Use review gates for consequential AI commerce operations
Generated prices, product claims, creatives, and campaigns can create commercial or compliance harm when publication is treated as a model side effect.
Context
The platform connects several AI-assisted workflows to deterministic commerce records and external providers.
Experiments and investigation
- ✓Model each output as a persisted draft with provenance.
- ✓Separate generation, validation, review, approval, and publication states.
- ✓Keep provider selection behind a task router rather than feature code.
Decisions
- ✓Require approval for consequential output by default.
- ✓Store model, task, references, validation, and reviewer state.
- ✓Use idempotent jobs when approved work crosses an external API boundary.
Result and current status
The operational foundation and review-state architecture are implemented; live provider acceptance remains environment-dependent.
Failure or limitation
Production quality and provider behaviour still require credentialed sandbox testing.
Lessons learned
- ✓Human review is an operating boundary, not a decorative confirmation dialog.
- ✓Deterministic commerce truth must remain separate from generated suggestions.
Next step
Complete sandbox acceptance with representative inputs, failures, retries, and reviewer feedback.