Nearly every leadership team now owns an AI mandate. The board wants a position, the minister or the CEO wants visible progress, and every vendor in the building has renamed its product line. The pressure is real and mostly healthy. What is not healthy is where the money goes first: models, copilots, and platform upgrades, bought before anyone has looked hard at the data those systems will run on.
We would put the pattern plainly. Organizations do not usually fail at AI because they chose the wrong model. They fail because the model met their data.
The pilot works. The rollout doesn't.
The stalled AI program has a recognizable shape. A pilot runs on a hand-picked dataset that one motivated team cleaned by hand. The demo is genuinely good. Then the rollout begins, and the system meets the estate as it actually is: the same customer or citizen represented differently in four systems, fields that changed meaning in 2019 without anyone recording it, extracts of extracts with no owner, and access rules that exist mostly as habit. The pilot's accuracy does not survive contact, and the program quietly becomes a working group.
None of this is an AI failure. Every one of those problems predates the model and would sink a warehouse migration or a BI program just as surely. AI simply exposes them faster and more publicly, because its outputs are visible to executives and, increasingly, to the public.
The model is the last mile. Readiness is the estate underneath it: data that is described, owned, governed, and reachable.
Readiness is not something you can buy
There is no platform tier called ready. Readiness is a property of the estate, and it decomposes into ordinary, checkable things. Somebody owns each dataset that matters, by name. The data is described well enough that a new team can find it and know what it means without folklore. Quality is measured where the data is produced, not patched where it is consumed. Access is a policy that can be read and audited, not a stack of tickets. Lineage exists, so when an answer looks wrong, someone can say where it came from.
Notice that nothing in that list mentions a vendor. Fabric, Databricks, and Snowflake can each host this discipline well, and none of them can supply it. That is why we treat the AI-readiness question and the platform question as separate decisions, taken in that order. An estate with owners, definitions, and governance can adopt almost any platform and any model family. An estate without them will underdeliver on the best platform money can buy.
This is also, quietly, good news. Model choices age in months. The readiness work ages slowly, transfers across vendors, and keeps paying whether your next initiative is a chatbot, a forecast, or a warehouse consolidation. It is the rare investment in this space that does not depend on guessing the future correctly.
The public-sector version is stricter
For government organizations, readiness carries obligations that private-sector playbooks skip. Decisions assisted by an AI system have to be explainable to the person they affect, to an auditor, and sometimes to a tribunal. Records of how a decision was reached fall under retention and disclosure rules. Privacy law constrains what data can be combined, and residency requirements constrain where it can live. Frameworks like the federal Directive on Automated Decision-Making make impact assessment a gate, not a courtesy.
Every one of those obligations lands on the data layer. You cannot explain a decision if you cannot trace what fed it. You cannot honour retention rules for data nobody has catalogued. This is why we tell public-sector clients that governance is not the compliance tax on their AI program. It is the program. Done properly, the readiness work and the accountability work are the same work.
Where to start
Not with a platform selection, and not with a moratorium either. The organizations that get this right start narrow and honest: pick the two or three decisions where AI plausibly earns its keep, then assess the data behind them. What exists, who owns it, what shape it is in, what law and policy say about using it. That assessment produces something a leadership team can actually act on: a short list of use cases that are ready now, a costed path for the ones that are not, and a defensible answer to the board's question of why the exciting demo is not in production yet.
Standing still has a cost too, and it is worth stating. The mandate does not wait for the estate. Teams under pressure will ship something, and if the governed path is not ready, the ungoverned path is what ships.
If you are somewhere in this story, between the mandate and the estate, write to us.