AI use case

AI traffic signal timing optimization

AI traffic signal timing optimization learns how vehicles move through intersections from sensor, camera or navigation data and recommends or applies new signal timings, so a city can cut stops and delay without building new road infrastructure. A traffic engineer approves each change in the recommendation model.

Last verified 18 November 2026 · 2 public deployments

13.5%
Reported traffic delay reduction
City of Boston, organization claim.
USD 1.5 million to USD 24.3 million
Indicative value per year
A city of 300,000 people that retimes 100 congested intersections with AI recommendations. Worked example, see how it is calculated.

What problem does it solve?

Traffic signals are timed from counts taken every few years, and most cities cannot retime more than a small share of their intersections each year. Between retimings, traffic patterns drift with new housing, roadworks and changed travel habits, so green time goes to streets that no longer need it. Drivers stop more often, wait longer and burn more fuel, and every stop and start adds emissions.

Two kinds of system address this. Adaptive control systems sit at the intersection, sense approaching vehicles and change the green time every few seconds, which needs detectors or cameras and a controller that accepts the commands. Recommendation systems work from data the city already has or can buy, such as navigation trends, and propose new timings that the city's engineers review and load into the signal controllers they already use. The evidence on this page covers the recommendation approach: Project Green Light, which models traffic with machine learning from Google Maps driving trends, and whose recommendations Boston has implemented at 114 intersections and Seattle has tried at five.

How does it work?

  1. Collect traffic data. The system reads vehicle counts from detectors and cameras, or driving trends from navigation data, for each intersection and the ones next to it.
  2. Model the flow. A model estimates how many vehicles arrive, how long they wait, how often they stop and how well adjacent signals work together, by time of day.
  3. Propose timings. The system either computes the next green phase itself every few seconds (adaptive control) or lists changes to cycle length, green split and offsets for the engineers (recommendation).
  4. Review and apply. A traffic engineer accepts or rejects each recommendation and loads it with the city's signal tools. In adaptive control, the engineers set the limits within which the controller acts.
  5. Measure and revert. The city compares delay, stops and travel time before and after, and rolls a change back when it brings no net benefit.
Audience
Back office
Autonomy
Assist
Adoption
Early adopters
Channels
Internal tools

What is it worth?

Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.

Value benchmarks for AI traffic signal timing optimization
KPIMedianReported rangeData pointsClaimed by
Traffic delay reductionToo few to pool
13.5%
11 organization

Value drivers: Speed and cycle time, Lower cost to serve, Customer experience.

Indicative value

A city of 300,000 people that retimes 100 congested intersections with AI recommendations

USD 1.5 million to USD 24.3 million

Value of driver time saved per year

How this is calculated

Formula: intersections * vehiclesPerDay * 365 * delayReduction / 3600 * valueOfTime. The low scenario uses every low input, the high scenario every high input.

InputLowHighBasis
Intersections retimed intersections, intersections100100The reference city.
Vehicles through each intersection per day vehiclesPerDay, vehicles per day15,00030,000Editorial assumption for a congested urban intersection, replace with your own counts.
Average delay saved per vehicle delayReduction, seconds per vehicle14Editorial assumption, kept below the 13.5% average delay reduction that Boston reports on this page (4 to 5 seconds on an assumed average delay of 30 to 40 seconds), because Boston's results are strongest at peak times.
Value of an hour of driver time valueOfTime, USD per hour1020Editorial assumption, replace with your transport authority's value of time.

What it leaves out: Driver time only. It leaves out fuel and emission savings, effects on buses, pedestrians and cyclists, the cost of the license and engineering time, and any intersection where the change was reverted.

Which organizations already do this?

2 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.

This library is not a list of Blits.ai customers. It documents public AI deployments from any vendor, each with its source.

City of Boston

United States · Government and public sector · 2025

ScaledGrade B

Boston's Traffic Management Center uses Project Green Light from Google Research, which models traffic patterns from Google Maps driving trends and recommends new signal timings for congested intersections. Traffic engineers review each recommendation and the city has implemented them at 114 intersections in 20 neighborhoods. The city had INRIX evaluate the phase 2 pilot and reports an average 13.5% reduction in delay and a 20% average reduction in unnecessary stops across newly retimed intersections, with the best results at peak times.

  • Traffic delay reduction: 13.5%, phase 2 of the pilot, newly retimed intersections
    "During phase 2 of the Project Green Light pilot, the City has seen an average 13.5% reduction in delay, with some locations seeing improvements of up to 24%, and a 20% average reduction of unnecessary stops across newly retimed intersections."
    Claimed by: organization

Seattle Department of Transportation

United States · Government and public sector · 2023

PilotGrade C

Seattle tested Google's Green Light machine learning system, which recommends signal timing changes from Google Maps driving trends, at five intersections from 2023. The city's transportation department told Scientific American that it has seen positive results, and also that it reverted one recommended change because it did not result in a net benefit. No before and after figures were published in the source. Transport for Greater Manchester, another pilot city, said its traffic engineers often chose to ignore the recommendations.

No outcome disclosed.

How do you implement it?

A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.

Data you need

  • Current signal timing plans and controller types for each intersection
  • Traffic data by time of day, from detectors, cameras or navigation trends
  • Priority rules for buses, emergency vehicles, pedestrians and cyclists

Systems to integrate

  • Central signal management system or controller programming tools
  • Detector and camera feeds, where adaptive control is used
  • Traffic management center workflow for review and sign off

Complexity: Medium

A recommendation system needs the city's signal timing plans in a form the vendor can read, an engineer review step and a way to load changes into the existing controllers. An adaptive control system also needs detectors or cameras and controllers that accept commands at each intersection, which is the larger part of the cost. Both need a before and after measurement, usually from a third party.

  1. 1

    Pick corridors with measurable delay

    Start with intersections that already have complaints or long measured delay, and keep a set of similar intersections unchanged as a comparison.

  2. 2

    Write down the priorities the model must respect

    Record bus priority, pedestrian crossing times, school zones and emergency routes as constraints, because a model that minimizes stops alone can work against them.

  3. 3

    Review every recommendation

    Have an engineer accept, change or reject each proposal and record the reason, so rejected ones also teach the vendor and the team.

  4. 4

    Measure before and after, and revert

    Compare delay and stops with an independent source, and roll back any change that brings no net benefit.

Guardrails

  • A licensed traffic engineer approves every change before it reaches a controller
  • Fixed minimum pedestrian crossing times and clearance intervals that the system cannot shorten
  • Bus and emergency vehicle priority stay in place
  • A rollback path for every timing change

KPIs to instrument

  • Average delay and travel time on the corridor, before and after
  • Number of stops per vehicle at each retimed intersection
  • Share of recommendations accepted, changed and reverted
  • Bus travel time and pedestrian wait on the same corridor

Human in the loop

Traffic engineers review and approve each recommendation, own the safety timings and decide when to roll a change back. In adaptive control they set the limits the controller works within.

Common failure modes

Recommendations that work against city priorities
A model that minimizes stops can shorten green time for buses or for a residential street the city wants to protect. Transport for Greater Manchester said its engineers often chose to ignore the recommendations. Encode the priorities as constraints and review rejections.
No net benefit
Seattle reverted a recommended change because it brought no net benefit. Measure every change against a comparison and keep a rollback.

What are the risks and rules?

EU AI Act

Depends on design

Annex III point 2 lists AI systems intended as safety components in the management and operation of critical digital infrastructure and road traffic. A recommendation tool that an engineer reviews, with safety timings fixed outside the model, is unlikely to be a safety component. An adaptive controller that changes signal phases on its own can be one, and is then high risk. The design and the control point decide the tier.

Rules that apply

Guidance

Controls to put in place

  • Documented engineer approval for every timing change
  • Safety timings stored outside the model and locked
  • Independent before and after evaluation
  • Log of recommendations, decisions and rollbacks

Frequently asked questions

How much does AI signal timing reduce delay?

Boston reports an average 13.5% reduction in delay and a 20% average reduction in unnecessary stops across newly retimed intersections, evaluated with INRIX, with improvements of up to 24% in delay at some locations. Results differ by intersection and time of day, and Boston saw the largest effect at peak traffic.

Do traffic engineers still decide the timings?

In Project Green Light they do. The system proposes changes and the city's engineers review and approve each one, and cities have rejected or reverted some. Seattle reverted a change that brought no net benefit, and Transport for Greater Manchester said its engineers often ignored recommendations.

Do cities need new sensors for AI traffic signals?

Not for the recommendation approach, which uses navigation data and the city's existing signal tools. Adaptive control needs detectors or cameras and controllers that accept commands at each intersection.

How to cite this page

Blits.ai AI Use Case Library, "AI traffic signal timing optimization", last verified 18 November 2026, https://www.blits.ai/ai-use-cases/traffic-signal-timing-optimization. Licensed under CC BY 4.0. Method: how we verify use cases.

Changelog
  • 18 November 2026: First published

Related use cases

Logistics and transportation

AI agent for freight dispatch and load matching

An AI freight dispatch and load matching agent coordinates moving a truckload for freight brokers and carriers. It reads quote and booking requests from shippers by email, EDI or API, prices them and ranks loads for carriers in a digital marketplace, books dock slots through the scheduling system and flags exceptions, so staff plan lanes and handle disputes instead of quoting by hand. Conversations with carriers and drivers, such as rate negotiation, vetting, check calls and appointment calls, are covered by the freight carrier voice agent.

Deployments
3 public, best grade B
Reported conversion uplift
12%
Uber Freight, organization claim
Government and public sectorLogistics and transportation

AI crowd density and flow monitoring for venues, transport hubs and pilgrimage sites

Crowd density and flow monitoring is AI that counts people on camera feeds and flags crowding, movement against the flow and congestion points so operators can act early. It estimates how many people are in a zone and which way they move, and leaves every decision, such as opening a gate, redirecting a stream or sending staff, to the operators in the control room.

Deployments
4 public, best grade B
Autonomy
Assist
Government and public sectorLogistics and transportation

AI road surface defect detection from vehicle cameras

AI road surface defect detection is computer vision on cameras mounted on council vehicles that finds, locates and rates potholes, cracks and damaged signs during normal rounds. Maintenance is then planned from a regular condition survey of the whole network instead of from resident reports and occasional manual inspections.

Deployments
3 public, best grade B
Autonomy
Copilot
Real estateGovernment and public sector

AI damp and mould risk prediction for social housing

AI damp and mould risk prediction is a machine learning model that scores homes by the likelihood of damp and mould, using repair history, property data and sometimes weather or sensor data. A social landlord inspects and fixes the highest risk homes before a tenant reports a problem or is harmed.

Deployments
2 public, best grade B
Reported accuracy
98%
Wolverhampton Homes, organization claim