Appearance
Guide: Human-in-the-Loop & Governance
Decision: Where humans must decide, and how governance scales with AI velocity.
Related: GOVERNANCE.md · processes/overview.md
Why human gates still matter with AI
AI increases drafting speed — not accountability. Irreversible or high-blast-radius decisions need a named human approver:
- Architecture commitments (ADRs)
- External API contracts
- Production promotion (especially T1)
- Security and data policy
- Incident command
Gate models
| Model | Pros | Cons |
|---|---|---|
| Heavy gates (enterprise regulated) | Audit trail; low drift | Slower; frustration if over-applied to T3 |
| Light gates (startup) | Speed | Architecture and security debt |
| Tier-based gates | Risk-proportional | Requires honest tier classification |
Recommended: Tier-based (GOVERNANCE.md) — T1 strict, T3 lighter.
RACI alternatives
| Approach | Pros | Cons |
|---|---|---|
| Full RACI per phase | Clear | Documentation overhead |
| DACI (Driver, Approver, Contributors, Informed) | Simpler decisions | Less familiar |
| Code owners only | Git-native | Misses PO/SRE/accountability |
Use RACI or DACI in your handbook; this repo provides a reference RACI in GOVERNANCE.
Human vs AI by activity
| Activity | AI role | Human role |
|---|---|---|
| Draft ADR | Full draft | Accept/reject/edit |
| Draft spec | OpenAPI skeleton | Approve contract |
| Write code | Generate + refactor | Steer; review PR |
| Write tests | Generate from spec | Review quality |
| PR review | First pass comments | Merge decision |
| Staging sign-off | Synthetics data | PO accepts behavior |
| Prod deploy T1 | Metric analysis | SRE/ARCH approve |
| Incident RCA | Log summary draft | SRE validates |
| Policy exception | Risk summary | SEC/EM approve |
Pitfalls
| Pitfall | Why it hurts | Mitigation |
|---|---|---|
| Rubber-stamping AI drafts | Governance theater | Time-boxed review; ARB for cross-team |
| Too many approvers | Bottleneck defeats AI speed | Tier-based approvers |
| Zero human review ("AI said OK") | Incidents | Merge requires human + CI |
| Approvers not trained on AI limits | False confidence | Onboarding module (SOP-011) |
| No exception process | Shadow IT bypass | SOP-012 with expiry |
| PO skipped on staging | Wrong product ships green | Explicit staging sign-off |
| Architect unavailable | De facto code-first | Delegate + ADR backlog SLA |
Recommended starting point
Six phase gates (G0–G5) with tier-scoped approvers; document in team handbook; record your choices in your ADRs.
Reference procedures: SOP index.