Waiting for the data to be ready sounds prudent. The state of the data is often the first answer worth having.
Most conversations about AI with an infrastructure owner reach the same sentence sooner or later, that the data needs sorting out first. The asset register has gaps and half the maintenance history sits in spreadsheets nobody fully trusts. The plan then becomes to clean everything and ask questions once the data is fit to answer them.
The difficulty is that cleaning data without a question has no way of knowing where to stop. Every attribute looks worth fixing until someone says which ones a decision depends on. The work spreads to fill the budget, and the questions that justified it are still waiting at the end of the queue when the money runs out.
The alternative is to ask the question now, of the data as it stands, and treat what fails as a result. If the question is which assets cost the most to keep in service, put it to the records you have. Where they cannot answer, the reason is usually specific, such as a class of assets with no installation dates or a condition survey that only ever covered half the estate. Each of those is a finding about how the organisation has looked after its information, and it points at exactly the data that decision needs.
The pattern in those gaps is rarely random. Data tends to decay where nobody needed it and survive where somebody did, so an attribute left empty for years without anyone noticing was not feeding any decision. Read that way, the records show what the organisation has actually relied on, which is usually a much shorter list than its data standard asks for. The shorter list is the obvious place for any cleaning to begin.
AI changes the arithmetic here. It makes a first attempt at the analysis cheap enough to run before anything has been cleaned, which turns that attempt into an audit of the data the question needed. Its output at that stage is too thin to act on, but it leaves a record of what was missing at the moment it was needed. That record makes a far better brief for a data programme than a general push for completeness.
Some records do have to be right before anything else happens. Safety-critical registers and statutory returns are obligations in their own right, and nobody should wait for a question before keeping them properly. The argument here is narrower. For the questions leadership actually wants answered, waiting for good data delays the work that would show what good needs to mean.
Data is only ever ready for a particular question, so the practical step is to ask one now and let the gaps show where to start.