Artificial intelligence (AI) can do tremendous work to solve complex problems in transit, from making sense of data to automating workflows. But from our experience applying AI across projects at Foursquare ITP, we’ve also seen the challenges agencies can face in effectively leveraging it.
In recent work for the American Public Transportation Association (APTA) on AI and machine learning (ML) applications in transit, we found that agencies are just starting to leverage AI for tasks such as transit lane enforcement, customer service, and maintenance. Through that research and other hands-on work, we’ve identified three foundational areas agencies should prioritize as they begin integrating AI into their operations:
- Building a source of truth through creating an organizational context for AI.
- Picking methods fit for purpose with machine intelligence beyond AI.
- Governing AI to understand organizational readiness.
A Source of Truth
In working with an AI chatbot, you may have done a bit of “prompt engineering” by telling the chatbot, “I’m a transit planning expert…” before your query. In the past year, however, the conversation has shifted to “context engineering,” or how to provide AI with the appropriate information about your problem. AI tools that go beyond ‘advanced search engines’ to those that have an agent working with you on important workflows need organizational context and access to data. But for many agencies, that information can be disjointed across different systems, spreadsheets, and loose documents. Planning leaders who want AI to answer questions like “Which three routes have the worst on-time performance this pick, and why?” will find that AI tools don’t immediately have the answers at hand, and even if they are provided, they may not have enough organizational context to handle data correctly. As a result, your on-time performance standards may not be applied correctly, or two conflicting vendor sources may tell different stories.
Open data standards like Transit Integrated Data Exchange Specification (TIDES) are the secret to unlocking AI’s effectiveness for transit agencies. TIDES provides a unified view of operational data, including vehicle locations, fare payments, and passenger activity. Similar to how the general transit feed specification (GTFS) enabled seamless transit routing for customers, TIDES can help agencies move quickly from data to actionable insights. With a defined and well-documented schema, agencies with data in TIDES format can more readily prompt AI tools for answers to analytical questions—skipping a dashboard, having the AI write analytical code, and moving straight to the answer. It’s a model of transit analytics that we’re excited to incorporate into our own tools, with more to come later this year.
For agencies that see AI adoption ahead, now is the time to start considering:
- Where can we put requirements for data in open standards for upcoming RFPs?
- Where does our agency have trouble resolving conflicting systems or vendor data?
- How can we begin integrating the data sources AI will need to retrieve context for answers?
Methods Fit for Purpose
While AI chatbots can be effective research assistants on their own, they cannot produce ridership forecasts, transit schedules, or other key analytical outputs that agencies need. To these ends, technical methods such as optimization and ML remain the best fit.
While some uses of AI can yield difficult-to-measure productivity benefits, ML and optimization can lead to practical outcomes. We’ve seen this in several of our recent and ongoing projects:
- Facility optimizations can identify where garage locations and vehicles should be best allocated. In our work with the Jacksonville Transportation Authority (JTA) and Maryland Transit Administration, this could save substantial deadheading costs.
- Our Route Optimization tool for high-demand corridors uses optimization to group origin-destination flows to reveal where routes can generate high ridership.
- Applying schedule optimization methods at the planning stage can provide a clearer sense of the operating costs of new services.
- Using custom ML models can help identify where to expand service based on an agency’s ridership data.
Key Terms
Generative AI (Gen AI) and Large Language Models: AI that creates content (text, code, images, video) based on an input.
Machine Learning (ML): AI that approximates cognitive tasks like clustering, forecasting, or classification.
Optimization: Methods that provide a solution to a problem given a set of constraints.
While AI chatbots can help someone build, refine, steer, or understand these methods, optimization and ML are tools in their own right.
When considering where to apply AI, broadly speaking, it’s worthwhile for agencies to understand:
- What are the analytical problems we need AI to solve?
- What are the problems we know optimal solutions are out there for, but haven’t yet had the time to solve?
- What are tasks where human cognition is useful but in short supply?
Governing AI
The methods above—ML, optimization, and large language models—come with a challenge familiar to any analytical endeavor: interpretability and governance. As a firm with transit planning at its core, Foursquare ITP frequently has our analyses evaluated by members of the public, stakeholders, and board members. A “just trust us” attitude doesn’t fly when any decisions made will impact transit service, major investments, or a person’s familiar bus route. In similar ways, agencies that use AI to automate workflows or provide research assistance will need to defend conclusions derived from AI.
We’ve learned that it’s valuable to have analyses that are readily interpretable; ones where you can tell a clear story of how inputs led to our recommendations. For example, a model that predicts that an area with more zero-car households needs more transit is easy to explain and plausibly true; however, a complex interaction among many factors driving transit demand can be harder to defend, even if well supported by the data.
Agencies that want to automate processes using AI or reach analytical conclusions will need to be able to peek under the hood to understand what data and context were consumed and how insights were generated.
This points to the underlying need for agencies to have:
- Analytical staff with strong skill foundations.
- IT departments that can manage audit trails of AI usage.
- AI workflows that have been scaffolded to return repeatable results where needed.
Looking Ahead
We’re excited to help build the future where AI agents can:
- Answer data questions in a trustworthy fashion;
- Leverage the right tools for the task; and
- Remain accountable and comprehensive to agency staff and members of the public.
There are important questions for agencies to answer as they determine how to responsibly and effectively integrate AI into transit planning. For agencies looking to learn more, we encourage you to read the recent Artificial Intelligence (AI) and Machine Learning (ML) in Public Transit Primer we authored with APTA and EBP.