Europe’s First Uber Robotaxi Still Has A Safety Operator

Europe’s first Uber-bookable autonomous ride launched in Zagreb this week. A licensed safety operator sat behind the wheel.

That detail is not a footnote. It is the whole operating model.

What the launch actually splits

Three companies built this service, and each holds a different piece. Pony.ai supplies the autonomous driving technology, its Gen-7 system. Verne owns the fleet and operates the service on the ground, and describes regulatory enablement as part of what it does. Uber puts the rides in its app and handles the rider experience.

Then there is a fourth role that belongs to none of them cleanly. A licensed safety operator has to sit in the car.

The safety operator is the part nobody planned for

Read the announcements and the technology gets the headline. Yet the service does not function today without a licensed person monitoring each ride, and the companies have not said when that person leaves.

Verne has been running a commercial robotaxi service in Zagreb since April. It reports thousands of completed rides, a figure that comes from Verne and has not been broken down. Fleet size is undisclosed. So is any timetable for removing the safety operator.

That is not a criticism of the deployment. It is a description of how serious AI deployments actually work, and the safety operator is the clearest part of the picture.

Automation created a job here

The convenient story about autonomy is that it removes the driver. In Zagreb the driving task changed shape and the safety operator role appeared beside it.

A safety operator is not a driver with less to do. The job is sustained attention without action, which is harder than driving, not easier. It requires licensing, training, a shift pattern and a supervisor. The safety operator also carries the liability that the technology cannot yet hold on its own.

Uber makes the same point in the announcement itself, saying it expects autonomous vehicles and human drivers to coexist for the foreseeable future.

Your organisation will not staff robotaxis. It will do the equivalent.

Where the safety operator role shows up at a desk

Deploy an AI system into an underwriting decision, a support queue or a hiring funnel, and the same shape appears. You will not call the role a safety operator. It does the same job.

Somebody reviews the outputs the system is least sure about. The escalation path, for when it produces something odd, belongs to a named person too. A third has the authority to switch it off. Those are jobs, with names, budget lines and reporting lines, and organisations routinely stand up the system without standing up the roles.

The result is a control that exists on paper and nowhere in the staffing plan.

The economics that get skipped

Here is where the business case usually goes wrong.

The savings model assumes the safety operator comes out. In Zagreb the human is still in the car, months after the commercial service began, with no announced exit date. That is a cost line, and it may run for years.

An honest AI business case prices the oversight function for the full period it will actually be needed, rather than for the pilot. Getting that wrong does not just miss a number. It creates pressure to remove the safety operator before the evidence supports it, which is exactly the decision you never want made under budget pressure.

Four questions before your next deployment

• Which specific person holds the oversight role, and is it in their job description rather than in a project document?
• What competence does that role require, and does your training plan build it?
• What would have to be true before that role is reduced or removed, expressed as evidence rather than as a date?
• Who carries the liability during the supervised phase, and does your vendor contract say the same thing you do?

The last one matters more than it looks. In Zagreb three organisations share the delivery, and the safety operator sits inside one of them. Uber states that any vehicle must meet applicable regulatory requirements and its own safety standards before operating on the platform, which is a gate rather than an answer to the liability question. Contracts across an AI value chain rarely say who answers when the supervised system gets it wrong.

Adoption is an organisational change

The Zagreb launch is a useful public example because every piece is visible. Technology from one company, operations from another, distribution from a third, plus a licensed human holding the whole thing together while the evidence accumulates.

Strip out the vehicles and that is a description of most enterprise AI deployments. The technology arrives quickly. The roles, the training and the accountability take longer, and they are what determine whether the deployment holds.

Treat the safety operator as a temporary inconvenience and you will underinvest in exactly the layer that makes the system safe to run. Treat it as the operating model and the rest follows.

What to do this quarter

Map the oversight roles your current AI systems require, then check them against your actual org chart. Not the intended state, the current one. Every place a safety operator function is implied, name the person who performs it.

Where a role exists in a policy document but not in a person’s objectives, that is your gap. It is also the cheapest thing on this list to fix, because it needs a conversation rather than a project. The job titles carrying these duties have already left the tech department.

Your AI lead can design the control. Employing someone to operate it is a separate decision, and it is the one that gets skipped. If your organisation is working out which roles AI adoption actually creates, the Future Prep Applied training programmes are built for the people who will hold them.