What to demand of an AI pilot before it goes to production

The distance between a demo people like and a system you can stand behind is not a question of model. It is the same distance that separates any prototype from any product: measurement, cost, limits and operations.

1. A set of real cases, not a demo

Before scaling anything you need a set of fifty to two hundred real cases with the correct answer annotated by somebody who knows the business. It is tedious work and it is what turns "seems to work well" into a number.

Without that set you cannot know whether a change of prompt, model or provider improves or worsens things. You are optimising blind, with an invoice attached.

2. Cost per operation, measured

The unit cost in the pilot multiplied by real volume produces a figure that sometimes changes the entire decision. Worth calculating before committing, not after the first monthly invoice.

The surprise usually goes in the good direction: for many classification or extraction cases, a small well-tuned model does the same job as a large one for a fraction of the price. But that is only discovered by measuring.

3. A clear limit on what the system must not do

Every AI system in production needs a defined answer for "I am not sure". Without it, the system always answers something, and answering confidently when it does not know is precisely the most expensive failure mode.

• A threshold above which the operation escalates to a person.
• Categories of query that are never answered automatically.
• A log of what was decided and on what information, so it can be reconstructed later.

4. Somebody who looks at the failures every month

The quality of an AI system is not a fixed property: it degrades when input data changes, when the provider updates the model or when the business changes a process. If nobody is assigned the monthly review of failed cases, the degradation gets discovered through a customer complaint.

What is not on the list

Which model to use is not on it, nor which provider, nor whether you need a vector database. Those are reversible and relatively cheap to change. The four above determine whether the system holds.