Rafael Mendoza on Why His Company Stopped Buying AI and Started Building Constraints

Rafael Mendoza can tell you the exact number of AI pilots Meridian Data ran between 2023 and 2025. It is nineteen. He can also tell you how many reached production. It is three.

"That is not a failure rate," he says. "That is a governance rate. We were not choosing badly. We were not choosing at all."

The pilot trap

Meridian, a mid-sized analytics provider serving insurers and logistics groups, had no shortage of enthusiasm. Every business unit wanted a model. Every vendor had a demo. And Mendoza, as chief technology officer, had a budget line that made saying yes easier than saying wait.

The problem surfaced in his own reporting. Nineteen pilots produced nineteen decks, each of which showed improvement against a baseline that the pilot team had defined.

If you let the team that builds the pilot also choose what the pilot is measured against, you have not run an experiment. You have run a demo with extra steps.

Three questions before a second cheque

Mendoza's response was a document his team now calls, without much affection, the gate. It is one page. It asks three questions of any pilot seeking further funding.

The first is who owns the workflow the model touches once the pilot team leaves. Not who sponsors it. Who is accountable on a Tuesday when it produces a bad output.

The second is what the baseline was before anyone knew a model was coming. Mendoza insists this be recorded before the pilot starts, by the operating team rather than the engineering team.

The third is what the failure mode costs. "Everyone can tell you the upside," he says. "Very few can tell you what a confidently wrong answer costs them, and that number decides whether you should deploy at all."

Saying no is the product

Of the six pilots Meridian has run since the gate went in, four reached production. Two were stopped. Mendoza treats the two stops as the more important result.

"A stopped pilot used to feel like a write-off," he says. "Now it feels like the system working. We spent a quarter and learned that the workflow was not ready. That is cheap. Deploying into a workflow that is not ready is not cheap."

He is careful not to present this as a technology position. He is not sceptical about the models. He is sceptical about organisations that treat model capability as the binding constraint when the binding constraint is almost always the process the model is dropped into.

What he tells other technology chiefs

Mendoza now spends a portion of each month talking to peers at similar companies, and he says the conversation is nearly always the same shape.

"They open with the model choice," he says. "Which model, which vendor, which hosting. And I ask them who owns the output, and there is a pause. That pause is the whole problem. You can change the model in an afternoon. You cannot change the fact that nobody in the business has agreed to be responsible for what it says."

His advice is deliberately unfashionable. Run fewer pilots. Write the baseline down first. Give every model a name on the door.

"None of that is interesting," he says. "All of it is why the three that shipped, shipped."