class, with evidence recorded at each. | Mandatory |
| Test coverage & CI | Automated tests for code and model behaviour, run in CI. SAST, dependency (SCA), secret and IaC scanning enforced. Built to OWASP Top 10, OWASP ASVS, and NIST SSDF. | Mandatory |
| Red-team before production | A red-team activity and vulnerability assessment before go-live; residual risks logged and mitigated. Coordinated with the CISO. | Mandatory |
| AI-assisted engineering | AI coding assistants only where Group-approved. Treat generated code as untrusted input and review it; the engineer remains accountable for what ships. Per the Vibe Coding Guidance. | Mandatory |
| Code quality baseline | Linting, review, and dependency and licence scanning enforced in the pipeline. No unreviewed code to production. | Mandatory |
| Environments & release | Separated dev, test, and production. Reproducible, automated, reversible deployments; changes are traceable. | Mandatory |
| Documentation & model card | Architecture, data, model card, known limitations, and run-book are current and version-controlled. | Mandatory |
| Monitoring & incident response | Production monitoring for quality, drift, and cost, with alerting and a defined incident and rollback process. | Mandatory |
| Handover & ownership | A named owner and a handover pack that lets another team operate the product without the build team. | Recommended |
Bring your product to office hours and we will help you classify it and scope the reviews.
