How to Industrialize from Pilot to Production
Most operational AI dies between the demo and the daily operation. This guide covers the engineering and governance of that passage: criteria written before the pilot, pilots run on live work, hardening before scale, staged gates, and the discipline to retire the old path. It is the least glamorous skill in the set and the one that decides whether anything real ships.
Developing
Start here. Build the foundation.- 1
Before launch, write what the pilot must prove: accuracy, exception rate, cost per transaction, volume, and what result kills it. Get the criteria signed by the stream owner. You are done when the later go or stop decision cites the document.
- 2
Put the pilot on real transactions with real data quality and real users, at enough volume to surface true exception rates. A pilot that cannot be run on live work yet is telling you something: the integration or data work is not done. The test: the pilot report includes the ugly cases and what happened to them.
Proficient
Build consistency and rhythm.- 3
Treat hardening as its own phase with its own owner: pipelines, integrations, exception routing, monitoring, fallbacks per failure mode. Test the exception paths at realistic volume. It works when the first scaled site's incident list looks like the pilot's, not worse.
- 4
Define the stages and re-check the production criteria at each gate, with rollback as a live option. Retire the legacy path as each stage passes. The signal of discipline: at least one gate has ever said no.
Mastered
Operate at the highest level.- 5
Turn the gates, readiness checklists, and cutover patterns into the standard path every program follows. When the next program runs it without you and moves faster, the playbook is real. Mastery is measured in program velocity: the tenth industrialization should be routine.
Common Pitfalls
Avoid the common failure modes.- Writing success criteria after seeing the results. Criteria that postdate the pilot are advocacy, not governance, and everyone in the room knows it.
- Scaling on momentum past a gate that should have said no. The problems a gate would have caught at one site arrive later at every site at once.
- Running old and new in parallel indefinitely. It doubles cost and lets the organization quietly revert; the old path has to be retired stage by stage.