Delivery · August 2026 · 4 min read
From prototype to production
The distance between a working demonstration and a dependable system is where most AI initiatives quietly die. The gap is not mysterious. It is a list of unglamorous engineering debts.
Every organisation now has one: the prototype that worked. It answered the questions in the demo, the room was impressed, and a date was pencilled in for rollout. Months later the pencil marks have faded. Nothing failed, exactly. The project just entered the long corridor between demonstration and dependability, and stopped.
The corridor is not mysterious. It is a list, and most of the list is engineering debts the demo was allowed to skip.
What the demo was allowed to skip
The demo ran on curated data; production runs on the data you actually have, with its gaps, duplicates and quiet disagreements between systems. The demo had an operator who wanted it to work; production has users with jobs to do and no patience for a tool that wastes their time twice. The demo failed privately; production fails in front of the operation, so failure behaviour, fallbacks and human override stop being edge cases and become the design. The demo was free to be wrong; production answers feed decisions, so evaluation must be continuous, measured against what the operation needs, not what a benchmark rewards.
None of these debts is exotic. Each is unglamorous, priceable work. The failure mode is not that organisations cannot do the work; it is that nobody listed it, so nobody budgeted it, so the prototype was judged as though the corridor did not exist.
Crossing deliberately
The crossing goes better when it is planned as its own project with its own definition of done: the system runs on real data, inside the real workflow, watched by real monitoring, with its quality measured and its limits documented. It goes better still when the people who will live with the system help shape it during the crossing, because adoption debt compounds faster than technical debt and is harder to refinance.
A prototype is a question asked cheaply: could this work? Production is a promise kept expensively: it works, still, today. The organisations that get value from AI are not the ones with the best demos. They are the ones that price the promise before falling in love with the question.
Written by Anthony Smith, Chief Technology Officer, Ballista.
Talking beats reading.
If you have a prototype that works and a production date that worries you, tell us what it must survive. We will help you price the gap honestly.
Your message goes to the people who would do the work.