Point a capable assistant at a messy estate and it answers with total confidence and a disputed number. Readiness is not a product you buy. It is four foundations.

AI does not fix bad data. It amplifies it.

AI does not fix bad data. It amplifies it, and it does so with total confidence. Point a capable assistant at a messy estate and it will answer a question with a number three of your departments would dispute, in a clean, persuasive sentence that makes people trust it more, not less. A confident machine on a weak foundation does not reduce your risk. It industrialises it.

So getting ready for AI is not about buying an AI. The model is the part you no longer have to build, supplied by Microsoft and others and improving without your involvement. What no vendor can supply is an understanding of your business: what a customer is, when revenue is recognised, which sites count as active. That lives in your definitions and your data model, and it is what decides whether AI turns out useful or dangerous.

The reassuring part is that readiness is the same work that makes ordinary reporting trustworthy. Do it once and you get both a reporting estate you trust and an AI you can switch on safely.

The four foundations

Agreed definitions come first. Write down, in plain business terms, what your core measures mean, and force the disagreements into the open. Most organisations find they have never actually agreed what "active customer" or "net revenue" means, and that different teams quietly use different versions. That conversation is uncomfortable, and it is the single most valuable part of the whole exercise, because every downstream number depends on it.

One semantic model is second. Each definition should live once, in a single shared model that every report and every assistant reads from. The model, not the dashboard, is what AI reasons over. Scatter your numbers across separate reports each with its own logic and an assistant has no single truth to stand on; it will confidently pick whichever it finds first.

Lineage you can show is third. The moment an assistant produces a figure, someone senior will ask how it was calculated and from which source. Without lineage, from source system through model to answer, you cannot say, and an answer you cannot trace is an answer you cannot defend.

Access and security is the fourth, and the one whose failure is worst. An assistant surfaces whatever it can reach. Row and object-level security, tied to your existing Microsoft identities, has to be correct before you let people ask open questions, or the friendly assistant becomes the most efficient route to a data breach in the building. The failure here is not a bad number. It is the wrong person seeing the right one.

Why the danger is quiet

Bad reporting tends to announce itself. The number looks off, someone checks, it gets fixed. Bad AI does the opposite. It launders a weak number through fluent language until it sounds authoritative, and human nature is to trust the confident, well-written answer and stop checking. The cleaner the prose, the less scrutiny it attracts. That is why readiness is a safety question, not just a quality one. The tool removes the very friction that used to catch errors.

Do it in order, and prove it small

Sequence matters, because each step depends on the one before. Agree the definitions for one important domain, with the people who argue about them. Build that domain into a single clean model with the definitions baked in. Add friendly names, plain descriptions, certification and security. Then let a small group ask questions against it, and check the answers hard before widening. A narrow thing that is trustworthy beats a broad thing that is merely plausible.

So this week, take one measure everyone quotes, "active customer" will do, and ask three people in different teams to define it without conferring. The gap between their answers is the work, and it is the work AI will expose in public if you skip it. Doing it properly is the foundation of our analytics strategy and Microsoft Fabric work, and the full readiness sequence is in our Getting Your Data Ready for AI guide.