- Home
- Services
- AI & Agent Engineering
- The AI prototype works in the demo and doesn’t reach production
AI in production
The AI prototype works in the demo and doesn’t reach production
An AI prototype that stays at the demo is a symptom, and its causes appear along the way to production: a use case chosen because it was easy to show, an agent that reads files prepared by hand, answers never measured, no plan for when it fails. Skipping a stage means arriving unprepared at the next, and meanwhile the model, the part that works, is the only thing anyone works on.
Book a callDoes this sound familiar?
- Everyone liked the demo, and since then the prototype has run only on the laptop of the person who wrote it.
- To get the prototype running, someone prepares a data file by hand every time.
- Did the last change to the prompt improve the answers or make them worse? It gets argued over on gut feeling.
- Before opening it to customers, someone asks what happens when it gets things wrong, and the question goes unanswered.
- Privacy and compliance are holding up the decision, because it isn’t clear what data would be sent to the model provider.
Why doesn’t the AI prototype reach production?
A prototype shows that the model can do the task on the cases chosen for the demo. Production needs it to do the task every day on real data, inside systems not set up to receive it. The causes below are the stages of that journey, in the order they appear.
The use case was chosen for the demo
The first prototype often starts in the department that wanted it, on the task easiest to show, without asking whether the systems and the people who will have to use it are ready. When it is time to take it to production, you find that the value lies elsewhere or that the data to support it is missing.
The agent isn’t connected to the real systems
Connecting the prototype to the database, the APIs and the existing workflows means working on the software they run on. As long as that work is postponed, the agent stays outside the flow it is supposed to work in.
There is no test set for the answers
The prototype was judged by eye, on the answers people liked in the meeting, and there is no collection of real requests with the expected answer to measure it against. Without those cases every change to the prompt or the model is trial and error, and signing off a release becomes a gamble.
Nobody has decided what the agent does when it gets things wrong
An agent will get things wrong, and production needs to know beforehand which cases go to a person, which actions wait for confirmation and which data can reach the model provider. If these choices are missing, the first question from whoever handles compliance is enough to stop everything.
What changes when the AI agent reaches production
The agent becomes a part of the software like any other. Prompt, configuration and model version live in the repository, every change passes through the test cases and reaches production through the pipeline. When it gets something wrong, the error stays in the logs with the version that produced it and becomes one more test case.
Opening the agent to a new department, or to customers, rests on the test cases and the logs, and the people who answer for data and compliance sign off on that evidence. The next idea starts from the same pipeline and the same checks.
Show us the prototype and where it stopped
We follow the prototype’s path to production and see whether the first step is connecting it to the systems or an AI Adoption Assessment to choose the right use case.
Message sent
Thanks for reaching out. Our team will get back to you within 24 hours with a tailored proposal