AI and the Economics of Transportation: Who Wins, Who Pays, and Who Gets Left Behind?

Engr. Kamran Abbas

BSc Civil Engineering

MS Transportation Engineering

Introduction: The Economic Question Behind Transportation AI
How AI Can Transform Transportation Economics
Where AI Creates Economic Value
The Real Cost of AI
The Productivity Paradox: AI Cannot Pave a Road
Efficiency Versus Equity
Data, Labour and the New Transportation Economy
AI Risk, Failure and Resilience
Market Power, Inequality and the Transportation AI Divide
How to Evaluate Whether AI Is Worth the Investment
The Future: AI as Transportation’s Decision Layer
Conclusion: Who Wins, Who Pays, and Who Gets Left Behind?
References

Quick Links

Health and Fitness Tools

Engineering Calculators

Finance Calculators

Economics Calculators

Percentage Calculators

Sciences Calculators

Math Calculators

Unit Converters

Introduction: The economic question is bigger than “Will AI save money?”

Transportation has always been an economic system.

A road, railway, bus network, port, airport or logistics corridor is not simply a physical asset. It is a mechanism for allocating time, space, energy, capital, labor and access to economic opportunity. When transportation performance changes, so does the economy around it.

A reduction in travel time can increase the effective labour market available to employers. More reliable freight movement can reduce inventory buffers. Better pavement maintenance can reduce vehicle operating costs. A safer intersection can reduce the economic burden of crashes. Better public transport can expand access to employment, education and essential services.

This is why transportation economics traditionally asks questions such as:

What does an intervention cost?
How much travel time does it save?
How does it affect reliability?
Does it reduce crashes?
Does it reduce vehicle operating costs?
Does it improve productivity?
What are its environmental and social effects?
Who receives the benefits?
Who bears the costs?

Artificial intelligence introduces a new economic capability into this system:

the ability to predict, classify, optimise and increasingly automate decisions at a scale and speed that conventional analytical systems often cannot match.

AI can forecast travel demand, detect infrastructure deterioration, optimise routes, coordinate traffic signals, identify safety risks, predict maintenance requirements, improve fleet scheduling and support logistics decisions. But this creates a much harder economic question. The question is not simply:

“How much money can AI save?”

It is:

“How does AI change the allocation of transportation resources—and who captures the economic value created by that change?”

That distinction is fundamental. An AI system may reduce an agency’s operating costs while increasing costs for passengers. It may increase freight efficiency while shifting congestion onto another corridor. It may improve average travel time while reducing accessibility for a small but highly dependent population. It may increase productivity while reducing demand for certain occupations. It may produce enormous analytical capability while creating new expenditures for sensors, computing, cybersecurity, software, data governance and specialist personnel. And it may improve economic efficiency without improving social welfare.

The economics of AI in transportation therefore cannot be reduced to technology adoption. It requires a framework for understanding productivity, costs, distribution, risk, resilience, competition, equity and long-term system effects.

Transportation economics is fundamentally about resource allocation

Traditional transportation economics can be understood as a resource-allocation problem.

Societies have limited:

public budgets;
road space;
transit capacity;
land;
engineering capacity;
energy;
labour;
construction resources;
maintenance resources;
institutional capacity.

The objective is not simply to maximise vehicle speed. A transportation investment can create value through multiple channels.

Major economic benefit channels

Economic channel	Typical transportation effect
Travel time	Less delay and shorter journeys
Reliability	Less uncertainty and reduced buffer time
Vehicle operating costs	Lower fuel, maintenance and operating costs
Safety	Fewer crashes, injuries and fatalities
Freight efficiency	Lower logistics and inventory costs
Infrastructure	Longer asset life and better utilisation
Productivity	Better access to labour, markets and services
Accessibility	Greater access to jobs, education and essential services
Environment	Lower emissions and external costs
Public-sector efficiency	Better allocation of scarce government resources

These categories are already established in transportation benefit-cost analysis.

For example, current USDOT guidance treats benefit-cost analysis as a systematic comparison of expected benefits and costs over the life of an investment. Transportation benefits can include travel-time savings, reliability, safety, operating-cost reductions and other societal effects. Importantly, USDOT distinguishes economic efficiency analysis from distributional analysis: a project can have a positive aggregate benefit-cost result while still creating unequal effects across groups.

AI does not replace this economic framework. It changes how efficiently the system can make and execute decisions within it.

AI is not simply another transportation technology

A new bridge adds physical capacity. A new bus adds service capacity. A new pavement treatment changes physical asset performance. AI is different. AI can influence how existing assets are used. That makes it closer to an economic coordination technology.

Consider a road network with:

1,000 km of roads;
thousands of vehicles;
hundreds of intersections;
multiple modes;
changing weather;
variable demand;
recurring congestion;
unpredictable incidents.

The physical network may remain unchanged while an AI system changes:

signal timing;
route recommendations;
incident detection;
maintenance priorities;
fleet deployment;
demand forecasts;
traffic management;
freight scheduling.

The economic opportunity therefore comes partly from increasing the productivity of existing infrastructure. This is important because transportation systems frequently suffer not only from insufficient physical capacity but also from poor utilization of existing capacity. AI can potentially attack the second problem.

The first major opportunity: reducing information and coordination costs

Transportation systems generate enormous quantities of information. Traffic counts. GPS traces. Transit smart-card records. Weather data. Crash records. Pavement condition data. Vehicle telemetry. Freight movements. Parking occupancy. Travel times. Camera imagery. Roadside sensor data. Maintenance records.

Yet information does not automatically become useful economic knowledge. The challenge is transforming data into decisions. This is where AI can create value.

From Data to Economic Decision-Making

A simplified chain is:

Data → Prediction → Decision → Operational change → Economic outcome

For example:

Traffic observations

↓

AI predicts congestion 30 minutes ahead

↓

Traffic management system adjusts signal timing

↓

Queue formation is reduced

↓

Travel delay and fuel consumption decline

↓

Economic benefit

The International Transport Forum has identified applications including predictive analysis, automated systems, infrastructure monitoring, demand analysis and transport-service optimisation. Its recent guidance for transport authorities emphasises that AI should be selected according to the task and policy objective rather than adopted simply because it is technologically advanced.

This distinction is critical.

AI has economic value only when improved prediction or automation changes a real-world decision sufficiently to produce a valuable outcome.

A highly accurate model that nobody uses has little economic value.

Where AI could create economic value across transportation

The economic opportunity is not confined to autonomous vehicles. It extends across almost the entire transportation ecosystem.

1. Traffic management

AI can support:

traffic-state prediction;
incident detection;
adaptive signal control;
congestion forecasting;
ramp management;
network optimisation;
traveller information.

The economic benefit can arise from reduced delay, improved reliability, lower fuel consumption and better utilization of existing road capacity. But there is an important caveat. Optimizing one intersection does not necessarily optimize the network. Optimizing one corridor does not necessarily optimize the region. And optimizing average travel time does not necessarily maximize social welfare. Transportation is a networked system, meaning local optimisation can produce system-level consequences.

Demand forecasting: better predictions, but not necessarily better decisions

AI can process large and heterogeneous datasets to improve demand forecasting. Traditional models often rely on structured variables such as:

population;
employment;
land use;
income;
vehicle ownership;
trip generation;
network characteristics.

Modern data-driven approaches can supplement these with:

mobile-device data;
GPS traces;
transaction data;
real-time traffic;
weather;
events;
social and economic indicators.

Research by the International Transport Forum highlights both the potential of big data for travel-demand modelling and the associated problems of bias, privacy, commercial sensitivity and data governance. The economic gain is potentially significant. Better demand estimates can improve:

transit scheduling;
fleet allocation;
road investment planning;
freight capacity;
parking management;
infrastructure prioritisation.

But prediction is not the same as explanation. An AI system may accurately predict that demand will fall on a particular transit route without explaining whether the decline is caused by:

fare increases;
poor reliability;
demographic change;
safety concerns;
service quality;
temporary disruption.

That distinction matters because the economically correct response may not be to reduce service. It may be to fix the underlying problem.

Predictive maintenance: perhaps one of AI's strongest transportation-economic applications

Transportation infrastructure is expensive to replace. The economic objective of maintenance is therefore not simply:

“Repair everything.”

It is:

“Intervene at the right time, on the right asset, with the right treatment.”

AI can potentially improve this decision. For roads, railways, bridges and other assets, predictive systems can combine:

inspection records;
deterioration history;
traffic loading;
weather;
material characteristics;
imagery;
sensor data;
maintenance history.

The objective becomes prediction of:

condition → deterioration → failure probability → intervention timing

This can improve lifecycle resource allocation. The OECD/ITF has specifically examined data-driven transport infrastructure maintenance and notes both its potential to make maintenance more effective and the risk that models trained on historical patterns can fail when future conditions differ from the past.

That warning deserves emphasis. A predictive maintenance model is not a crystal ball. Climate conditions can change. Traffic loads can change. Construction practices can change. Materials can change. Maintenance policies can change. Extreme events can occur outside the historical dataset.

Therefore:

Prediction quality must be evaluated under future uncertainty, not merely historical accuracy.

The economic value of AI should be measured through lifecycle economics

A common mistake is to compare:

AI purchase price against estimated annual savings.

That is incomplete. A transportation AI system has a lifecycle. A more realistic economic model is:

NPV = Σₜ₌₀ᵀ (Bₜ − Cₜ)/(1 + r)ᵗ

where:

(B_t) = benefits in year (t);
(C_t) = economic costs in year (t);
(r) = discount rate;
(T) = analysis period.

The benefit-cost ratio can be represented as:

BCR = PV(Benefits) / PV(Costs)

But the cost side must be much broader than software licensing.

AI lifecycle costs

Cost category	Examples
Hardware	Sensors, cameras, edge devices, servers
Connectivity	Network and communications infrastructure
Data	Collection, cleaning, storage and licensing
Computing	Cloud, GPUs, inference and model training
Software	Licensing, development and integration
Integration	Legacy-system integration and APIs
Cybersecurity	Security controls, monitoring and incident response
Personnel	Data scientists, engineers and operators
Training	Workforce training and organisational change
Governance	Auditing, documentation and compliance
Monitoring	Model performance and drift monitoring
Maintenance	Hardware/software updates
Vendor dependence	Switching and migration costs
Decommissioning	Replacement and system retirement

The USDOT’s current benefit-cost guidance similarly stresses full lifecycle project costs rather than narrow acquisition costs.

The economic lesson is straightforward:

An AI system is an infrastructure investment, not merely a software subscription.

The hidden cost: AI requires infrastructure of its own

AI is often presented as a digital technology that can optimise physical infrastructure. But AI itself depends on infrastructure. It needs:

electricity;
computing;
telecommunications;
data centres;
sensors;
data storage;
networks;
cybersecurity;
skilled personnel.

The energy component is increasingly economically relevant.

The International Energy Agency estimates that data centres consumed around 415 TWh of electricity in 2024, approximately 1.5% of global electricity consumption, and projects electricity consumption by data centres to reach roughly 945 TWh by 2030 in its base case.

AI’s physical economic footprint

Indicator	2024	2030 base case
Global data-centre electricity consumption	~415 TWh	~945 TWh
Approximate share of global electricity	~1.5%	just under 3%

Interpretation: AI-enabled transportation does not operate in an economically weightless digital environment. Its benefits must ultimately be considered alongside the infrastructure and energy required to provide the computational capability. This does not mean AI is economically undesirable. It means its benefits should be compared with its full resource requirements.

The productivity paradox: AI cannot pave a road

This is one of the most important limitations. Transportation is a physical system. AI can optimise traffic on a road. It cannot create road capacity out of nothing. AI can predict pavement deterioration. It cannot replace the pavement. AI can optimise bus scheduling. It cannot manufacture buses. AI can predict a bridge’s deterioration. It cannot physically strengthen the bridge. AI can optimise freight routing. It cannot eliminate a missing rail connection or a deficient port. This produces an important distinction:

Digital efficiency versus physical capacity

f(Physical Capacity, Operations, Information, Demand, Institutions)

AI primarily acts on the latter components. If physical infrastructure is the binding constraint, the marginal value of additional algorithmic optimization may eventually decline.

This creates a practical economic principle:

The return on AI is likely to be highest where poor information, coordination or operational inefficiency is the binding constraint—and lower where physical capacity is the dominant constraint.

AI does not eliminate Congestion Economics

One of the most tempting assumptions is that sufficiently intelligent traffic management could “solve congestion.” That is unlikely. Congestion is partly an economic phenomenon. When travel becomes easier, cheaper or more reliable, people and firms may respond by changing:

departure times;
routes;
destinations;
travel frequency;
mode;
vehicle use;
location decisions.

Improved road performance can therefore generate additional demand. This is one reason transportation planners distinguish between operational improvements and broader travel-demand responses. AI may make a network more efficient while simultaneously making travel more attractive. The result can be a rebound effect.

Simplified feedback loop

AI optimisation

↓

Lower travel cost

↓

More attractive travel

↓

Additional demand

↓

Part of the original capacity gain is consumed

This does not mean AI failed. It means transportation is a dynamic economic system rather than a static engineering problem.

Efficiency versus equity: the problem AI cannot optimise away

Now we reach the central economic tension. Suppose an AI model is instructed to minimise the cost of a public transport network.

It may recommend:

fewer buses on low-demand routes;
reduced service during off-peak periods;
consolidation of stops;
dynamic pricing;
fleet redistribution.

From a narrow efficiency perspective, these decisions may be rational. But transportation services generate benefits that are not always visible in passenger counts.

A low-demand route may serve:

older people;
people with disabilities;
low-income households;
students;
shift workers;
people without private vehicles;
people travelling to healthcare;
communities with few alternatives.

The route’s passenger count does not fully measure its social value.

This is a classic distinction between:

private or operational efficiency

and

social welfare.

A mathematically optimal solution is therefore not necessarily a socially optimal solution.

Average efficiency can hide distributional harm

AI should not be judged solely by efficiency. Its economic value should be evaluated against cost, risk, safety, accessibility, reliability and minimum equity requirements.

Consider two AI strategies.

Strategy A

Average travel time decreases by 8%. But the lowest-income users experience a 2% increase.

Strategy B

Average travel time decreases by 5%. But low-income users experience a 4% improvement. A purely aggregate optimisation algorithm may prefer Strategy A. A policymaker concerned with distribution may prefer Strategy B. This is why transportation AI should not be evaluated using a single objective function.

A more realistic objective may be:

$\max W = α E + β A + γ S + δ R - λ C - μ R_{k}$

Where:

W = overall social/economic welfare
E = economic efficiency
A = accessibility
S = safety
R = reliability/resilience
C = lifecycle cost
Rk = AI/system risk
α,β,γ,δ,λ,μ = policy weights

While retaining the equity constraint, the fuller formulation is subject to:

$Eq u i t y \geq Eq u i t y_^{m i n}$

The precise weights should not be hidden inside an algorithm. They are ultimately policy choices. That is an important governance principle.

AI can optimise the objective it is given. It cannot decide whether society chose the right objective.

Accessibility should become an Economic Outcome—not merely a side metric

Traditional transport analysis often focuses heavily on:

vehicle kilometres;
travel time;
speed;
traffic volumes.

But mobility has a deeper economic dimension:

What opportunities can people actually reach?

A person may have a road nearby but still lack effective accessibility because of:

high fares;
unreliable service;
unsafe walking conditions;
disability barriers;
poor connectivity;
inconvenient schedules;
insecurity;
inadequate information.

Therefore AI-enabled transport economics should increasingly evaluate effective accessibility, not merely movement.

A useful conceptual representation is:

A_physical × F_affordability × F_reliability × F_security × F_usability

This matters economically because accessibility affects:

employment opportunities;
education;
healthcare;
labour-market participation;
household expenditure;
social inclusion;
productivity.

An AI system that improves network efficiency while reducing effective accessibility for vulnerable users may therefore create a positive operational result but a negative distributional result.

The data advantage problem: AI may create a new infrastructure divide

AI performance depends heavily on data.

Large metropolitan authorities may possess:

extensive traffic sensors;
connected-vehicle data;
long historical records;
detailed maps;
automated passenger-counting systems;
large engineering departments;
data scientists;
dedicated technology budgets.

Small municipalities may have:

limited sensors;
incomplete crash records;
inconsistent asset inventories;
fragmented databases;
limited computing capacity;
few specialised personnel.

This creates a potentially important economic divide.

The AI adoption chain

Physical infrastructure

→ Sensors & connectivity

→ Data

→ Computing

→ AI capability

→ Operational capacity

→ Economic benefits

A failure anywhere in this chain can reduce the value of the entire system.

The World Bank’s 2026 World Development Report makes a similar point at the economy-wide level: AI benefits depend on complementary infrastructure, skills, data and institutions, and countries lacking these complements risk falling further behind.

This has a direct transportation implication:

The digital divide can become an infrastructure-management divide.

AI concentration may also become a transportation-economics problem

There is another emerging issue. Transportation authorities may increasingly depend on external providers for:

cloud computing;
mapping;
AI models;
mobility platforms;
predictive analytics;
connected-vehicle data;
software infrastructure.

If a small number of firms control critical layers of the technology stack, transport agencies may become dependent on them.

This creates risks involving:

vendor lock-in;
switching costs;
proprietary data;
opaque pricing;
interoperability;
long-term licensing;
model access;
data portability.

Recent OECD research finds that AI markets are evolving unevenly, with concentration around key inputs such as computing, cloud infrastructure, data and foundational models creating potential competitive concerns.

For transportation authorities, this means procurement cannot be reduced to:

“Which AI system has the highest accuracy?”

It must also ask:

Who owns the data?
Can the authority retrieve its data?
Can another vendor use the same dataset?
What happens if the supplier exits the market?
Can the model be independently audited?
Are APIs open?
What are the switching costs?
What happens after the contract expires?

Public procurement becomes an economic instrument

Traditional procurement often asks:

“Does the product meet the technical specification?”

AI procurement needs additional questions. The International Transport Forum’s 2025 guidance specifically highlights the need for transport authorities to update procurement practices for AI, including requirements concerning transparency, risk management, data rights, auditing and explainability.

A transportation authority should therefore evaluate AI procurement across at least five dimensions:

Dimension	Core economic question
Performance	Does the system actually improve outcomes?
Cost	What is the full lifecycle cost?
Risk	What happens when predictions fail?
Competition	Can the authority avoid vendor lock-in?
Public value	Does the system support policy objectives?

This leads to a powerful principle:

The cheapest AI system is not necessarily the lowest-cost option.

A cheap platform that creates vendor lock-in, cybersecurity exposure or expensive migration costs may be substantially more expensive over its lifecycle.

AI and labour: transportation jobs will change, not simply disappear

Transportation employs millions of people across:

driving;
logistics;
maintenance;
engineering;
planning;
operations;
inspection;
administration;
infrastructure construction;
traffic management.

AI can automate or assist some tasks.

Examples include:

image-based inspection;
document processing;
route planning;
dispatching;
demand analysis;
predictive maintenance;
traffic monitoring;
engineering data analysis.

But automation does not automatically equal unemployment. The International Labour Organization’s recent research finds that AI is more likely in many occupations to augment workers and transform tasks than simply eliminate entire occupations.

The important economic question is therefore:

Which tasks disappear, which tasks expand, and who possesses the skills required for the new tasks?

The transportation workforce may become more technically specialised

Future transportation organisations may need more:

data engineers;
AI/ML specialists;
cybersecurity engineers;
systems engineers;
digital-twin specialists;
data governance professionals;
AI assurance specialists;
model validation engineers;
technology procurement specialists.

At the same time, some traditional tasks may decline. This creates a skills-transition problem. The World Economic Forum’s Future of Jobs Report 2025 estimates that broader technological and other macro trends could affect roughly 22% of today’s formal jobs by 2030, with both substantial job creation and displacement expected. It also reports rapidly growing demand for AI, big-data and cybersecurity skills alongside continued importance of human capabilities.

The lesson for transportation agencies is not:

“Replace engineers with AI.”

It is:

“Make engineers capable of working with AI.”

Human engineering judgement remains critical because transportation decisions involve safety, uncertainty, public accountability and physical consequences.

The AI productivity paradox

There is another economic puzzle. An organisation can invest heavily in AI without immediately experiencing proportional productivity gains. Why? Because technology requires complementary changes. An agency may purchase an advanced AI platform but still have:

poor data;
fragmented databases;
outdated IT systems;
weak procurement;
insufficient staff;
poor organisational processes;
no model-governance framework;
inadequate cybersecurity;
no mechanism to act on predictions.

The World Bank’s 2026 research strongly reinforces this complementarity principle: AI adoption produces stronger results where infrastructure, skills, institutions and data systems are already capable of supporting it.

This suggests:

AI Benefit ≠ AI Technology

Rather:

f(AI, Data, Infrastructure, Skills, Institutions, Processes, Governance)

This is one of the most important economic insights in the entire AI debate.

AI can reduce costs—but may also shift them

Consider predictive traffic management. If congestion decreases on a major arterial, that creates benefits. But the algorithm may divert traffic to:

residential streets;
lower-income neighbourhoods;
local roads;
pedestrian-heavy areas.

The system has not eliminated the externality. It has redistributed it. This is a fundamental transportation-economic problem.

Efficiency can be spatially redistributed.

The same principle applies to:

ride-hailing;
freight routing;
parking pricing;
congestion pricing;
transit service allocation;
autonomous vehicle routing.

Therefore, AI evaluation should include spatial distribution of impacts, not merely network-wide averages.

AI can create rebound effects

Suppose AI reduces the cost of freight routing.

That could increase:

vehicle utilisation;
delivery frequency;
e-commerce activity;
network efficiency.

But cheaper logistics can also stimulate additional freight demand. Likewise, if AI makes car travel more reliable, people may be more willing to drive.

Therefore:

An efficiency improvement can increase total system activity.

This is not unique to AI.

It is a classic issue in economics.

But AI can accelerate it because it can make transportation systems more responsive and efficient.

Consequently, policymakers should distinguish between:

Efficiency effect

Less resource required per trip

and

System-level effect

More trips generated because travel becomes cheaper or easier

Both can occur simultaneously.

AI and transportation emissions: potentially positive, potentially ambiguous

AI can help reduce environmental impacts through:

route optimisation;
freight consolidation;
predictive maintenance;
traffic management;
improved vehicle efficiency;
infrastructure optimisation.

OECD analysis identifies digital technologies as capable of improving transport efficiency, including through AI-enabled digital twins, predictive maintenance and freight-routing optimisation. But the outcome is not automatically positive.

AI may also:

increase travel efficiency and therefore induce additional travel;
increase vehicle utilisation;
support faster logistics growth;
consume substantial computing resources.

Therefore the correct question is not:

“Is AI green?”

It is:

“What is the net system-level environmental effect of the AI intervention compared with the alternative?”

Safety is an economic benefit—but AI safety itself has an economic cost

Road crashes generate enormous economic losses through:

fatalities;
injuries;
medical costs;
lost productivity;
property damage;
emergency response;
congestion;
legal and administrative costs.

AI can potentially contribute to safety through:

computer vision;
infrastructure-risk mapping;
conflict detection;
predictive crash analysis;
driver assistance;
infrastructure monitoring.

The ITF has examined AI specifically for proactive road-infrastructure safety management, including the use of computer vision and predictive models to identify high-risk locations before crashes occur. But safety systems introduce their own economic risks.

False positives can produce:

unnecessary interventions;
wasted enforcement resources;
excessive maintenance;
operational disruption.

False negatives can be much more serious.

A system that incorrectly identifies a dangerous location as safe can create catastrophic consequences.

Thus:

The economic value of AI safety systems must account for both expected benefits and failure consequences.

AI failure is an economic event

Traditional transportation infrastructure fails physically. AI systems can fail informationally.

Examples include:

inaccurate predictions;
data drift;
sensor failure;
cyberattacks;
model degradation;
unexpected conditions;
adversarial inputs;
software bugs;
poor transferability.

The OECD/ITF has explicitly warned that data-driven infrastructure systems can become unreliable when future conditions differ from the historical data used to train them.

This creates a new category of transportation risk:

model risk.

Model risk should be treated similarly to other engineering risks.

A system should have:

performance thresholds;
fallback modes;
monitoring;
redundancy;
human oversight;
incident response;
periodic validation;
documented limitations.

NIST’s AI Risk Management Framework emphasises characteristics such as validity, reliability, safety, security, resilience, accountability, transparency, explainability, privacy and fairness as components of trustworthy AI.

AI needs an economic value-of-risk calculation

A sophisticated transportation AI evaluation should therefore consider Expected Benefits Expected Costs and Expected Risk Losses.

where:

Σᵢ P(Fᵢ) × C(Fᵢ)

with:

(P(F_i)) = probability of failure scenario (i);
(C(F_i)) = consequence of that failure.

For high-consequence transportation applications, this can fundamentally change the investment decision. An AI system with slightly lower expected efficiency but significantly lower failure risk may be economically preferable.

Accuracy is not the same as economic value

This deserves its own section.

Suppose:

AI Model A: 95% prediction accuracy

AI Model B: 90% prediction accuracy

It is tempting to choose Model A.

But suppose:

Model A costs $10 million;
Model B costs $2 million;
both lead to nearly identical operational decisions.

Model B may have greater economic value.

Conversely, suppose:

Model A is 90% accurate;
Model B is 85% accurate;
Model A prevents high-cost failures that Model B misses.

Model A could be economically superior by a large margin.

Therefore:

The correct metric is not model accuracy. It is value generated per unit of total lifecycle cost and risk.

AI should be compared with non-AI alternatives

This is where many technology projects go wrong. An agency identifies a problem and asks:

“How can AI solve it?”

The economically correct question is:

“What is the least-cost intervention capable of achieving the desired outcome?”

The alternatives might include:

conventional statistical modelling;
additional staff;
better inspection procedures;
improved signal timing;
better asset management;
conventional sensors;
infrastructure investment;
policy changes;
pricing;
operational reform.

AI should compete against these alternatives. A sophisticated algorithm should not automatically win simply because it is sophisticated. The ITF’s latest guidance explicitly notes that sometimes not using AI is the right choice. That is an important statement.

The AI counterfactual

Every transportation AI project should establish a counterfactual:

What would happen if we did not deploy AI?

Without a counterfactual, it is difficult to determine whether benefits are actually attributable to AI. A proper evaluation should compare:

Outcome_No AI

not simply:

Outcome_AI

This sounds obvious, but it is frequently overlooked in technology deployment. For example, if pavement failures decline after an AI maintenance system is introduced, the decline might actually be caused by:

lower traffic;
better construction;
higher maintenance budgets;
favourable weather.

The causal contribution of AI must be isolated.

From pilot projects to measurable economic evidence

Transportation authorities often begin AI deployment through pilots. That is sensible. But pilots can create a dangerous illusion. A successful demonstration is not necessarily a successful investment.

A pilot may work because:

experts provide exceptional oversight;
data scientists manually clean data;
vendor personnel are present;
the deployment area is unusually controlled;
funding is temporarily abundant.

Scaling may expose:

higher maintenance costs;
data-quality problems;
workforce shortages;
cybersecurity issues;
interoperability failures;
vendor dependence.

Therefore AI deployment should follow an evidence ladder:

Stage 1 — Feasibility

Can it work?

Stage 2 — Technical validation

Does it perform reliably?

Stage 3 — Operational validation

Can the organisation actually use it?

Stage 4 — Economic evaluation

Do benefits exceed lifecycle costs?

Stage 5 — Distributional evaluation

Who gains and who loses?

Stage 6 — Scale evaluation

Does the business case survive expansion?

This approach aligns closely with ITF’s recommendation to deploy cautiously, assess risks, monitor performance and continuously adjust AI systems.

AI and small versus large transportation agencies

One of the greatest policy challenges is scale.

Large agencies can afford:

dedicated AI teams;
high-quality data systems;
advanced sensors;
cloud infrastructure;
procurement expertise;
cybersecurity teams.

Small agencies may struggle to justify the fixed costs. This creates economies of scale. A large authority can spread AI infrastructure costs over millions of trips. A small authority may not. Possible solutions include:

shared regional platforms;
national data platforms;
open standards;
pooled procurement;
shared computing;
common model libraries;
technical assistance;
interoperable systems.

The objective should not be to force every agency to build its own AI stack. It should be to prevent technological capability from becoming concentrated only in wealthy jurisdictions.

Developing economies face a different AI opportunity

For many developing countries, the most economically attractive transportation AI applications may not require frontier AI. They may involve relatively inexpensive systems for:

road-condition monitoring;
traffic counting;
public transport planning;
route optimisation;
crash-risk mapping;
maintenance prioritisation;
logistics coordination.

The World Bank’s 2026 World Development Report argues that developing economies can often benefit more from adopting and adapting existing AI tools than attempting to build frontier AI infrastructure themselves. It identifies electricity, connectivity, skills, data and institutions as critical complements. This is highly relevant to transportation. A country does not need a trillion-dollar AI ecosystem to obtain economic value from AI-assisted road management.

It needs:

the right problem, good enough data, adequate infrastructure and institutional capacity to act on the result.

But AI may also widen international inequality

This creates a paradox. AI can lower the cost of expertise. That could help countries with shortages of:

engineers;
planners;
inspectors;
analysts.

But advanced economies may simultaneously adopt AI faster because they already possess:

better infrastructure;
more data;
stronger institutions;
greater computing access;
more skilled workers;
larger investment budgets.

The IMF’s 2025 analysis finds that differences in AI exposure and preparedness could cause AI’s growth effects to be substantially larger in advanced economies than in low-income countries, potentially widening cross-country income differences. Transportation could reproduce this pattern. The future may therefore contain:

AI-rich transportation systems

versus

AI-poor transportation systems.

That is not merely a technology gap. It can become a productivity gap.

AI and transportation economics: a new distribution question

This brings us back to the central question. Suppose AI creates $100 million in annual transportation-system benefits. Who receives them? Potential beneficiaries include:

road users;
transit passengers;
freight operators;
logistics companies;
vehicle manufacturers;
technology vendors;
employers;
landowners;
governments;
taxpayers.

But the distribution may be highly unequal.

For example:

AI improves freight routing

Freight companies save money. Consumers may eventually receive some of the savings through lower prices. Shareholders may capture some of the value. Employees may benefit if productivity increases wages. But the distribution is not automatic.

This is why:

Economic efficiency and distributional incidence are separate questions.

Who captures the AI dividend?

A useful conceptual framework is:

Productivity Gain + Resource Savings + Risk Reduction

But the distribution of that dividend depends on:

f(Market Power, Ownership, Labour Institutions, Pricing, Regulation, Public Policy)

This is why the economic impact of AI cannot be determined by engineering performance alone. Two cities could deploy the same AI technology and produce very different social outcomes because their:

governance;
labour markets;
procurement systems;
pricing structures;
public-service objectives

are different.

AI can change the economics of infrastructure investment

One particularly interesting future possibility is that AI could alter the relative attractiveness of:

new physical capacity

versus

better management of existing capacity.

Suppose a city faces a congested corridor.

Traditional thinking might suggest:

Build another lane.

But an AI-enabled alternative could include:

adaptive signals;
dynamic lane management;
transit priority;
demand forecasting;
incident prediction;
freight scheduling;
parking management;
dynamic pricing.

If these interventions deliver similar outcomes at lower lifecycle cost, AI could change the investment hierarchy. But this should not become an excuse to avoid infrastructure investment when physical capacity is genuinely inadequate.

The right economic question is comparative:

BCR (AI Operations) vs. BCR (Physical Expansion)

The winning intervention should depend on the net social benefits, not technological fashion.

AI could make transportation investment more dynamic

Traditional infrastructure planning often evaluates projects at specific points in time. AI could support continuous prioritization. Imagine a road authority with 10,000 km of network. Instead of relying primarily on periodic inspections and fixed maintenance schedules, it could continuously update:

deterioration probabilities;
traffic exposure;
crash risk;
climate exposure;
maintenance costs;
expected service life.

The investment programme could therefore evolve continuously.

This moves transportation asset management toward:

dynamic portfolio optimisation.

The road network becomes an investment portfolio in which each asset has:

expected deterioration;
risk;
replacement cost;
user impact;
expected return from intervention.

AI can potentially improve the ranking of these interventions.

But optimisation can become dangerous when objectives are incomplete

Imagine an algorithm instructed to:

“Minimise transportation operating costs.”

It may find a technically excellent solution.

But what if the objective excludes:

accessibility;
safety;
equity;
environmental effects;
resilience;
privacy?

The algorithm may optimise the wrong system.

This creates what might be called the:

Objective-function problem

AI does not eliminate value judgments. It hides them inside optimization criteria. Therefore transportation engineers and policymakers must determine:

What should be optimised?

before asking:

How should it be optimised?

A state-of-the-art Transportation AI Economic Framework

A mature evaluation should examine at least eight dimensions.

Dimension	Key question
Efficiency	Does AI improve system performance?
Cost	Are lifecycle benefits greater than lifecycle costs?
Equity	Who gains and who loses?
Safety	What happens under model failure?
Resilience	Does the system remain useful under abnormal conditions?
Competition	Does deployment create vendor dependence?
Labour	How are tasks, skills and jobs affected?
Sustainability	What is the net environmental/resource effect?

This produces a broader objective:

Lifecycle Cost Risk

subject to:

Equity ≥ Minimum Acceptable Threshold

This is closer to the real problem transportation authorities face.

The economic lifecycle of transportation AI

                 ┌──────────────────────┐
                 │   TRANSPORT PROBLEM  │
                 └──────────┬───────────┘
                            ↓
                 ┌──────────────────────┐
                 │  BASELINE / COUNTER- │
                 │       FACTUAL        │
                 └──────────┬───────────┘
                            ↓
                 ┌──────────────────────┐
                 │   AI ALTERNATIVE     │
                 └──────────┬───────────┘
                            ↓
        ┌───────────────────┼───────────────────┐
        ↓                   ↓                   ↓
   Performance          Lifecycle Cost       Risk
        ↓                   ↓                   ↓
   Time / Safety       Hardware / Data      Failure /
   Reliability         Software / Skills    Cybersecurity
        └───────────────────┼───────────────────┘
                            ↓
                 ┌──────────────────────┐
                 │ NET ECONOMIC BENEFIT │
                 └──────────┬───────────┘
                            ↓
                 ┌──────────────────────┐
                 │ DISTRIBUTIONAL TEST  │
                 │ Who gains? Who pays? │
                 └──────────┬───────────┘
                            ↓
                 ┌──────────────────────┐
                 │ SCALE / DON'T SCALE  │
                 └──────────────────────┘

The key point is that AI deployment is the middle of the economic analysis, not the end of it.

A practical economic scorecard for transportation AI

Before approving an AI project, a transportation authority could ask:

Economic

What is the measurable baseline?
What is the counterfactual?
What benefits are expected?
What costs are expected?
What is the NPV?
What is the BCR?
How sensitive are results to assumptions?

Operational

What decision will AI change?
How frequently?
Who acts on the output?
What happens when the system is unavailable?

Data

Who owns the data?
How representative are the datasets?
What biases exist?
How often will data be refreshed?

Safety

What are the failure modes?
What is the worst credible consequence?
Is human oversight required?

Equity

Which groups benefit?
Which groups may lose?
Are impacts spatially concentrated?

Labour

Which tasks will change?
What training is required?
Are workers being augmented or displaced?

Competition

Is the solution interoperable?
Can the authority switch suppliers?
Are APIs and data portable?

Sustainability

What computing resources are required?
What physical infrastructure is required?
What is the net energy/environmental effect?

The most important shift: from AI adoption to AI governance

The next stage of transportation AI should not be defined by:

“How much AI can we deploy?”

It should be defined by:

“Where does AI create demonstrable public value?”

The ITF’s 2025 guidance is particularly important here. It recommends a proportionate approach, safe-to-fail deployment, risk assessment, continuous monitoring and recognition that in some circumstances the correct decision is not to use AI. That is a much more mature position than technology-first thinking.

Transportation agencies should therefore move toward:

problem-first → evidence-first → economics-first → AI where justified.

Not:

AI-first → find a problem for it to solve.

The future: AI may become transportation's decision layer

The long-term significance of AI may not be a single application such as autonomous vehicles. Its deeper impact could be the creation of a continuous decision layer across the transportation system.

Imagine a future network in which AI continuously estimates:

traffic state;
travel demand;
infrastructure condition;
crash risk;
transit demand;
freight demand;
energy consumption;
weather risk;
maintenance needs;
accessibility;
network reliability.

The system could continuously update investment and operational recommendations.

This would transform transportation management from:

periodic planning

toward:

continuous evidence-based adaptation.

But such a system would also increase the importance of governance.

The more decisions depend on algorithmic systems, the more important it becomes to understand:

who controls them;
what objectives they optimise;
what data they use;
how they fail;
how decisions can be challenged;
who is accountable.

The deeper economic question: Efficiency for Whom?

This is ultimately the question that conventional engineering optimisation can overlook.

Suppose AI makes a transportation network:

10% faster;
8% cheaper to operate;
5% more reliable.

Those are impressive results.

But suppose simultaneously:

low-income users face higher fares;
rural routes are reduced;
certain neighbourhoods experience more traffic;
transport workers lose bargaining power;
the agency becomes dependent on a single technology vendor.

Is the system economically successful? There is no purely technical answer. It depends on what society means by economic success.

If success means only:

maximum output per unit of input,

the answer might be yes.

If success means:

improved social welfare, accessibility, safety, productivity and fairness within sustainable fiscal and environmental limits,

the answer becomes much more complicated.

The new Transportation-Economic principle

The traditional transportation engineering mindset often asks:

Can we make the system work better?

AI introduces a more powerful question:

Can we make the system allocate scarce resources better?

Transportation economics adds another:

Can we make that allocation improve overall welfare?

And equity adds the final challenge:

Who receives the improvement?

These questions should be considered together.

Conclusion: The AI Dividend Is Not Automatically a Public Dividend

AI could become one of the most important productivity technologies in transportation since the development of modern traffic management, digital mapping and large-scale logistics systems. Its potential is substantial.

AI can help transportation organisations:

predict demand;
optimise networks;
improve fleet utilisation;
identify infrastructure deterioration;
prioritise maintenance;
improve logistics;
detect safety risks;
allocate resources;
improve reliability;
support better planning.

But none of these benefits are automatic.

AI also introduces:

lifecycle costs;
energy consumption;
cybersecurity risks;
model uncertainty;
data dependency;
vendor concentration;
workforce disruption;
privacy concerns;
institutional requirements;
equity challenges.

And perhaps most importantly, AI can make an inefficient transportation system more efficiently pursue the wrong objective.

That is why the economic evaluation of transportation AI must go beyond:

“How much money does it save?”

A complete assessment should ask:

What does it cost?

What does it save?

What alternative could produce the same result?

What physical constraints remain?

What happens when demand responds?

Who benefits?

Who pays?

Who owns the data?

Who controls the algorithm?

What happens when the model is wrong?

What happens to workers?

Does the system improve accessibility?

Does it improve safety?

Does it increase resilience?

Does it create vendor dependence?

Does it reduce or merely redistribute externalities?

And, ultimately, does it increase social welfare?

The greatest economic opportunity presented by AI may therefore not be automation itself.

It may be better allocation of scarce transportation resources.

But the greatest economic danger is the assumption that efficiency automatically equals welfare. It does not. An AI system can make a transportation network faster without making it fairer. It can make logistics cheaper without making the economy more resilient. It can make an agency more productive without making its decisions better. It can reduce average travel time while worsening accessibility for the people who depend on transport the most. And it can generate enormous economic value without guaranteeing that the people who create or need that value actually receive it. The future of AI in transportation economics will therefore be determined not only by the intelligence of algorithms, but by the quality of the economic objectives, institutions, governance and engineering judgement surrounding them.

The real question is not whether AI can make transportation more efficient. It can. The harder question is whether we can make that efficiency economically productive, socially equitable, physically meaningful and resilient enough to deserve the name “progress.”

Key evidence at a glance

Research finding	Why it matters for transportation economics
AI can support predictive analysis and automated transport functions	Creates opportunities for operational and resource-allocation efficiency
Data-driven infrastructure maintenance can improve targeting but is vulnerable to changes outside historical data	AI benefits depend on model robustness and future uncertainty
Big data can improve travel-demand modelling but introduces bias, privacy and governance issues	Better information is valuable only when its limitations are understood
EU AI adoption in transport was reported at 8% in 2024 versus 13% across the EU economy	Transportation AI adoption remains relatively early and uneven
Data centres consumed about 415 TWh in 2024 and could reach about 945 TWh by 2030 in the IEA base case	AI has a real physical and energy cost
AI benefits depend on infrastructure, data, skills and institutions	Technology adoption without complementary capacity can produce weak returns
AI can widen productivity differences between economies with different levels of preparedness	Transportation technology could reinforce geographic inequality
AI markets show potential concentration around computing, cloud, data and foundation models	Transport authorities need interoperability and procurement safeguards
ILO research indicates augmentation and transformation are often more likely than complete occupational replacement	Transportation workforce policy should focus on task transformation and reskilling
USDOT distinguishes benefit-cost analysis from distributional analysis	A project can be economically efficient while producing unequal impacts

Final takeaway

AI may become a powerful transportation productivity technology—but productivity is not the same thing as welfare.

The next generation of transportation economics therefore needs to evaluate AI across efficiency + lifecycle cost + risk + accessibility + equity + labour + competition + resilience + sustainability.

The winners will not necessarily be the transportation systems with the most AI. They will be the systems that use the right AI, for the right problem, at the right scale, with the right economic objective—and can demonstrate that the resulting benefits are actually greater than the costs and risks.

References & Further Reading

OECD/ITF (2025). AI for Transport Authorities: Principles and Practical Guidance. International Transport Forum, OECD.
Official publication — OECD/ITF
OECD/ITF (2021). Data-driven Transport Infrastructure Maintenance. International Transport Forum Policy Papers, No. 95, OECD Publishing.
Official publication — OECD
OECD/ITF (2021). Artificial Intelligence in Proactive Road Infrastructure Safety Management: Summary and Conclusions. ITF Roundtable Reports, No. 187.
Official publication — OECD/ITF
OECD/ITF (2021). Big Data for Travel Demand Modelling: Summary and Conclusions. ITF Roundtable Reports, No. 186.
Official publication — OECD/ITF
OECD/ITF (2019). Governing Transport in the Algorithmic Age. International Transport Forum Policy Papers, No. 82.
Official publication — OECD/ITF
World Bank (2026). World Development Report 2026: The Promise of Artificial Intelligence. World Bank.
Official World Bank report
Cerutti, E. M., Garcia Pascual, A., Kido, Y., Li, L., Melina, G., Mendes Tavares, M., & Wingender, P. (2025). The Global Impact of AI: Mind the Gap. IMF Working Paper 25/76.
Official IMF publication
Rockall, E. J., Mendes Tavares, M., & Pizzinelli, C. (2025). AI Adoption and Inequality. IMF Working Paper 25/68.
Official IMF publication
International Labour Organization (2025). Artificial Intelligence Adoption and its Impact on Jobs. International Labour Organization.
Official ILO publication
International Labour Organization (2025). Generative AI and Jobs: A Refined Global Index of Occupational Exposure. ILO Working Paper 140.
Official ILO publication
World Economic Forum (2025). The Future of Jobs Report 2025. World Economic Forum.
Official World Economic Forum report
U.S. Department of Transportation (2025/2026). Benefit-Cost Analysis Guidance for Discretionary Grant Programs. U.S. Department of Transportation.
Official USDOT guidance
Federal Highway Administration (FHWA). Transportation Systems Management and Operations Benefit-Cost Analysis Compendium. U.S. Department of Transportation.
Official FHWA resource
Tabassi, E. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1. National Institute of Standards and Technology.
Official NIST publication
Autio, C., Schwartz, R., Dunietz, J., Jain, S., Stanley, M., Tabassi, E., Hall, P., & Roberts, K. (2024). Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile. NIST AI 600-1.
Official NIST publication
International Energy Agency (2025). Energy and AI. IEA, Paris.
Official IEA report
OECD (2026). Competition in the Age of AI: Initial Evidence from Microdata. OECD Artificial Intelligence Papers, No. 64.
Official OECD publication

AI and the Economics of Transportation: Who Wins, Who Pays, and Who Gets Left Behind?

Table of Contents

Quick Links

Introduction: The economic question is bigger than “Will AI save money?”

Transportation economics is fundamentally about resource allocation

Major economic benefit channels

AI is not simply another transportation technology

The first major opportunity: reducing information and coordination costs

From Data to Economic Decision-Making

Where AI could create economic value across transportation

1. Traffic management

Demand forecasting: better predictions, but not necessarily better decisions

Predictive maintenance: perhaps one of AI's strongest transportation-economic applications

The economic value of AI should be measured through lifecycle economics

AI lifecycle costs

The hidden cost: AI requires infrastructure of its own

AI’s physical economic footprint

The productivity paradox: AI cannot pave a road

Digital efficiency versus physical capacity

AI does not eliminate Congestion Economics

Simplified feedback loop

Efficiency versus equity: the problem AI cannot optimise away

Average efficiency can hide distributional harm

Strategy A

Strategy B

Accessibility should become an Economic Outcome—not merely a side metric

The data advantage problem: AI may create a new infrastructure divide

The AI adoption chain

AI concentration may also become a transportation-economics problem

Public procurement becomes an economic instrument

AI and labour: transportation jobs will change, not simply disappear

The transportation workforce may become more technically specialised

The AI productivity paradox

AI can reduce costs—but may also shift them

Efficiency can be spatially redistributed.

AI can create rebound effects

Efficiency effect

System-level effect

AI and transportation emissions: potentially positive, potentially ambiguous

Safety is an economic benefit—but AI safety itself has an economic cost

AI failure is an economic event

AI needs an economic value-of-risk calculation

Accuracy is not the same as economic value

AI should be compared with non-AI alternatives

The AI counterfactual

From pilot projects to measurable economic evidence

Stage 1 — Feasibility

Stage 2 — Technical validation

Stage 3 — Operational validation

Stage 4 — Economic evaluation

Stage 5 — Distributional evaluation

Stage 6 — Scale evaluation

AI and small versus large transportation agencies

Developing economies face a different AI opportunity

But AI may also widen international inequality

AI and transportation economics: a new distribution question

AI improves freight routing

Who captures the AI dividend?

AI can change the economics of infrastructure investment

AI could make transportation investment more dynamic

But optimisation can become dangerous when objectives are incomplete

Objective-function problem

A state-of-the-art Transportation AI Economic Framework

The economic lifecycle of transportation AI

A practical economic scorecard for transportation AI

Economic

Operational

Data

Safety

Equity

Labour

Competition

Sustainability

The most important shift: from AI adoption to AI governance

The future: AI may become transportation's decision layer

The deeper economic question: Efficiency for Whom?

The new Transportation-Economic principle

Conclusion: The AI Dividend Is Not Automatically a Public Dividend

What does it cost?

What does it save?