Most Companies Do Not Have an AI Problem. They Have a Data Readiness Problem.

When an AI initiative stalls, the postmortem almost always blames the model, the vendor, or the use case. Look closer and the failure is usually two layers down. The model was fine. The use case was fine. The data the whole thing depended on was scattered across systems that do not talk to each other, inconsistent in ways nobody had measured, and governed by rules that existed only as habits in specific employees' heads.

That is not an AI problem. That is a readiness problem, and it existed long before anyone said the word "model" in a meeting.

The uncomfortable inventory

Here is an exercise I put in front of operators considering an AI investment. Before evaluating a single vendor, answer four questions honestly:

1. What data do you actually own? Not what your systems theoretically log. What you could produce, today, in a usable format, with a straight answer about where it came from.
2. What condition is it in? Duplicates, gaps, conflicting definitions of the same field across departments, the customer record that exists four times with three spellings.
3. Who is allowed to see what? If you cannot state your access rules, you do not have governance. You have hope, and an audit finding waiting to happen.
4. Where would intelligence actually land? Which specific decision, made by which specific person or system, would change if better information arrived?

Most organizations can answer none of these crisply. Which means the AI conversation they are having is premature by exactly one layer of infrastructure.

Why the industry skips this step

Nobody sells data readiness with any enthusiasm, because readiness is unglamorous and demos badly. A model that writes a paragraph is a demo. A pipeline that reconciles your inventory data across three warehouses is a diagram. So the market pushes companies straight to the visible layer, and the invisible layer stays broken underneath it.

The result is a pattern I have now seen enough times to call it a rule: an AI system built on unconditioned data does not fail loudly. It fails plausibly. It produces answers that look right, sourced from data that is quietly wrong, and the organization trusts the output precisely because it arrived through an impressive interface. A broken spreadsheet at least looks broken.

Readiness is a progression, not a project

The way out is to treat data as something that moves through stages, each one a precondition for the next. The progression I use is the one TacTech is built around: owned data, then conditioned data, then governed intelligence, then operational delivery.

The order matters more than the labels. Conditioning before governance wastes work, because you clean data nobody is allowed to use. Governance before delivery produces a compliant warehouse that changes nothing. Intelligence before any of it produces the plausible failure described above.

The stage almost everyone skips is the last one: delivery into actual operations. Data readiness is not finished when the data is clean. It is finished when the intelligence built on it reaches the person or system making the decision, in time to change the decision. Anything short of that is inventory, not infrastructure.

What this means if you are deciding now

If you run a company and the AI question is on your desk, the sequence is straightforward, if not easy. Audit what you own. Condition what matters, and only what matters: readiness for a specific use beats readiness in general, and boiling the whole data estate is its own failure mode. Define who can see what before you build anything that can see everything. And pick the single decision where delivered intelligence would pay for the whole effort, because that decision is the honest business case.

Do the layers in order and the model choice at the end becomes almost boring. Skip them and no model, at any price, can save the project. The companies that get value from AI in the next few years will not be the ones with the best models. Everyone will have good models. They will be the ones whose data was ready to meet them.