← All articles

Mikhail Polyakov

AI in city public transport: what it already does and where it still needs a human

Level: BeginnerReading time: 13 min

Topics: AI, Transport

A city in the evening seen from above: a tram with an empty cab, a stop and an underground metro station, with glowing data lines and recognition frames above them and a thread running from the tram to a dispatch console

The short answer: in a big city's transport, AI has four roles, and they are at different stages. It drives: a driverless tram in Moscow has been carrying passengers since September 2025, and a driverless metro train, according to data as of August 2026, is being tested without passengers, with a driver in the cab. It counts: in the Moscow Region, according to a statement by the regional ministry in August 2026, a neural network spots queues at bus stops. It watches: in February 2026, the Moscow Transport Department reported that a system that detects falls onto the tracks was being tested at one metro station, and a detailed public account of AI cameras at a station was published by Wired magazine, based on TfL documents about a London trial in 2022-2023. And it plans: in Moscow there was a two-week pilot in which AI drew up the timetable, the bus deployment plan and the driver shifts in a single calculation. In every case a human stays close by. For the driverless tram, at its launch in September 2025, the law required a staff member in the passenger area or the cab; in the cab of the metro train under test, according to the mayor in July 2026, a driver sits as required by law; some of the camera alerts are passed to staff (in the London trial, 19,000 out of more than 44,000); and an extra bus is put on the route by the operator after it receives a demand following a signal from the system. Almost all the figures on this topic come from the city and the transport operators themselves; the sources I found contain no independent checks. Below I go through each role: what is confirmed, what is still a test or a plan, and what an ordinary company can learn from it.

Driving: a tram without a driver and a metro train under test

Tram. On 3 September 2025, a driverless tram began carrying passengers in Moscow on route No. 10, "Shchukinskaya metro station - Kulakova Street". According to the mayor's office, it makes stops, opens and closes the doors, follows traffic lights, gives way to pedestrians, switches the points and keeps to the timetable on its own.

There is still a person on board. At the launch, the mayor's office explained that a tram service employee sits in the passenger area or the cab: they watch the traffic, check that passengers have paid and, in an abnormal situation, open or close the doors, but they do not intervene in driving. At that time, federal law required their presence. The city said back then that, together with the Ministry of Transport, it was preparing changes that would allow an operator without a tram driving licence to sit in the cab. The sources for this article contain no adopted changes.

A year later, on 3 September 2026, the Moscow Department of Transport gave these figures:

  • the city has four driverless trams, and some of them still run in test mode: they study the route and build a precise map;
  • including the tests, they have covered more than 46,000 kilometres, and passengers have made more than 140,000 trips;
  • more than 4,200 runs were completed without a single traffic rule violation, according to the Transport Department's own count.

Next come the plans: by the end of 2026, equip another 11 trams (15 in total) and launch driverless trams on two more routes; by 2030, make about two thirds of Moscow's trams driverless, and 90% by 2035. According to the mayor's office, the launch was made possible by an experimental legal regime, that is, special rules for the experiment adopted by the Russian government at Moscow's initiative.

Metro. In January 2026, testing of a driverless metro train began on the Big Circle Line. According to the Moscow mayor in July 2026, it covered more than 4,000 kilometres in half a year, so far without passengers. The AI accelerates and brakes the train, opens and closes the doors and keeps track of the timetable. There is a driver in the cab, as required by law, and, according to the mayor, they have never had to intervene.

So that the train knows exactly where it is, tags have been installed at all 29 stations of the Big Circle Line; an antenna on the lead car reads them, and the train stops with an accuracy of up to 30 centimetres. In February 2026, the Transport Department reported that a track fall detection system was being tested at Pechatniki station: a camera, a lidar (a laser distance sensor) and other sensors detect an object or a person on the rails within a second, the signal goes to the train and to the dispatcher, and the train stops in advance. At the same time, the city promised to gradually install it at other stations after successful tests; the sources I found contain no more recent information about it. The city plans the first trips with passengers for 2027, and by 2030 it wants to make the Big Circle Line fully driverless, all 94 trains.

Railway. On 28 August 2024, a Lastochka train with a driverless system began carrying passengers on the Moscow Central Circle. How this train works now and how many there are, the sources I found do not say. According to Russian Railways on that day, this is the third level of automation: the automation drives the train, while a driver in the cab monitors the situation and opens and closes the doors when passengers get on and off.

What is important to understand here: the kilometres and violation-free runs are counted by the same city that launched the project. That does not mean the figures are wrong, but the sources I found contain no independent safety report.

Watching: what AI cameras showed in the London Underground

London has a detailed public account of video analytics (software that looks for events in camera footage on its own) in transport. Wired magazine obtained, on request, documents from Transport for London (TfL) about a trial at Willesden Green station, which ran from October 2022 to the end of September 2023. AI was connected to the old CCTV cameras and taught to look for 11 kinds of events: from a person on the tracks and an abandoned bag to fare evasion.

Results according to the documents:

  • There were a great many alerts. Over the year, the system produced more than 44,000 alerts, 19,000 of which were sent to station staff in real time. Most of the alerts were about fare evasion, 26,000. At first, staff were supposed to respond to them, but, as the documents say, because of the large number of alerts (more than 300 on some days) and the high detection accuracy, the system was set up to confirm them automatically.
  • The system made mistakes. It flagged children following their parents through the ticket gates as possible fare dodgers, and it could not tell a folding bike from an ordinary one. It failed to detect aggression; one of the reasons given in the documents was a lack of training data. Instead, the system raised an alert when a person raised their arms.
  • There were benefits too. According to TfL itself in the documents, 59 alerts about wheelchair users allowed staff to give them the necessary assistance at a station with no wheelchair facilities. There were almost 2,200 alerts about people beyond the yellow line and 39 about people leaning over the tracks. According to the documents, during the trial staff made far more announcements asking passengers to step away from the yellow line.
  • Privacy questions. Faces in the footage were blurred at first, and data was kept for up to 14 days. Six months in, it was decided to show the faces of people suspected of fare evasion and to keep those recordings longer. There were no signs about the trial at the station, which TfL itself confirmed.

The conclusion I draw from this: besides accuracy, it matters how many alerts people are ready to handle. In London, fare evasion alerts ended up being confirmed automatically, without a staff member deciding on each one.

Counting: queues at bus stops in the Moscow Region

In the Moscow Region, a neural network watches video from 649 cameras of the "Safe Region" system and counts people at bus stops, according to the regional transport ministry, with an accuracy of up to 95%. If a queue stays at a stop for a long time, a signal goes to the Regional Management Centre, and problem spots are shown on a heat map. The transport operator receives a demand to add vehicles to the route or to run larger buses. According to the ministry, the system has analysed 662 stops and helped clear almost 80,000 long queues.

How exactly a "cleared queue" was counted, and how many of them would have cleared on their own with the next bus, the statement does not say. But the setup is telling: here AI only spots the problem and shows it on the map, while the bus is put on the route by the operator.

Planning: timetable, buses and driver shifts in one calculation

A transport operator has three linked plans: the timetable, which bus covers which runs, and the driver shifts. They are usually drawn up one after another, even though moving one run can change the neighbouring intervals, the vehicle's schedule and the boundaries of a shift. That is why plans drawn up separately frequently have to be fitted together by hand, as the publication that reported the pilot below writes.

At the end of August 2026, it became known that an AI model from the Opturan platform (created by ZashchitaInfoTrans, a federal state enterprise under the Ministry of Transport), which calculates all three plans at once, had been tested in Moscow. The pilot ran for two weeks on nine routes of one Mosgortrans site. According to the Ministry of Transport, bus idle time between runs fell by 34%, drivers had fewer night hours, and two buses were freed up and moved to other routes. Separate modelling for the same routes showed a possible 5.2% reduction in the payroll, or 13.4 million roubles a year. This is a calculation, not savings actually achieved, and the figures come from the developer and the ministry themselves.

In my view, of the four roles, this one is the least visible to passengers and the closest to ordinary business.

Where the boundary is: what is still left to people

If you put the examples together, the boundary runs in three places.

  1. Responsibility. At the launch of the driverless tram in September 2025, the law required a staff member on board (in the passenger area or the cab), and in the metro train under test, according to the mayor in July 2026, a driver sits in the cab as required by law. At the launch in September 2025, the city said that, together with the Ministry of Transport, it was preparing changes that would allow an operator without a tram driving licence to sit in the cab. What their status is now, the sources I found do not say, and they do not mention removing the person from the vehicle altogether.
  2. Handling alerts. The London experience shows that AI cameras can produce alerts by the hundreds per day, and handling them becomes a separate job. What to do about a queue at a bus stop is also decided by people.
  3. Checking the figures. The mileage, the violation-free runs, the cleared queues and the scheduling savings in this article were all counted by the operators themselves. That is not enough to decide whether to adopt something at your own company; you need a measurement on your own data.

What a company can take from this

The transport experience is useful not only to transport operators. It can be applied to any company with shifts, routes or cameras: a delivery service, a field service team, a warehouse, a shop.

1. Start with planning, not with a driverless vehicle. Of the examples above, in my view, the easiest thing to transfer to an ordinary company is calculating the timetable and shifts. If you have shifts, call-outs or routes, ask a scheduling service provider for a pilot on your data.

  • Who does it: the manager together with the dispatcher or whoever currently draws up the schedule.
  • How to check: before the pilot, record idle time, overtime and the number of vehicles or people per shift over two to four weeks. After the pilot, compare the same indicators on the same routes or sites.

You can work out the payback like this: monthly benefit = paid hours saved × hourly cost including taxes + the cost of the freed-up vehicle or position. The vehicle or position counts if you have actually given it up or used it for work for which you would otherwise have had to hire or buy. Costs = the monthly subscription, plus the one-off cost of your staff's hours spent on setup. Reducing idle time saves money when it leads to fewer paid hours or vehicles. If the people and vehicles stay the same, it is spare time, not savings.

An example with made-up numbers; substitute your own. The subscription is 40,000 roubles a month, and setup took a one-off 30 staff hours at 1,000 roubles each, that is, 30,000. Idle time fell by 60 paid hours a month at an hourly cost of 600 roubles, which is 36,000 a month: less than the subscription, so the service does not pay off. Another scenario: in addition, one vehicle is no longer needed, and you returned the leased vehicle, no longer paying 90,000 a month for the lease, fuel and insurance. Then the benefit is 126,000 a month, 86,000 remains after the subscription, and the one-off 30,000 for setup is recovered in the very first month. If the vehicle simply sits idle, those 90,000 are not in the calculation.

2. Measure video analytics by the alerts you can handle. Before launching AI cameras, decide who responds to the alerts and how many alerts a day that person can realistically check.

  • Who does it: the site manager together with whoever will handle the alerts.
  • How to check: during the first week, write down how many alerts came in each day, and manually review at least 50 of them, taken from different days and different types of events: how many turned out to be real. If there are more alerts than can be handled, narrow the list of events before people start confirming everything.
  • When this does not fit: if the camera has to recognise people's faces, show the project to a lawyer before launch: they will tell you which personal data protection measures are needed. A sign about the cameras does not replace them, but it is worth warning people on site.

3. Keep a person with the right to stop it. As the mayor's office described it at the launch, the staff member on the driverless tram opens or closes the doors in an abnormal situation and deals with any issues that arise, and in the metro train under test, a driver sits in the cab for safety. If your AI sends emails to customers, changes prices or schedules on its own, appoint a person who checks the result and can stop everything, and write down in which cases they do so.

4. Check the supplier's figures on your own data. How to measure the benefit of AI in hours before and after, rather than in the number of closed tasks, I covered in the article Developers and analysts working with AI: how their work differs and how to check the results. The same approach works for any pilot.

Summary

As of the beginning of autumn 2026, AI in Moscow's public transport carries passengers in a driverless tram, and in the Moscow Region, according to a ministry statement in August, it counts queues at bus stops. The driverless metro train, according to data as of August 2026, is being tested without passengers; the track fall detection system, according to a Transport Department report in February, was being tested at one station; one-calculation scheduling has gone through a two-week pilot; and the detailed experience with AI cameras at a station remains the London trial of 2022-2023. In the sources I found, there is a person close by everywhere: the staff member on the tram and the driver in the metro train, the people who receive the camera alerts, the operator who puts a bus on the route. In my view, what transfers most readily to an ordinary business from this experience is not a driverless vehicle but calculating shifts and timetables, checked on your own data, and an honest count of how many AI alerts people manage to handle.

I work on AI agents and automation. If you would like to see my projects or discuss your own task, take a look at my portfolio.

Sources