Your data isn't wrong. It's just not fit-for-purpose.

Enterprises have spent twenty years making data report-ready, and by that measure the work often succeeded. Dashboards are correct, KPIs reconcile, auditors are satisfied.

Then the first AI agent goes live and breaks on records that never bothered anyone before.

Reports and agents ask different things of the same data, and only one of them notices when the answer looks wrong. A report is read by a person who sees an empty field, assumes millimeters when the unit is missing, and calls a colleague when two fields contradict each other.

An agent does none of this. It reads the field, takes the value, and acts on it.

Agents are greedy. They take what they find, they fill the gaps with assumptions, and they rarely realize or tell you when the assumption was wrong.

Three zeros in a material master

Here is what that looks like in a real SAP system. In the SAP MARA table, GROES ("size") is a free-text field. LAENG, BREIT and HOEHE hold length, width and height as numbers.

Same picture in system after system. And most of the business is fine with it. The monthly report on material counts doesn't query the field. The buyer reads "80 X 80 X 120", recognizes the product, and orders it.

A packaging agent walks straight into it. It needs length, width and height, and it finds three zeros. From there, three things can follow:

1. It uses 0 and plans a pallet that cannot exist.

2. It parses GROES and guesses the unit. Millimeters? Centimeters? A factor of ten is a truckload.

3. It guesses the order of the numbers. Is 80 x 80 x 120 L × W × H, or W × H × L?

And is 80 x 80 x 120 the product, or the pallet it ships on? Nothing in the record says.

No rule is broken here, so no dashboard turns red. The data isn't wrong. It simply cannot be understood by anything that wasn't in the room when it was created.

Readiness is always readiness for something

That same table gets three different verdicts on the same day. AI-ready for monthly reporting. Conditionally ready for procurement with approval. Not ready for an autonomous packaging agent.

Which means no dataset is "AI-ready" in general. And that readiness can be measured, per use case, rather than pursued as a state of perfection.

One more thing follows from the MARA example: this gap has to be closed where the record is created. Most data quality tooling lives in the analytical layer. By the time a record reaches a warehouse it is a copy, and correcting a copy changes nothing about what the agent reads out of SAP tomorrow.

Score your own data this week. Our 2026 whitepaper From Report-ready to AI-ready includes the five dimensions of AI readiness and a 20-question scorecard you can apply to one dataset and one use case. Plus, the maturity bands tell you whether to automate, pilot, or fix first.