AI in public transit operations

Staffing. The American Public Transportation Association's workforce study, published in October 2022, reported that 96% of agencies it surveyed were experiencing a workforce shortage, and 84% of those said the shortage affected their ability to provide service. A shortage turns every morning into an exception-management exercise: open work, extraboard assignments, canceled trips and the communication that follows.

Disruptions. Vehicle failures, detours, weather and incidents change service in real time. The information has to reach dispatch, operators, customers and the record, and each of those hand-offs is usually manual.

Paratransit. Under 49 CFR 37.131, an agency must schedule ADA paratransit trips requested the previous day, and may not limit service through operational patterns such as substantial numbers of significantly untimely pickups, trip denials or missed trips. Reservations, confirmations, same-day changes and no-shows generate a large volume of customer contact and scheduling work.

Customer contact. Questions about arrivals, detours, lost property, fares and complaints. Many answers already exist in vehicle-location and schedule data that the customer cannot see.

Records. Incident reports, safety events, maintenance records and reporting to funders all depend on staff rekeying information from one system or form into another.

Customer contact is the usual first choice. Answering a detour question with live data is call deflection that resolves, and the complaints that need a person reach one faster.

Safety. Agencies covered by 49 CFR Part 673 run a Safety Management System with four components: safety management policy, safety risk management, safety assurance and safety promotion. An AI system that changes how operations work belongs inside that system's hazard identification and risk assessment, and it should never sit on the critical path of a safety decision without a person.

Labor. Under 49 U.S.C. 5333(b), commonly called section 13(c), recipients of federal transit funds must have protective arrangements in place for affected employees, and the Department of Labor must confirm those arrangements before the Federal Transit Administration releases funds. Those arrangements include preserving collective bargaining rights and protecting employees against a worsening of their positions. Part 673 brings labor into safety directly as well: it defines the Safety Committee as "the formal joint labor-management committee on issues related to safety that is required by 49 U.S.C. 5329 and this part." Work assignment rules also live in the labor agreement. An AI project that changes how represented employees work needs labor at the table from the start.

Accessibility. A paratransit error is a compliance exposure as well as a service failure. Scheduling suggestions stay suggestions until a scheduler accepts them.

Public accountability. A public agency has to explain its decisions. Every action the AI takes needs a log a supervisor, an auditor or a board member can read.

The working design is management by exception: AI handles the routine and surfaces the deviations, with human in the loop built into each queue and guardrails that stop the system at its limits.

Start outside the safety-critical path, in one mode or one division. Customer inquiries, complaint triage or incident report intake all work. Bring labor representatives in before design, not after. Set baselines first: response time, repeat contacts, report completeness, time from incident to coded record, and hours supervisors spend on administration. Run long enough to include a service change or a bad-weather week, because transit operations are judged on their worst days.

In Manthle's model, the engineers build around the agency's existing scheduling, vehicle-location and customer systems and operate the result with agency staff. The broader method, including how a working pilot becomes an ongoing operation, is set out in the guide to AI for business operations, and the queue design follows exception management.