A demo is not a deployed system. The demo skips auth, a database, access control, monitoring, evaluation, and a deployment pipeline, and building those to production standard is as much engineering as the model itself. Teams that do not budget for that work discover the production timeline is dominated by the infrastructure they did not plan for. This is the guide to the part after the demo works.
| The demo | Production | |
|---|---|---|
| Data | A clean extract, prepared once by hand | The real source, inconsistent and permissioned, read live |
| Correctness | Looked right on the examples shown | An evaluation suite, a human gate on the hard cases, a rollback |
| Integration | Output in a notebook or a dashboard | Inside the screen people already use, changing real behavior |
| Operations | Runs when you run it | Monitoring, cost controls, retries, on-call, versioning |
| Ownership | Whoever built the demo | Someone whose job is that it keeps working |
The most common way an AI project dies is that its output lands somewhere nobody works. A model that writes a great answer into a dashboard no one opens has changed nothing. Shipping means the output arrives in the CRM, the ticket, the queue, the existing screen where the work actually happens, and someone changed how they work because of it. If no one has said which screen changes, the project is not scoped to ship.
Every production AI system needs to know what happens when it is unsure or wrong: the invalid record, the failed lookup, the low-confidence answer. The teams that ship decide those cases on purpose and put a human gate exactly there. The teams that stall let the model improvise past them, which is fine in a demo and expensive against a real deadline. Deciding the failure cases is not overhead. It is most of the engineering.
This is not analyst work and it is not a proof of concept. It is a senior engineer inside your codebase and data, accountable for the system reaching production, handling the hard cases on purpose. That role has a name now, the forward deployed engineer, and it exists precisely because the demo-to-production gap is an engineering problem, not a model problem. When AI agents are the right tool, the same discipline applies: climb only as high as the task demands.
The full method is in Past the Demo: a CTO's guide to AI that actually ships. To find which workflow to ship first and what it takes, start with the AI Readiness Assessment, then a forward deployed engineer builds it to production in 90 days.