AI for UK Smart Cities: Control Before Scale

A “smart city” is not a dashboard full of live dots. It is a set of public decisions: which junction gets priority, which bridge receives inspection, where energy is reduced, which neighbourhood receives attention, and what happens when a sensor or supplier fails.

AI can help forecast demand and detect patterns across infrastructure. It can also make an opaque preference look objective or turn a bad data feed into a physical action. The safe design unit is therefore a public outcome with an owner, lawful authority, evidence, limits, fallback and route for challenge—not a model looking for somewhere to be deployed.

This guide reflects UK policy and regulator guidance available on 31 July 2026. Public-law, planning, transport, equality and data-protection requirements depend on the authority, nation and use case, so each deployment needs its own legal analysis.

Map the Chain From Sensing to Public Effect

Begin by drawing four stages. A team that cannot explain the final effect should not procure the first sensor.

A model that only proposes an inspection is not equivalent to one that changes traffic signals automatically. Classify the system by its highest-consequence action, including integrations that may be added later. Closed-loop cyber-physical control demands stronger testing, separation and manual recovery than an analytical dashboard.

Define the Outcome Before the Procurement

Write a one-page outcome contract. It should state the public problem, current baseline, affected groups, authority responsible, intervention, expected mechanism, measures and unacceptable outcomes. “Use AI to optimise traffic” is not an outcome. “Reduce unreliable bus journey time on these corridors without worsening pedestrian delay, emergency access or collisions” can be tested.

Select the simplest method that can answer the question. A fixed timing plan, statistical threshold or better maintenance process may outperform an ML system while remaining easier to explain and operate. The government’s AI Playbook emphasises lawful and ethical use, security, meaningful human control, lifecycle management, openness and choosing the right tool rather than defaulting to AI.

Budget for sensors, integration, assurance, staff, communications, maintenance and exit—not only a model licence. The public service must remain operable after the pilot budget or vendor contract ends.

Establish Lawful, Necessary Data Use

Public space is not a data-free zone. A face, number plate, device identifier or repeat movement pattern may be personal data even when the project describes it as “footfall”. Combining nominally anonymous streams can increase identifiability.

For processing under the public-task basis, ICO guidance says the task or official authority must have a clear basis in law, the processing must be necessary, and a less intrusive reasonable route defeats that necessity claim. Record the exact power or function, the purpose and why each field is needed. A supplier cannot choose the authority’s lawful basis on its behalf.

The ICO requires a Data Protection Impact Assessment for high-risk processing, including systematic large-scale monitoring of publicly accessible places. Its DPIA guidance also flags innovative technology, large-scale profiling, biometric data and joined datasets. Start the DPIA while alternatives are still possible, not after cameras are installed.

A useful minimisation sequence is: can the outcome be measured without identifying anyone; can counting happen at the edge; can raw data be discarded immediately; can zones replace precise paths; and can the retention period be shortened? Test re-identification and linkage risk, access controls, deletion and supplier support access. Publish clear notices and a contact route in formats people can actually use.

For the wider compliance pattern, see data privacy and AI compliance.

Test Equality by Place and Group

A citywide average can hide a neighbourhood that becomes harder to reach. The Public Sector Equality Duty requires relevant public authorities to have due regard to eliminating prohibited conduct, advancing equality of opportunity and fostering good relations. Government PSED guidance also stresses understanding and monitoring actual effects.

Complete an equality impact assessment early enough to change the design. For transport, examine disabled pedestrians, wheelchair users, older people, school routes, night workers and people without smartphones. For planning or maintenance prioritisation, test whether historic complaint or inspection data under-represents places where residents have less access to reporting channels.

Measure errors and outcomes by geography and relevant groups where lawful and feasible. If collecting protected-characteristic data would itself be disproportionate, use carefully chosen accessibility audits, qualitative engagement and area-level indicators; document the limitations. A model should not be declared fair merely because it does not ingest a protected field.

Publish the Decision System, Not a Slogan

The Algorithmic Transparency Recording Standard is mandatory for central government departments and specified frontline arm’s-length bodies, not currently for every council. However, the government’s Data and AI Ethics Framework recommends ATRS use across the broader public sector, including local government, even where it is not mandatory.

A useful public record states:

• the service problem and why an algorithm is used;
• where it influences a decision or acts automatically;
• the data categories and provenance;
• performance, limitations and known differential effects;
• the senior responsible owner and operational team;
• human review, override and appeal routes;
• impact assessments and supplier roles; and
• the date, model version and change history.

The detailed ATRS guidance expects intelligible explanations and risk information. Update the record when a pilot becomes operational or when data, model, purpose or process changes materially. Transparency can omit genuinely sensitive security details without reducing the public explanation to “AI improves efficiency”.

This connects directly with AI in government and public services.

Procure the Ability to Govern and Leave

An authority cannot assure a system it is contractually forbidden to inspect. Build the following into market engagement, tender evaluation and contract management:

• access to data provenance, evaluation methods, known limitations and change logs;
• authority approval before material model, data or subprocessor changes;
• security architecture, vulnerability handling and incident-notification periods;
• audit access and assistance with DPIA, equality and transparency records;
• ownership and permitted reuse of authority, resident and derived data;
• open interfaces and export in documented formats;
• defined service levels for safety-critical support;
• deletion, migration and operational continuity on exit; and
• rights to suspend automated action and use the manual process.

The Procurement Act guidance collection organises the commercial lifecycle around planning, defining the requirement, procurement and contract management. Evaluate whole-life operating risk and public value, not only demonstration accuracy or initial cost. A small pilot should not quietly create a citywide data dependency.

Secure the Connected Place as Operational Technology

Connected-place systems move sensitive data and may control physical infrastructure. The NCSC’s Connected Places Cyber Security Principles cover understanding impacts and risks, governance and suppliers; designing architecture, exposure reduction, data protection, resilience and monitoring; and managing privileges, supply chains, lifecycle and incidents.

Apply network segmentation so a compromise in public Wi-Fi, a dashboard or one sensor fleet cannot directly reach signal control or building plant. Give devices unique credentials, remove default access, minimise remote administration and log privileged actions. Maintain a complete asset and software-component inventory, patch policy and supported-life dates.

Design fail-safe behaviour for missing, delayed or implausible data. A traffic controller should revert to an approved safe plan; a predictive-maintenance alert should not disable an asset; a building optimiser should respect hard environmental and life-safety limits. Run incident exercises that include the operator, cyber team, supplier, communications team and emergency partners. Test restoration from trusted configurations, not only detection.

For the built-environment design layer, continue with AI in architecture and urban design.

Pilot in Shadow Mode Before Touching Infrastructure

Consider a proposed traffic-signal model. This is a workflow example, not a claim about a particular city.

First replay historical data and test unusual days, roadworks, sensor outages and emergency routes. Then run the model live in shadow mode: it proposes timings while the approved controller remains in charge. Compare proposals with the baseline across bus reliability, pedestrian delay, queue spillback, cycling, emergency access and safety proxies, by corridor and time period.

Next permit bounded recommendations to trained operators, with reasons and limits. Only after safety, equality, cyber and operational acceptance should the authority consider automatic control—and then within a defined area, schedule and command envelope. Every command should be attributable to a version and input state. If the feed drifts, latency rises, a threshold is breached or the operator loses visibility, the system returns to the approved safe plan.

Public engagement should happen before the decision is irreversible. Ask affected communities about outcomes, blind spots, notices and challenge routes; do not ask them merely whether they “support AI”.

Measure Public Outcomes and Guardrails Separately

Use a scorecard with both service and harm indicators:

Do not collapse these into one optimisation score. A small improvement in traffic flow cannot mathematically cancel a safety breach or unlawful surveillance.

Release Gates for Production

A senior responsible owner should sign production only when:

• the public outcome, statutory function and accountable owner are documented;
• non-AI and less intrusive options have been assessed;
• the data map, lawful basis, DPIA and retention controls are approved;
• equality and accessibility impacts have been tested with affected users;
• baseline, shadow and stress testing cover representative and adverse conditions;
• physical actions stay within independently defined safety limits;
• manual operation and fail-safe recovery work without the model or vendor;
• security architecture, supplier access, patching and incident exercises pass;
• procurement terms provide evidence, change control, portability and exit;
• an understandable public record and challenge route are live; and
• outcome and guardrail metrics have named thresholds, owners and stop triggers.

Repeat the gates after a material data, model, sensor, supplier, geography or purpose change. A pilot’s approval is not a permanent licence to expand.

The credible smart city is not the one with the most automation. It is the one that can explain each public effect, prove that the system works across communities, keep essential services safe under failure and stop the technology when the evidence no longer supports it.