Integration & Governance

AI accelerates every role, but only inside guardrails. This page defines the rules that keep AI use safe, compliant, and accountable. These apply to everyone.

The non-negotiables

A human owns every AI output. Nothing merges, ships, or reaches a stakeholder without a named accountable human.
No secrets or client IP in prompts unless the tool is an approved, contractually-covered enterprise instance that does not train on our data.
Verify before trust. AI output is a draft. Code is run and reviewed; facts are checked; APIs are confirmed against real docs.
Disclose where required. If a client contract requires disclosure of AI use, we comply.

Data handling tiers

Data class	Examples	Allowed AI usage
Public	OSS code, public docs	Any approved tool
Internal	Our non-secret code, process docs	Enterprise/team-tier tools only
Confidential	Client code, PII, credentials, business secrets	Only approved, contractually-covered instances with no training; redact where possible
Restricted	Keys, tokens, regulated data	Never paste into any AI tool

Human-in-the-loop checkpoints

AI may draft freely, but these decisions always require a human:

flowchart LR
    AI[AI output] --> C{Affects security,<br/>money, scope, or<br/>external comms?}
    C -->|yes| H[Human reviews & approves]
    C -->|no| L[Lightweight human check]
    H --> OUT[Proceed]
    L --> OUT

Merge to main → human Quality Engineer approval (AI review assists).
Release / deploy to prod → human sign-off (Architect + Stakeholders as PO).
Scope or priority change → Stakeholder (Product Owner) decision.
External communication → human author owns the message.
Architecture decision → Architect owns the ADR.

Quality controls on AI itself

Prompt patterns are shared, not reinvented. Reusable prompts for stories, reviews, and test generation live with each role so quality is consistent.
AI suggestions are measured. We track accept/reject rates and defect-escape rates to know whether a tool actually helps (see Field Notes).
Tool access is provisioned and reviewed. Approved tools, enterprise tiers, and offboarding are managed centrally.

Security & compliance integration

AI sits inside, not around, our existing controls:

SAST, SCA, and secrets scanning run in CI regardless of how code was written.
AI-generated dependencies go through the same license and vulnerability checks.
Generated IaC passes policy-as-code checks (e.g. OPA/Conftest, tfsec) before apply.

When NOT to use AI

When you can’t verify the output and the cost of being wrong is high.
When the input would breach a data boundary above.
When it would obscure accountability (“the AI decided”).
When a simple deterministic tool does the job better and more cheaply.