5.4.5. From Experimental to Production-Grade
💡 First Principle: A pilot is allowed to be fragile because it serves a few forgiving users under close watch. Production serves everyone, every day, under real conditions, so it must be reliable, secure, governed, supported, monitored, and affordable. The transition from experiment to production is where most of the remaining work lives, and skipping it is how successful pilots become production incidents.
What a pilot typically lacks, and production requires:
| Area | Pilot reality | Production-grade requirement |
|---|---|---|
| Governance | Informal approval | Risk classification, governance sign-off proportional to tier, documented decisions, entry in the AI inventory |
| Security and privacy | Sandbox data, broad access | Security review, least-privilege access, data protection, privacy assessment |
| Reliability | Best effort | Defined service levels (availability, response time), capacity planning, fallback when AI is unavailable |
| Monitoring | Manual checks | Performance, drift, cost, and responsible AI monitoring with thresholds and owners |
| Support model | Builders answer questions | Defined support tiers, incident response, on-call ownership, runbooks |
| Integration | Manual inputs and outputs | Integration with core systems and workflows |
| Change management | Volunteer early adopters | Training, communication, process documentation for all users |
| Cost control | Small, unmanaged | Budgets, attribution, optimization (3.2.5) |
| Lifecycle management | Fixed model version | Change control for model, prompt, and data updates; re-validation; retirement plan |
The exam guide emphasizes governance and operational requirements in this transition. The governance requirements come from Phase 4 (oversight, risk tier, accountable owner, responsible AI testing). The operational requirements are about running the system reliably: monitoring, support, service levels, incident response, and lifecycle management.
A useful practice is a production readiness review: a checklist-based gate, scaled by risk tier, that an initiative must pass before launch. It turns "is it ready?" from opinion into evidence.
⚠️ Exam Trap: When a pilot is moving to production, the best answer adds governance requirements (risk classification, sign-off, oversight, documentation) and operational requirements (monitoring, support model, service levels, incident response, cost controls, change management). Options that simply expand the pilot to more users, or that treat production as a technical deployment step only, are distractors.
Reflection Question: A generative AI pilot used by 25 enthusiastic analysts is about to roll out to 3,000 employees. List the five production-grade requirements you would insist on before launch, and explain which ones the pilot probably did not need.
