Everyone warned me that AI would expose our data problems. That prediction turned out to be true and, in the way it usually gets discussed, completely misleading.
Yes, when we pointed AI tools at parts of our data, the cracks showed. Inconsistent records, duplicated sources, fields repurposed years ago, datasets whose documentation lived in someone’s head. No surprises for anyone who has worked…well, anywhere.
But here’s what the warnings miss. AI didn’t expose a data problem. It exposed how long our people had been quietly working around one.
That distinction matters more than it first appears. The data issues weren’t news to the people close to the work; they’d known for years. What they’d done about it was what good, conscientious people always do: compensate. They knew which report to distrust and which spreadsheet held the corrected numbers. They knew you had to cross-check that system against this one before believing either. They carried translation tables in their heads and applied silent fixes on the way through. The organisation’s data looked functional because a layer of human workaround was continuously repairing it, invisibly and for free.
AI broke the illusion, because AI consumes data without the folklore. It doesn’t know the report lies or that the real figures live in Sandra’s spreadsheet. The workarounds that made the problem survivable also kept it unmeasured and unfunded, and the moment a machine bypassed the humans, the true state surfaced.
So the honest accounting isn’t “AI revealed bad data”. It’s that our data debt had been serviced for years by unacknowledged human effort, and we’d been reading that effort as data quality. Which reframes the fix, too. It’s not just cleansing datasets. It’s harvesting the workaround knowledge, the folklore is a map of every defect, before the people carrying it move on. Their knowledge is the remediation backlog, already prioritised by operational pain.
Ask your teams what they routinely correct before trusting the numbers. The answers will be long, specific and slightly embarrassing. Every one is a defect your dashboards have never shown you, and your agents will find them all.