An AI pilot proves that a bounded idea deserves further investment. It does not prove that the system is ready for production or scale. Production readiness requires evidence across quality, controls, integration, exceptions, ownership, economics, and adoption.
Gate 1: Task quality
The system should meet predefined criteria on representative cases, difficult cases, and cases it should refuse or escalate. Track important failure categories, not only an average score.
Ready means the team understands the capability boundary and accepts the remaining error profile for the intended use. Not ready means the result depends on handpicked examples or subjective impressions.
Gate 2: Human oversight
Define the reviewer, review trigger, criteria, authority, service level, and retained evidence. Test whether reviewers can actually detect important failures with the information and time available.
Ready means oversight is operationally credible. Not ready means “a human checks it” is the complete control description.
Gate 3: Data and permissions
Confirm approved data sources, purpose, access, quality, retention, lineage, and behavior when data is missing or stale. Test user permissions and connected-system permissions.
Ready means the production data path is authorized and predictable. Not ready means the pilot used manually curated data or elevated access that users will not have.
Gate 4: Integration and exceptions
Test identity, source and destination systems, duplicate actions, timeouts, unavailable services, malformed inputs, retries, recovery, and rollback. Map how a person handles exceptions.
Ready means the complete workflow can fail safely and recover. Not ready means only the model interaction has been demonstrated.
Gate 5: Security, privacy, and governance
Complete the required risk classification, vendor review, threat assessment, privacy review, prohibited-use definition, monitoring plan, incident process, documentation, and approval.
Ready means the evidence matches the risk tier. Not ready means control functions are being asked for a last-minute exception.
Gate 6: Ownership and support
Name the business process owner, system owner, evaluation owner, support path, incident authority, and change approver. Define maintenance and retirement responsibilities.
Ready means questions and failures have somewhere to go after the project team leaves. Not ready means the vendor or temporary project is described as the owner.
Gate 7: Workforce adoption
Train users and managers on the actual workflow, approved environment, data rules, review criteria, difficult cases, and escalation. Test whether they can perform the task without the implementation team guiding every step.
Ready means the new process fits real work and managers can reinforce it. Not ready means the adoption plan is a launch announcement and feature demonstration.
Gate 8: Operating economics
Calculate benefits and the complete run cost: licenses or model usage, integration, review, support, monitoring, rework, training, and incident handling. Use ranges for uncertain assumptions.
Ready means the economics are plausible at the expected adoption and quality levels. Not ready means the case counts theoretical hours saved but ignores human review and operating cost.
Gate 9: Production monitoring
Define performance, workflow, control, incident, cost, and adoption measures. Establish thresholds for investigation, pause, rollback, and re-evaluation.
Ready means the organization can see when the system or context changes. Not ready means monitoring is limited to uptime and model errors.
Gate 10: Controlled launch and scale decision
Start with a defined user population, workflow scope, review cadence, and rollback path. Compare production results with the baseline and pilot assumptions.
The next decision can be to scale, improve, narrow, replace, or stop. Scale should be earned by sustained evidence, not by the desire to declare the pilot successful.
Build a production evidence packet
Keep the essential evidence together: intended use, prohibited use, owner map, system and data flow, risk tier, evaluation, human oversight, approvals, training, support, monitoring, economics, known limitations, rollback, and decision record.
This packet reduces ambiguity when the model, platform, workflow, regulation, or team changes. It also makes later reviews more efficient.