AI Doesn't Hallucinate. Your Data Does.

Somewhere this quarter, an executive watched an AI system deliver a wrong answer with total confidence. Maybe it reported a customer count that didn't match finance's number. Maybe it forecast demand for a product that was renamed a year ago and now exists twice in the system. The room did what rooms do: someone said the model was hallucinating, everyone nodded, and the AI initiative absorbed another quiet blow to its credibility.

Hallucination is a real phenomenon in generative AI. It is also the most convenient misdiagnosis in enterprise technology, because it assigns the failure to the mysterious machine instead of to anything a leadership team could have prevented.

Here is what we find when we audit these systems: the model was usually telling the truth. The truth just happened to be about the data.

The model is a mirror

An AI system connected to your business does not consult reality. It consults your systems of record, and it treats them as reality. If sales and finance define a customer differently, the model doesn't detect the disagreement and split the difference. It picks up one definition, or worse, both, and produces answers that are internally consistent with data that isn't. If a product was renamed and now lives in your warehouse under two identities, the model sees two products. When it reports on either one, it is being perfectly faithful to what you gave it.

A model that automates your data is also automating your data's disagreements, at scale, with confidence, in front of your board.

This is the mechanism behind a number that should alarm any executive approving AI budget: MIT research found that roughly 95 percent of enterprise generative AI pilots produce no measurable return. Harvard Business Review has reported that only 3 percent of companies' data meets basic quality standards. Those two findings are not separate facts. The second is the explanation for the first.

Why nobody catches it until production

In a demo, this failure mode is invisible. Demos run on curated data, narrow questions, and forgiving audiences. The conflicting definitions, the duplicate identities, the metrics that quietly changed meaning last quarter: none of it surfaces until the system is answering real questions from people who know what the numbers should say.

That is why the discovery so often arrives six months and several hundred thousand dollars into an implementation. Not because anyone was careless, but because nothing in a typical AI project plan ever asks whether the foundation can support what's being built on it. The technology evaluation asks whether the model works. It does. The question nobody assigned was whether the data means what everyone assumes it means.

What readiness actually looks like

In our audit work at Standard North, we score organizations across nine dimensions of data maturity, producing the Standard North AI Readiness Score, or SNARS, on a scale from 0 to 4.0. The pattern that emerges is remarkably consistent. Companies below Tier 3 on that scale do not have AI problems. They have semantic problems, ownership problems, and history problems that AI converts into visible, confident, expensive wrongness.

The organizations that succeed with AI are not the ones with the best models. Everyone's models are increasingly the same. They are the ones where a customer means one thing, where metrics hold their definitions over time, where somebody owns the data and can prove where it came from. None of that is glamorous. All of it is what the model inherits.

So before the next post-mortem blames the machine for hallucinating, it's worth asking the harder question: what did we feed it?