Chapter 02

CAM and Urban Governance

What is being deployed, and why cities must define the terms

2.1From autonomous vehicles to connected autonomous mobility

For the better part of a decade, the public debate about autonomous mobility focused almost entirely on the vehicle. Which sensor arrays would prove most reliable? Which companies have vehicles that can operate outside of closed courses? Do they fail in the snow? These are not unimportant questions. Answers drift naturally towards safety regulation and technical readiness certification. Regulators and technologists alike found these questions relatively comfortable – they have quantifiable, technical answers and clear regulatory precedents.

They are, however, a red herring for the institutions that will be responsible for managing the diffusion of autonomous vehicles; the city governments that will ultimately be accountable for autonomy’s impact on communities and public spaces.

As companies actually deploy autonomy into cities this year, the full operating picture of urban movement is coming into focus. It combines autonomous driving with connectivity infrastructure, fleet routing, shipping, siting depots (where vehicles charge, idle, and receive servicing), dynamic curb prioritization, fleet-to-fleet and fleet-to-infrastructure communications protocols, real-time data, and liability. In this paper, we refer to this new operating picture as connected autonomous mobility, or CAM ¹ This paper addresses ground-based connected autonomous mobility. The governance questions raised by autonomous air mobility – drones, vertiports, low-altitude airspace – are adjacent and increasingly material, but sufficiently distinct in institutional and regulatory structure that they merit separate treatment. . To observe it, one needs to take an uncomfortable step back from the self-driving car itself, and regard a much wider field of view.

Reframing from self-driving cars to CAM means more stakeholders will be at the table. Different regulatory frameworks apply. It means a spectrum of decisions must be made, in tandem and on a faster timeline. A city preparing to manage self-driving cars is comfortable passively adopting federal safety and licensing frameworks. A city preparing to govern CAM is building the capacity to meet a systemic challenge that spans transport, logistics, grid utility, data analysis, urban planning, budgeting, and public messaging.

Figure 2.1 Who is at the table. Governing connected autonomous mobility puts labor, utilities, developers and real estate, the fleet operators and the public around the same decision, with the city's own share of it splitting again across transportation, planning, the mayor's office and the auditor. The circles overlap because the responsibilities do; no single one of them holds the whole decision.

2.2What is being deployed?

As of early 2026, deployment has moved decisively beyond demonstration. Waymo operates more than 3,000 vehicles – 3,067 in its December 2025 federal recall filing^EV3227, and about 3,900 by its June 2026 filing^EV3485 – and its weekly ride volume now compares with the bus operations of a mid-sized city.

In China, Baidu’s Apollo Go platform operates what is likely the world’s largest robotaxi service, with authorized trials in cities including Beijing, Wuhan, Chongqing, and Shenzhen^EV3676. More than 1,000 fully driverless vehicles were operating across Chinese platforms by the end of 2025. Europe has seen fewer commercial deployments^EV60, but several EU member states have enacted the regulatory frameworks that will enable them.

The deployments differ less in the questions they raise than in the order they raise them. Every autonomous fleet will eventually confront the same set: who gets the curb and on what terms, who is displaced from the driving seat, where the vehicles charge and what that does to the grid, what data the city can see. What varies by use case is which of those questions arrives first and with urgency.

Passenger services – robotaxis, on-demand shuttles, first and last mile connectors – surface curb access, transit interaction and accessibility first, because those are the conflicts a rider or a bus driver notices immediately. Goods movement – autonomous delivery, middle-mile freight, yard-to-hub – surfaces loading access, time-of-day routing, depot siting and grid load first, because a fleet running sixteen to eighteen hours a day is an energy and land question before it is a traffic one. Service fleets in controlled environments – airports, ports, campuses, hospitals – look lower-risk until they need to extend onto public streets, at which point workforce transition and liability arrive together, and an operating design domain defined for a fenced site meets the public right of way.

None of these sequences holds for long. A passenger fleet that carries parcels off-peak inherits the goods-movement questions by default; a campus shuttle that crosses one public road inherits the passenger questions. The typology below is a map of where each use case bites first.

A governance framework designed around any single use case will be inadequate for the others. The cities and authorities building their CAM governance frameworks in 2026 need to build for the full range of use cases, accounting for the ways in which a single deployment may hybridize multiple use cases.

Table 2.1 Deployment typology: each type creates distinct institutional demands across the Governability Stack.

Type	Examples	First Questions Raised	Primary domains
Passenger services	Robotaxis, on-demand shuttles, first/last-mile	Curb access, service obligations, transit interaction, accessibility	Flows, Space
Goods movement	Autonomous delivery, middle-mile freight, yard-to-hub	Loading access, time-of-day routing, depot siting, grid load	Space, Power
Service fleets	Municipal vehicles, campus, port, airport	Workforce transition, liability, extension to open streets	Flows, Data
Staged pilots	Geofenced trials, hybrid human-AV	Evaluation criteria, exit ramps, precedent management	All four

2.3Technical ODD and the Institutional ODD

The concept of an Operational Design Domain (ODD) is familiar to anyone in the automotive industry. It refers to the conditions under which a given driving system is designed to operate safely: the range of weather conditions, road types, traffic densities, and operational scenarios within which the system is validated and certified. A prosaic example of the ODD is the weight of a trailer that a car can safely pull. If an autonomous vehicle is operating outside its ODD – on an icy road in a storm when it was designed for clear urban conditions – it should hand control back to a human or pull over. The ODD is, in essence, a statement of the technical limits of competence.^EV08

This paper introduces a parallel concept that is missing from most governance discussions: the Institutional ODD. Where the technical ODD describes the conditions under which the system can operate safely, the Institutional ODD describes the conditions under which a city can govern that system effectively – with appropriate authority, adequate information, competent oversight, and legitimate public accountability. A city with a weak iODD hosts CAM operations it cannot observe or shape.

The Institutional ODD

The Institutional ODD refers to the governance conditions required for CAM to operate in a city on terms consistent with public goals. It sits alongside technical ODD but asks an entirely different question.

Technical ODD asks: Can the automated driving system operate safely under these environmental conditions?

Institutional ODD asks: Can this city govern fleet operations effectively under its current political, spatial, infrastructural, and informational conditions?

The Institutional ODD is defined across four domains:

Flows – Does the city have the authority, data, and coordination capacity to manage what moves, when, in what volume, and with what knock-on effects for transit, freight, congestion, and active travel?

Space – Can the city allocate and reallocate curb space, depot sites, loading zones, and charging infrastructure dynamically, with enforceable conditions, in response to real-time demand?

Power – Does the city have the inter-institutional relationships with utilities and network operators to coordinate charging load, depot siting, and grid capacity with fleet deployment decisions?

Data – Does the city have the legal authority, technical infrastructure, and operational capacity to require structured data from operators, audit compliance, and act on what it finds?

A city that does not know its Institutional ODD is operating blind. A city that has scored it has the starting point for a governance program.

iODD shifts the governance question from a binary – are we ready or not? – to a structured diagnostic. Most cities will score well on some domains and poorly on others. London, for instance, has strong Flows governance through the statutory authority of Transport for London (TfL), but weak Power governance through the almost total disconnect between TfL and UK Power Networks. Phoenix has a relatively permissive licensing environment but very limited Data capability. With a solid understanding of their specific gaps, cities can close them in the right order.

2.4Institutional constraints

One of the most important lessons from the first generation of commercial CAM deployment is that institutional constraints remain binding even when technical performance improves substantially. The Cruise case in San Francisco is the defining example. By October 2023, Cruise had scaled to 24-hour driverless operations across most of the city^EV07 – a technical achievement built on years of development and billions of dollars. But when a vehicle was involved in a pedestrian injury incident, the company’s response revealed how thin the city’s governance infrastructure was. The city’s oversight depended entirely on Cruise’s own disclosure, and that disclosure was incomplete and delayed. The regulator’s response – a full suspension of approximately 150 robotaxis – was dramatic, and arguably disproportionate. But it was the only enforcement tool available.^EV07 A city with a proactively designed compliance structure could have had independent access to the data. It could have had proportionate enforcement options short of full suspension. And it could have been able to distinguish between a data-disclosure failure and a safety-culture failure, rather than being forced to treat both as the same thing.

Figure 2.4 Where deployment outruns the institutional ODD. Technology deployment advances on its own schedule; a city's authority, information and oversight reach different distances in each domain. The hatched area between the two is where risk accumulates.

The city’s oversight capacity was designed for a different operating environment, and the gap only showed at the point of failure. Governance infrastructure has to be built before it is tested, not after.

London: governance design before scale

In 2017 and again in 2019, Transport for London declined to renew Uber’s private hire vehicle license. The grounds were substantive: a 2019 investigation found that a change to Uber’s systems had allowed unauthorized drivers to upload their photographs to other drivers’ accounts, enabling at least 14,000 trips to proceed with the wrong driver.^EV09 TfL’s ability to detect and act on this failure depended on the quality of Uber’s operator reporting. That reporting was incomplete, and TfL’s independent audit capability was limited.

The case did not end Uber in London – a magistrates’ court granted the company a new license in 2020 with strengthened conditions, including enhanced DBS (Disclosure and Barring Service) checks, mandatory reporting of incidents, and independent data verification. But it demonstrated a lesson in governance frameworks that applies directly to CAM: the city’s ability to exercise oversight depends critically on the quality and completeness of what operators disclose. A CAM governance framework that relies on voluntary operator transparency will discover its limits at the worst possible time.

Most frameworks do not address legal liability – a gap that both the Cruise and Uber cases highlight. When an autonomous vehicle causes harm, who is responsible? In the United Kingdom, the Automated Vehicles Act 2024 establishes the Authorized Self-Driving Entity (ASDE) framework: the entity that authorizes the self-driving system – the developer or operator – holds primary liability when the system is in control, replacing the driver liability model of conventional vehicle law. This is a clarification that reduces legal ambiguity for both operators and cities. But it does not resolve the governance question of how cities access the data needed to assess whether an incident occurred within or outside the vehicle’s operational design domain – or whether the operator’s disclosure of that determination can be independently verified. Liability frameworks and data governance frameworks are not the same thing; both are required, and the absence of either creates the conditions for a governance failure of the kind that materialized in San Francisco.

2.5Different cities, different governance postures

The global deployment landscape reveals markedly different governance postures, each with its own strengths and risks. Phoenix has operated as one of the most permissive environments for autonomous vehicle (AV) deployment, with Arizona’s state-level regime providing a minimal regulatory floor and the city itself contributing relatively light touch permitting. The result has been commercially useful for Waymo – Phoenix became its primary testing and early commercial environment – but the governance lessons are limited because the city has not needed to manage conflict at scale. Waymo’s Phoenix operation has not yet encountered the curb pressure, transit interaction, grid stress, or political opposition that will define the harder governance questions in denser cities.

New York represents the other end of the spectrum: layered permits, safety-driver requirements, tightly constrained testing geographies, and a city government that has demonstrated it is willing to impose hard limits on fleet operators when public interest requires it. The 2018 pause on new for-hire vehicle licenses – a decision that capped the transportation network company (TNC) fleet at approximately 80,000 vehicles while the city assessed the impact on congestion and driver income – was the first of its kind in a US city and politically difficult to achieve. But it demonstrated that a city with the institutional will and legal framework can exercise meaningful control over fleet deployment. The tools exist. The political conditions for using them must be built.

The London governance scaffold

London’s governance position for CAM is stronger than most comparable cities, but its strengths are unevenly distributed across the Institutional ODD. Transport for London has statutory authority over the strategic road network, a twenty-year track record in operating the congestion charge with multi-agency coordination, and a mature private hire vehicle licensing regime that has already been tested at Uber’s commercial scale. The data model – Oyster and contactless records covering over four billion annual journeys, real-time network performance monitoring, and a public-facing data portal – gives TfL a sensing capability that most city transport authorities would struggle to recognize.

The weakness is in the Power domain. Greater London is served by two distribution network operators (DNOs): UK Power Networks covers most of the city, and SSEN Distribution serves parts of west and north-west London. Neither has a formal coordination structure with TfL in relation to CAM. Both face grid constraints in the areas where CAM depots are most likely to be sited. Fleet depot siting decisions, which carry grid load implications (see Chapter 3), are not currently subject to any joint transport-energy planning process. A megawatt-scale grid connection in the UK takes fifteen to eighteen months end-to-end (Chapter 3). The first commercial CAM fleet to seek depot space in London will arrive in a market where the grid constraint is already binding. Building the Power domain governance – specifically, a formal joint planning process between TfL, UK Power Networks, and SSEN Distribution on depot siting, grid hosting capacity, and managed charging standards – is a high-priority institutional gap for London’s CAM readiness.

2.6The global governance landscape

The regulatory frameworks taking shape across major jurisdictions share a common challenge and diverge sharply on how they address it: how much authority should national, regional or state governments retain over CAM deployment, and how much should flow to cities? The answer varies enormously, and the variation has direct consequences for how much of the governance agenda described in this paper is available to urban authorities at all.

The United States operates the most permissive and fragmented regime. Federal authority through the National Highway Traffic Safety Administration (NHTSA) covers vehicle safety standards, but fleet operations – how vehicles are deployed, at what density, with what service obligations – are governed primarily at the state level, with substantial variation. Arizona’s light-touch permitting enabled Waymo’s early commercial scaling. California’s more structured system, with mandatory incident reporting, passenger disclosure arequirements, and the ability to suspend operating authorizations, provides a meaningful governance floor while still enabling commercial scale. Cities within either state are largely left to construct local governance using whatever instruments they already possess – business licensing, traffic regulation, curb management – without a bespoke statutory framework for CAM. Operators have every incentive to concentrate deployment in the most permissive states, and cities within those states inherit operating conditions they did not negotiate.

The fragmentation runs deeper than the federal-state divide. Under the Dillon’s Rule doctrine that applies in many US states, municipalities can only exercise authority that the state legislature has explicitly granted – they cannot act in the absence of a specific statutory mandate. In states that follow this doctrine, a city that wants to impose data-sharing requirements on a CAM operator, set service area boundaries, or establish equitable access conditions must first secure state enabling legislation. Several states have moved further still: Florida, Indiana, and Utah have enacted pre-emption laws that prohibit local governments from regulating autonomous vehicle operations altogether, reserving that authority at the state level. The effect is to make local governance of CAM structurally impossible in a large portion of the United States, regardless of the political will of city administrations. The contrast with Home Rule states – where cities derive broad inherent authority from their charters, without needing state-by-state permission to act – is stark, and maps imperfectly onto the geography of where CAM deployment is occurring. Cities that need governance tools most may be the ones that have the least legal authority to create them. City administrations and local communities bear the costs of market failure, and they lack the tools to mitigate those costs.

The state-level variation now extends to two operative reference frameworks rather than one. California has operated the dominant US model since 2018: a vehicle-as-permitted two-stage architecture, in which each vehicle class and each commercial service tier require separate authorization from the DMV and CPUC – the state Department of Motor Vehicles and Public Utilities Commission respectively – with broad regulatory discretion in enforcement. Texas SB 2807, which entered enforcement on 28 May 2026, operates an alternative operator-as-licensee model: a single state-level license covers an operator’s entire Texas fleet, fees are minimal and renewals non-expiring, and administrative-action thresholds sit at the high bar of serious bodily injury under the Texas Penal Code.^EV107 The two frameworks differ on authorization-stage rigor, fee structure, enforcement triggers, and public-complaint channel design, and operators that successfully obtain authorization under one face materially different operating conditions than under the other. Beyond the state regulatory framework, the variation extends to whether cities can govern autonomous mobility at all. Many US states pre-empt municipal authority over vehicle operations to varying degrees; Oregon is one of the small number that do not, allowing cities to convene locally negotiated operating frameworks under their own authority. The implication is structural: in some US states a city’s authority is what state law leaves untouched, in others a city can move first and set the terms – and this variation should be visible to any operator, investor, or city peer working across multiple US jurisdictions.

The United Kingdom is building the most deliberately coherent national framework. The Automated Vehicles Act 2024 creates a statutory basis for operating authorization, establishes liability frameworks for automated systems, and lays the groundwork for data-sharing obligations and third-party insurance requirements. The enabling regulations will specify the conditions under which fully driverless commercial deployment can begin^EV3165; the first autonomous services have meanwhile entered under existing private-hire licensing, with safety drivers aboard – TfL granted the first fifteen such licenses in August 2026. This gives UK cities a more robust legal standing than their US counterparts. The Act explicitly preserves local authority powers over road space and conditions of access. Whether local institutions will build the capacity to use that scaffold remains a question.

The European Union’s approach folds the AI Act’s high-risk requirements – risk management, data governance, transparency, and human oversight – into existing vehicle type-approval regulation rather than applying them directly, alongside the proposed AV-specific framework under negotiation; a provisional agreement in May 2026 deferred those requirements to August 2028^EV98. The result will be the most comprehensive ex-ante compliance regime globally: slower to enable commercial deployment, but more robust in its baseline governance guarantees. For cities in EU member states, this means less governance work to do from scratch, but also less freedom to design bespoke local frameworks that go beyond the EU floor.

China has authorized Level 4 trials in more than 20 cities^EV06 under a state-directed model that combines rapid permitting with centralized data requirements and explicit national industrial policy objectives. Baidu’s Apollo platform, DiDi’s autonomous unit, and several other operators are scaling under frameworks that are designed to maximize deployment speed rather than minimize governance risk. The governance construct is top-down and, in important respects, unlike anything available in democratic systems: operators are required to share data with the state as a condition of operating, but that data flows to central rather than local government.

The risk of this model became visible in early 2026 when a simultaneous software fault across a fleet of robotaxis in a major Chinese city triggered a mass operational failure: vehicles lost routing capability simultaneously, halted across intersections and main roads, and could not be overridden or recalled without manual intervention.^EV89 The result was street congestion across multiple city blocks for several hours. The incident illustrates a risk that speed-first governance frameworks leave unaddressed: what happens when a system fails not as an isolated vehicle incident but as a networked fleet event, across an entire urban operating environment, at once. Governance frameworks designed for individual vehicle safety incidents have no established instrument for a networked failure of this kind.

The lessons for European and North American cities are therefore indirect: China demonstrates that rapid deployment and structured data access are compatible, but the institutional model is not transferable – and the failure mode it is creating is one that urban governance frameworks everywhere will need to prepare for.

Gulf states – the UAE and Qatar^EV59, and Saudi Arabia – are pursuing autonomous mobility as part of national economic diversification strategies, with timelines and ambition that exceed most comparable programs elsewhere. The governance frameworks are centrally designed, with limited democratic consultation and no established tradition of local authority involvement in mobility regulation. The cities themselves are new enough to be designed around the technology. Gulf deployments may set commercial precedents in data sharing, service standards, and operator-government relationships. Operators may seek to export these to jurisdictions with different political expectations. Understanding the governance terms under which global operators first scaled matters for every city those operators subsequently approach.

Regardless of jurisdiction, the frameworks being locked in during 2026 and 2027 will shape the operating conditions for a decade or more. The differences between regimes create risks – jurisdiction shopping, regulatory arbitrage – but also opportunities for learning. The cities that engage actively with their national frameworks, extend them where local authority permits, and share learning across jurisdictions will be better positioned than those that wait for national frameworks to tell them everything they need to know.

2.7Counter-arguments

Three objections to the governance-first argument come up regularly.

Objection 1: Cities cannot coordinate across multiple agencies

The claim is that multi-agency, cross-domain governance of the kind the Governability Stack describes is beyond the practical capacity of city administrations that struggle with basic service coordination in normal times. The concern is legitimate; the history of failed urban technology governance is long. That is not a reason to give up.

The counterevidence is equally long. London’s congestion charge has operated for over twenty years, coordinating TfL, the Metropolitan Police, the Mayor’s office, and thirty-two borough councils under a clear statutory framework with formal data-sharing agreements and transparent revenue allocation. Stockholm’s congestion pricing was designed, implemented, tested in a pilot, evaluated, and confirmed by referendum – all within a two-year window, across multiple agencies and political levels. Emergency management systems – Gold/Silver/Bronze in the UK, Incident Command in the United States – demonstrate that real-time, multi-agency, cross-domain coordination works at city scale when the governance model is designed for it. These precedents are survivors of a much larger population of standing cross-domain boards that quietly failed, and the design features the survivors share are the specification: a single accountable executive, and either a statutory mandate or a live emergency trigger.

Objection 2: Governance cannot keep pace with the technology

This is the objection that this paper’s own Chapter 1 raises – and the response needs to be more nuanced than simple disagreement. It is true that technology deployment cycles are faster than institutional reform cycles. It is true that the pacing gap is real and structural rather than merely contingent. But the inference that governance should therefore wait for the technology to stabilize before engaging is backwards. The pacing gap is an argument for starting earlier, not later. Every month that passes without governance action is a month in which path dependencies are being established that will require more effort to shift.

The no-regrets actions described in Chapter 8 are specifically designed to address this objection. They are chosen because they are robust across multiple technology scenarios, because they do not require the city to predict which deployment pathway will dominate, and because they build institutional capabilities that will be valuable regardless of how the technology develops.

Objection 3: operators will choose the most permissive jurisdiction

When ride-hailing was scaling into urban markets, operators concentrated early commercial activity in regulatory environments that required the least from them. But the argument is often overstated as a constraint on governance.

Operators need cities as much as cities need operators. The commercial viability of autonomous fleet operations depends on population density, trip demand patterns, and physical infrastructure that only certain cities provide. Phoenix was permissive, and it was also flat, dry, and gridded. Permissiveness bought Waymo a testbed, but San Francisco is a counterargument. Dense, high-demand cities – which are the cities where the governance questions are hardest – are also the cities where the commercial returns are potentially largest. Operators cannot avoid them, but they often defer them – enter later, after more permissive markets have absorbed the early capital. Deferral is a real threat, and it is partially self-limiting. That can only be framed, hypothetically, as a “cost” to a city if early service is a net benefit, and that depends on governance.

Consider what ‘deterring investment’ has meant in practice. London’s congestion charge, introduced in 2003, was widely predicted to drive economic activity out of central London. The opposite happened. A well-designed governance instrument that creates predictability, transparency, and fair operating conditions does not deter serious operators; it attracts them, because serious operators prefer governed markets to ungoverned ones. When there is a level playing field, with the same data-sharing requirements for every operator, the operators who can perform under structured conditions will win.

Financial regulation faced the same objection. Strong rules were expected to drive business offshore. To some extent they did. But the response to financial regulatory arbitrage has not been to stop regulating. It has been to build coordination mechanisms between regulators, to raise the floor across jurisdictions, and to make the costs of operating in an ungoverned environment clearer.

On all three objections our position is the same: proactive governance is the condition under which deployment goes faster and further. Without it, rollout is a sequence of governance failures – safety incidents, political backlash, grid limits, legitimacy crises – each inviting a moratorium or reversal that sets the technology back by years, or an acceleration that sacrifices public good. Operators can navigate clear operating conditions, predictable data obligations, and transparent enforcement. Under those conditions, autonomy can have a positive impact on cities.

2.8The agency claim

This chapter has built to a single claim that the rest of the paper develops in operational detail. Cities will be subjected to a governance framework designed by someone else – or be entirely unable to govern urban autonomy – unless they actively define the terms of access, the conditions for data sharing, and the compliance obligations for fleets operating in the public right of way.

Building governance capacity starts with understanding the city as an operating environment in its own right. CAM will interact with a system of flows, space, power, and data in ways that are partly predictable and partly not. The agency claim is conditional: it holds where a city has statutory authority to license (or meaningful proxies for licensure), data collection and analysis tools, and the staff headroom to do the work. The Institutional ODD is a framework to score that capacity systematically.