feliXart
All articles

Why do AI projects stay in pilot?

A successful pilot does not mean production will follow. The distance between them is not model quality; it is enterprise requirements.

Published: 24 August 2026

The short answer

The short answer: a pilot works because it runs on selected examples under supervision. Production asks for measurement, limits, an audit trail and predictable cost. Until those four exist there is no case for promoting the pilot, and the project remains a good demonstration.

01

Which process should AI be applied to?

Pilots usually start on the most impressive process rather than the most suitable one. Suitability has three tests: does the work repeat often enough, is the cost of an error bearable, and can correctness be measured?

Where all three do not hold, no investment case appears however well the model performs. Automating something done three times a month does not repay setup and running cost.

That is why the first output of a readiness assessment is process selection — and sometimes the answer is that you do not need this yet.

02

How is AI accuracy measured? (evals)

In a pilot quality is judged by impression: a few examples are tried and the result looks good. In production that is not enough, because the system meets hundreds of different inputs every day.

Without measurement these questions have no answer: we changed the prompt — did quality fall? The model version updated — is it still doing the same job? Did a scenario that worked last month break this month?

An enterprise AI system depends on a fixed evaluation set built from real questions. Every change is scored against it and a version below the threshold does not ship. Deciding to go to production without that means putting an unmeasured system live.

03

What is an AI guardrail or policy layer?

In a pilot nobody asks the model a hard question, because everyone knows what it is meant to do. In production a real user pushes the boundary precisely because they do not know where it is.

What the model may not say, where it may not steer, how it masks personal data and how it states that it does not know something all have to be written down. Without that layer a corporate risk is produced one day, and that is the day the project closes.

In our IPO application the boundary was set like this: speak only from the data supplied, never invent a figure or a ratio, state a gap explicitly as missing, and never steer an investor to buy or sell.

04

How are AI data protection and residency handled?

A pilot runs on the initiative of IT or a single department. The production decision is made once legal and IT sit down, and four questions are asked there: where does the data go, how long is it kept, is it used to train the model, and is there a record of who reached what?

Without written answers the project does not move. Looking for those answers after the pilot usually means rebuilding the architecture, which is why data residency and the audit trail belong in the first design.

05

How is the cost of an AI system calculated?

Pilot volume is low, so cost stays invisible. In production, budget approval does not come without a cost per transaction.

The answer is to measure the pilot at real volume and derive that figure. Long context and long answers are cost directly; using a different model per task and cutting unnecessary context lowers it several times over in most systems.

Frequently asked

How long should a pilot run?

Long enough to produce a measurable result. The output decides, not the calendar: once an evaluation set exists, quality has been scored and cost measured at real volume, the pilot has done its job.

Who is responsible if the model invents something?

The party that built the system. That is why grounding, citation, output validation and human approval are matters for the contract. No irreversible action is left to the model.

Do we need to train a model on our own data?

For most companies, no. Fine-tuning costs data, time and money, and must be repeated whenever the information changes. Grounding keeps the knowledge outside the model, so the system updates the moment a document does.


Let’s talk about where you are.

Tell us what you want to build — and we will say so plainly if you do not need it.