AI data readiness
AI data readiness is the degree to which your data is good enough for AI to handle one specific use case. It is judged per use case on four questions: does the data exist, can it be reached, does it mean the same across systems, and is it complete and recent enough. A full clean-up beforehand is not required.
Readiness belongs to a use case, not to a company
The question "is our data ready for AI?" has no answer until you say what for. Data that is more than good enough to sort incoming quote requests can be unusable for forecasting demand. That is also how data quality is usually defined: data is of high quality when it is fit for its intended use. The same dataset can be ready and not ready at the same time, depending on the question you put to it.
The myth that stalls projects
"First all our data has to be in order." It sounds responsible, and it usually leads to a migration or clean-up programme with no end date. In the meantime every AI use case waits. A first use case typically touches a limited set of fields. Those need to be right. The rest can wait until a use case actually needs it.
Four questions per use case
| Question | What you check | Warning sign |
|---|---|---|
| Does the data exist? | Is it recorded anywhere at all | The knowledge sits in someone's head or inbox |
| Can it be reached? | Is there an export or connection you are allowed to use | The data is locked in a system nobody can get into |
| Does it mean the same everywhere? | Do systems and departments use the term the same way | Sales and finance count an order at different moments |
| Is it complete and recent enough? | Enough history, no gaps in the fields that matter, updated often enough | The fields you need are half empty or a month old |
The third question is the one most often missed. When a term means different things in different systems, a semantic layer records the meaning once, on top of the systems you already have.
An example
Suppose a wholesaler wants AI to enter orders that arrive by email into the ERP. That needs product codes, customer numbers and agreed prices. If those exist, can be reached and are current, the data is ready for this use case, even if the sales history from five years ago is messy. If the same wholesaler wants to forecast demand, that history suddenly matters, and the answer may well be no.
The useful answer per use case comes in three forms: ready, ready after a limited repair, or not yet.
Who should judge it
Not a data engineer on their own. Whether a term means the same thing everywhere is a question for the departments that use it, and the answer has consequences for how people work and are measured. That is why an AI Readiness Scan looks at the data behind each opportunity before the opportunities are ranked.
How to get the data behind one use case in order in practice is covered in this article on preparing your data for AI agents.