Uber's €825M fine: automate the decision, keep the human

If you build automation that can cut someone off — a driver, a seller, a customer, an employee — the expensive part is not the model. It's the appeal path you didn't build. Reuters reported on August 21 that the Dutch Data Protection Authority has fined Uber €825 million ($966 million) for deactivating driver accounts by automated system without meaningful human review. That is the second-largest GDPR penalty ever issued, behind Meta's €1.2 billion in 2023.

What actually happened

Per Reuters, which reviewed the August 17 decision, Uber's systems suspended European drivers between 2020 and 2022 on suspicion of fraud — flags like unnecessary detours and unfinished trips. Some drivers with low customer ratings were deactivated permanently. The regulator found Uber breached two rights: the right not to be subject to solely automated decisions with significant consequences, and the right to be informed.

That first one is GDPR Article 22, and it is narrower than most people assume. It doesn't ban automated decisions. It says that when a decision materially affects someone and no human meaningfully reviewed it, the person is entitled to human intervention, an explanation, and a way to contest the outcome. Uber says it strongly disagrees, notes its current policies include human review and dispute handling, and plans to appeal.

Note the timeline. The conduct ended in 2022. The fine landed in 2026. Regulatory exposure on an automated workflow does not expire when you fix the workflow.

Why automated decision-making rules matter for your business

You are probably already in scope. You don't need a fleet of drivers. A fraud rule that auto-cancels orders, a scoring model that rejects applicants, an agent that suspends a marketplace seller, a churn model that silently downgrades an account — all of it is a decision about a person with a real consequence. If a human never actually looked, "the system did it" is the finding, not the defense.

A human in the loop means a human who can say no. Rubber-stamping isn't review. The reviewer needs the inputs the model used, the authority to reverse it, and enough time budget to use that authority. If your ops team clears 400 flags an hour, you have a queue, not a control.

Log the reason, not just the result. Article 22's information right is the one people fail quietly. When your agent deactivates something, write the triggering signals, the model version, the threshold, and the reviewer to an append-only record. That log is what turns a five-minute answer into a five-month investigation when it doesn't exist.

Build the appeal before you ship the automation. A contest path — a form, an SLA, a named owner, a reversal that actually restores state — is a week of work up front and effectively unbuildable under regulatory pressure two years later. We build it into the first version, because the alternative is retrofitting an audit trail onto decisions nobody recorded.

Sources: Reuters via Insurance Journal, NL Times.