A bus route can look healthy on paper while struggling in the streets.
In Salvador, Brazil, researchers found that broad measures of passenger demand could obscure differences from one part of a route to another. Their agentic artificial intelligence system, SUNTInsight, was designed to let transit managers drill into those patterns using ordinary language, turning a sprawling transportation dataset into more targeted operational decisions.
Public transportation agencies have no shortage of data. In Salvador, researchers had access to a dataset covering about 700,000 passengers, approximately 2,000 vehicles, nearly 400 lines and almost 3,000 stops and stations, according to a research article published in May.
The challenge was turning that volume of information into decisions that transit managers could use. Transportation agencies often rely on static dashboards and manual queries to examine their operations, creating a gap between the data they collect and the insights they can act on. The researchers behind SUNTInsight set out to narrow that gap with an agentic AI system that could let managers explore the data using natural language.
Dissecting this data was important because broad averages can conceal problems occurring in specific parts of a transportation network. A route may appear to have manageable demand overall while particular stops or segments experience pressure at certain times. Finding these patterns traditionally requires managers to know what data to retrieve and how to query it.
Developing Agentic AI for Public Transportation
We’d love to be your preferred source for news.
Please add us to your preferred sources list so our news, data and interviews show up in your feed. Thanks!
SUNTInsight was designed to put that analysis behind a conversational interface. A manager can enter a request in ordinary language, and the system determines whether it needs to access the underlying database. For data requests, the AI model generates structured query language (SQL), retrieves the relevant information and sends the results back for analysis. The system then displays the query and results alongside visualizations, allowing the manager to inspect what the system did, according to the research article.
To test this system, researchers examined a bus route during a morning rush period. A request for data on March 5, 2026, for Line 1637 between 7 a.m. and 8 a.m. produced 428 records from 103 stops. The resulting visualization showed that occupancy varied along the route, with heavier demand concentrated in particular neighborhoods rather than distributed evenly across the entire line.
That finding changed the operational response suggested by the original dataset. Adding buses across the entire route could address some of the pressure while adding capacity in areas where it was not needed. The AI-powered analysis instead pointed toward more targeted measures, including short turns and segment-specific headway control on high-demand sections.
Working Out the Kinks
SUNTInsight was not without its teething issues in development, however. Several smaller language models struggled with tool calling and sometimes generated responses as though they had accessed a database when they had not, the research article said. Larger reasoning-focused models performed more consistently.
Security was a concern during development as well. Because SUNTInsight executes AI model-generated SQL and code, the researchers restricted the application’s database permissions, isolated generated code, and logged prompts, queries, outputs, errors and performance data. These controls were intended to address both accidental model errors and attempts to manipulate the system through malicious prompts.
The result is a system designed to bring data retrieval, analysis and visualization into a single workflow while leaving decisions with human operators. In the Salvador case, that approach helped surface localized demand patterns that could be obscured by route-level averages. SUNTInsight now serves as a practical co-pilot for transit management, illustrating how agentic AI can connect operational data with more targeted decision-making while maintaining human oversight.
For all PYMNTS AI coverage, subscribe to the daily AI newsletter.
Add as Preferred Source
