The Data-Quality Platform Fixes the Wrong Problem

In March 2025, Salesforce asked 552 US decision makers about the data behind their choices. Confidence in its accuracy had dropped 27% in two years. KPMG had already measured a version of that same doubt back in 2018, before ChatGPT, before Copilot, before the AI wave took the blame for a problem it did not create. The distrust in our numbers is older than the tooling we keep buying to repair it.

The proposed data professional’s oath has a value for exactly this doubt: interpersonal trust, the second of eight I am walking through post by post as I read the paper properly. I will plant a flag on it. The data-quality platform funded after a bruising quarter of “the numbers are wrong” complaints fixes the wrong problem. Trust comes back when someone can explain the number in plain words to the person who has to act on it.

The case for the quality budget

Let me steelman the other side first, because it deserves it. Bad data genuinely breaks trust. A margin number that shifts between Monday and Wednesday, with nobody able to say why, teaches a room to stop asking and start guessing. Tested transformations, readable lineage, and owned definitions are worth every euro they cost, and nothing in this piece argues against funding them. If your pipeline is genuinely broken, fix the pipeline.

Seven years, same doubt

All true. But KPMG’s 2018 figure put high trust in your own organisation’s analytics at 35%, and that doubt has outlived every quality platform bought in the seven years since. Both surveys come from vendors, so I hold the numbers loosely; the direction is what I believe, and the direction says the next platform will not buy the trust back either. Two decades of accuracy tooling had already run its course by 2018, and the doubt was still there waiting for a cause to blame.

What actually brought people back

Two experiments carry the argument better than either survey can. In 2015, Dietvorst, Simmons and Massey let participants watch a statistical model and a human forecaster make the identical error. The model lost more confidence than the human did, and the participants then put real money on the worse forecaster, the one they had just watched be wrong in exactly the same way. In 2018 the same team found something more useful: participants kept using an imperfect model once they were allowed to adjust its output, even slightly. Accuracy alone had not restored their trust. A sense of control did.

Those were lab forecasts, and the step to a production dashboard is mine to own, not the researchers’. But it matches the meeting rooms I keep finding myself in. Machine-made numbers sit on the shortest leash in the building, and people come back once they get their hands on the number, not once the number gets a little more accurate.

The oath splits trust in two

The oath lists interpersonal trust and trustworthiness as two separate values, and that split is doing more work than it first looks like it is. Onora O’Neill has argued since 2002 that trustworthiness takes competence, honesty and reliability, plus evidence other people can actually check. A number nobody can explain gives the room nothing to check, so the CFO who withholds her trust from it is not being difficult. She is being exactly as careful as she should be, her triad, my mapping onto a dashboard.

I read the oath’s split like this: trustworthiness is a property your work can have on its own, sitting quietly in the pipeline whether anyone looks at it or not. Interpersonal trust is a relation with the specific person who has to act on your number. A quality platform, however good, only ever delivers the first. It cannot sit across the table and answer the question the room is actually asking.

The budget question

So put one question next to every quality line item this budget round, before the accuracy sprint and the lineage tooling get signed off: for every number we ship, who can explain it in plain words to whoever must act on it? Value 1 of the oath, dignity, asked who the columns decide about. Value 2 asks whether the room can still check the number once it is on the screen. Fund the tooling where the data is genuinely broken. Fund the explanation either way, because the platform was never going to buy that part back on its own.

Happy to think this through with you.