Companies that run on software have a habit when a new rule arrives: change as little as possible and keep the business model. Five European platform cases, from Glovo and Deliveroo to Airbnb, Uber and Meta, show why that stops working once decisions are made by code. A patch to software is stored, dated and linked to other parts of the system, so it grows in a way a patch to a policy never does.
Why a patch on software grows
A revised policy memo is one document. A patch to software is a change to code, a contract and a workflow, and each of them leaves a record. If a regulator asks, those records show what the company knew and when it chose to do the least.
A change in one place also moves things nearby. Glovo let couriers refuse orders freely, which left it short of couriers at busy times, so it added bonuses and pay changes, and those needed their own corrections. Each correction looked cheaper than a redesign. Added together, they cost more than one.
When the fix holds
Deliveroo, in the same business, faced a much narrower question in Britain: were its couriers required to make deliveries themselves? It gave couriers a real right to send a substitute and did not discourage them from using it. The tribunal looked at the contract and at what happened in practice, accepted the right as genuine, and rejected the union's claim.
That fix held because it was small, it was real, and it touched nothing else in the app. Airbnb's handling of EU tax reporting is similar. It freezes a host's payouts until the host supplies tax details, which is a separate step that never touches ranking or pricing. The test is whether the rule reaches the decisions the business runs on. Spain's rule reached nearly all of Glovo's.
The patch most AI users will reach for
For AI, the most common patch will be a person who approves. Uber used it. Four drivers challenged account bans for suspected fraud, and Uber said staff had reviewed each case. In April 2023, an Amsterdam court ruled against Uber for three of the four drivers. Uber had not shown what the reviewers saw or whether they could change the result, and the court called their involvement not much more than a symbolic act.
The bigger penalty came later. In August 2026, the Dutch data protection authority fined Uber about €825 million over driver accounts that software deactivated between 2018 and 2022 without meaningful human involvement. Uber is appealing.
The reviewer who approves whatever the AI already decided is the pattern we described in September. It is easy to put in place, and it leaves a log that shows how little the reviewer did.
For example, a staffing firm uses AI to rank 600 applications for warehouse jobs and adds a step where a coordinator approves the shortlist. The coordinator sees only the top 20, with no reasons, and has never changed a list. On paper there is a human in the loop. In the log there is a click that has never made a difference.
The same step could be real. The coordinator sees the reasons for each ranking and a sample of the rejected applications, and can move people up. Checking 50 rejections a week at 30 seconds each is about 25 minutes. Every change the coordinator makes is recorded.
What to do before approving a fix
Before you approve a workaround, write down what will make you redesign instead. It could be a letter from a regulator, rising complaints, or a fix that needs its own fix, and one named person decides. Then count how often your human reviewer changes the AI's answer. A review that has never changed an outcome will not look like a review to an inspector.