Amtrak: Using an AI Assistant to Scale Self-Service
Amtrak used Julie across phone and web entry points to handle recurring travel questions before customer-service teams took on the cases that needed human judgment.
Passengers ask schedule or ticket questions
Passengers ask about schedules, fares, train status, reservations, or booking.
Julie identifies the request
Julie interprets a spoken or written request and identifies the service task.
AI answers or supports the task
The assistant provides a concise answer or directs the passenger into the relevant booking flow.
Complex issues go to support staff
Requests outside the defined flow continue to customer-service teams.
Before: Passenger inquiry → Human answer → Humans handle booking and exceptions After: Passenger inquiry → Julie interprets → Self-service or escalation → Humans handle complex issues
- Start with high-volume questions that have reliable answers.
- Treat booking and service data as part of the workflow, not an afterthought.
- Keep a measurable path from automation to human support.
View process and full evidence
Nine sections covering context, ICP, positioning, channels, results, replicability, and next steps
Overview
Amtrak introduced its voice application, Julie, to manage enquiries about schedules, fares, train status, reservations, and bookings. A U.S. Department of Transportation briefing says the initiative began around 2000, when Amtrak was receiving more than 84,000 passenger calls per day and needed to interpret references to more than 500 destinations in 46 states.
The same government briefing and a 2021 Verint interview describe Julie as an AI-based service across phone and web. The U.S. DOT material reports about 20 million calls per year, while Verint reports web-assistant figures from an interview with Amtrak's Manager of E-commerce; these figures describe different stages and channels.
Context
The initial constraint was not only volume. Julie had to recognize natural spoken requests across a broad destination grammar, then provide an automated spoken reply. The U.S. DOT says the application later expanded beyond IVR to a website chatbot and text messages.
This made the early design problem one of controlled coverage: reduce pressure from repeatable travel questions without assuming every passenger request could be resolved automatically. That is why the system focused on concise answers and booking assistance rather than trying to replace the whole customer-service operation.
Ideal customer
Julie serves passengers who need travel information, a service-status update, a reservation, or a path into booking without waiting for an agent.
GTM problem
The practical problem was to separate predictable travel enquiries from the broader call load, so customer-service capacity could focus on work that did not fit a concise automated response.
Positioning
The U.S. DOT describes Julie as taking a natural-language phone request, filtering possible inputs, and returning an automated response. Verint describes the web assistant as providing answers and transaction help, including redirection into Amtrak's booking tool.
The operating model is: use automation for repeatable information and transaction entry points, then retain human capacity for unresolved or non-standard cases. The boundary matters more than the interface alone: it limits the assistant to work where a concise, current response is possible.
Channels
Julie operated by phone and later through Amtrak's website and text messages. Verint says its web deployment was intended to serve a website with more than 375,000 daily visitors.
Reusing the assistant across those service entrances makes self-service available at the point passengers already choose to ask for help. The expansion also has operational value: the same knowledge can address a voice request, a web question, or a text-based request without a separate first-line process for each channel.
Results & evidence
The U.S. DOT reports that Julie answered about 20 million calls per year, averaged about 55,000 calls per day, peaked around 95,000 calls per day, and solely handled about 25% of all phone calls. In the Verint interview, Amtrak's E-commerce Manager said the web assistant answered about 10,000 queries per day, reduced unnecessary contact-center escalation calls by 20%, and achieved more than 90% "appropriate" responses, defined there as an adequate first response.
Verint also says Julie data was used to identify website-content, functionality, user-interface, and product-suggestion issues, turning support interactions into a feedback loop for the wider customer journey.
The evidence supports high-volume self-service capacity and a feedback loop for service improvement. These are service-operation indicators rather than a complete ROI claim: they show how much work Julie handled and how interaction data informed content and interface improvements.
Replicability
Replicability · Medium
Applicable when
- Inquiry and transaction flows can connect to reliable, current data.
- Customers can access self-service through the main support touchpoints.
- High-risk and complex cases have a clear human escalation path.
- The team measures self-service completion, handoffs, and service quality over time.
Not transferable
- The complexity of rail schedules, ticketing, and operations does not transfer directly to other businesses.
- The public sources do not fully disclose Julie's architecture, query mix, or cost structure.
Risks
- Incorrect live schedule or ticket data can directly affect passenger decisions.
- Better self-service metrics do not by themselves prove higher satisfaction or revenue.
- Peak demand may exceed the coverage of existing self-service flows.
Next steps
Identify the repeated enquiries currently consuming human capacity, the authoritative data each one needs, and the conditions that require an agent. Codeleo can turn that service map into a staged AI workflow, define measurable first-response and escalation criteria, and use the resulting interaction data to improve both automation and the surrounding customer journey.
Sources
- Artificial Intelligence and Machine Learning for TransportationU.S. Department of Transportation, ITS Joint Program Office · 2026-08-17
- All Aboard - A Q&A with Allen Sebrell, AmtrakVerint · 2026-08-17