Beyond the Midnight Closing: One AI Agent, Five Bus Brands, Round-the-Clock Coverage

About Grupo Flecha Amarilla

Grupo Flecha Amarilla has been moving people across central-western Mexico since 1932, when it began running buses out of León, Guanajuato. Nine decades later it operates five intercity passenger brands — Primera Plus, the group’s flagship line, along with Coordinados, Flecha Amarilla, Unebus, and Roll and Bits — plus a parcel service that rides on the passenger network: packages travel in the holds of buses already making the trip, terminal-to-terminal in under 24 hours.

That is a lot of separate businesses sharing one set of customers. And sooner or later, those customers pick up the phone.

A Three-Year Search

When ChatGPT first launched, Juan Pablo Martín Yamín, Chief Business Officer at Grupo Flecha Amarilla, immediately recognized the shift ahead for mobility and passenger transport. He launched an internal project to build a conversational AI agent for passenger inquiries, but the investment was too large and the effort was shelved.

“In hindsight, the timing saved us. A year later, robust solutions existed — but I was still skeptical about implementation,” says Juan Pablo. “Every vendor tells you integration is easy, that it’s all APIs. I laugh, because I know our systems are not unified in architecture, and there is no data governance.”

That changed when Juan Pablo heard HappyRobot co-founder Javi Palafox speak at an industry event. He went looking for him afterwards, found his brother Pablo instead, and ended up talking to both. What he saw was a platform that could handle complex voice interactions at low latency, paired with Forward Deployed Engineers who could work directly with legacy telephony infrastructure.

He came back with a pilot in mind.

What One Phone Line Had to Cover

In central-western Mexico, bus schedules face conditions no one can control: highway incidents, road disruptions, and blockades. When a route changes at 11:00 PM, ticket-holding passengers have one lifeline: the phone.

Flecha Amarilla’s buses operate 24 hours a day — but until this year, its contact center closed at midnight. In that gap, about 200 calls a day went unanswered. During peak travel periods like Semana Santa and the December holidays, volumes spiked sharply, and even with seasonal hiring, many of those calls went unreturned.

Passengers were left without real-time information during the late-night disruptions. A passenger who can’t find out at 1:00 AM whether their bus is running either makes an unnecessary trip to the terminal or doesn’t travel with the group at all.

The hours were only half of it. One contact center answers for every brand in the group, and each carries its own routes, fares, and policies. A caller might ask about one brand’s schedule and another’s fare in the same conversation. Whoever picks up has to know all five.

Building "MarIA"

The pilot Juan Pablo came back with was an AI agent that would sit on the main customer line and take the calls the human team couldn’t reach. They called her MarIA — the name María, with IA, the Spanish initials for artificial intelligence. He handed the project to Simón Fusco, the group’s Market Intelligence Manager and Salesforce Solution Leader — an obvious call, since the contact center and customer service report to him, and he owns the group’s Salesforce CRM, which gave him, in his words, a 360-degree view of where the problem actually sat.

Simón took the business side of the build and became the bridge to everyone else. What MarIA says, why she says it, and how she responds all came from him. Flecha Amarilla’s own IT team learned the platform and handled the IVR configuration and integration work alongside him. In the early months every conversation with HappyRobot routed through Simón; today the two teams work directly.

The division of labor was clear. Simón brought operational knowledge: how the five brands differ, what passengers usually ask, and which calls have to reach a person. HappyRobot brought the platform and the build. Antonio González Ferreira was the Forward Deployed Engineer embedded with Simón’s team throughout. “Antonio understood what Flecha Amarilla is,” says Simón. “He knows the business. He developed the new version himself, improved it, showed it to us — and we said: yes, this is what we need.”

What started as a narrowly scoped pilot is no longer treated as one. Simón’s stated goal is for MarIA to do what a human agent does — the whole job, not a subset of it.

Chosen, Not Assigned

Rather than routing callers to an AI agent by default, the team placed MarIA as an option inside the existing Interactive Voice Response (IVR) menu on the main customer line (+52 477 710 0060). Callers press 1 to reach her — they have to choose her. Around 20% of inbound traffic does, which works out to roughly 4,000 calls a week.

The coverage gap closed quickly. Unanswered calls fell from around 200 a day to fewer than five, with MarIA working the overnight hours when the human team is offline. Of the conversations she handles, 60% close without an escalation — routine route, schedule, and policy inquiries now resolve on their own, which leaves the human team for the complex cases that cluster around holiday surges. Today, Flecha Amarilla holds MarIA to the same KPIs and accuracy standards as its human agents.

The First Real Test

It came sooner than anyone expected.

In late February 2026, an internal incident in Mexico disrupted road travel across several states. Passengers called all at once — people mid-journey, people due to travel the next morning, people trying to work out whether to cancel or find another route. It remains the highest-volume moment in MarIA’s history.

The human contact center saturated within hours — but MarIA didn’t. She kept answering, at a volume no amount of emergency staffing could have absorbed on that notice, and callers who had the option of waiting for a person kept choosing her instead.

“The contact center was blocked by the sheer number of calls,” says Simón. “People did choose MarIA.”

What she didn’t have was information. Her scope was schedules, routes, fares, and policies, and nobody had yet built her a source for live incident updates, so for the questions specific to that week she transferred what she could. The conclusion the team drew was that MarIA needed more to work with.

A dedicated incident knowledge base is now on the roadmap — with defined responses and a named owner keeping it current, so the next time an incident disrupts the network, MarIA has the information she needs.

“People aren’t afraid of AI agents when they actually solve their problems. Callers just want quick, accurate answers.”

— Antonio González Ferreira, Deployment Lead LATAM, HappyRobot

Retiring the Phone Menu

Simón Fusco and Jaime Jaime Von Nacher, the group’s Chief Commercial Officer, flew to San Francisco in September to review progress with the HappyRobot team. They ended up at a whiteboard instead, mapping the call flow to see what would change if MarIA answered first rather than waiting behind the menu.

The 90 seconds a caller spends working through options would become about 20 seconds of “Hola, soy MarIA, ¿en qué puedo ayudarte?” — and from there she would carry straight on into the request rather than handing it off. Jaime called it an eye-opener.

By the end of the session they’d decided to remove the menu altogether. The approval to let MarIA take every inbound call and route it herself has since come through.

What’s Next

Simón calls the next version MarIA 2.0, and the integration work is the bulk of it. Flecha Amarilla’s IT team is building it alongside HappyRobot’s engineers.

• End-to-end reservations and sales: expanding MarIA from information lookup to booking and selling tickets.
• Proactive WhatsApp notifications: pre-trip reminders two days and one day out, live alerts when a route changes mid-journey, and a post-trip thank-you with a short satisfaction survey — over WhatsApp and SMS rather than another app to download.
• Conversational commerce: integrating MarIA with the group’s e-commerce platform, so passengers can request a trip, receive options, and buy tickets with QR codes inside a WhatsApp chat.
• An agent per business unit: beyond the contact center — technical support, bus rental, franchises, and parcels. Packages ride on the buses and arrive terminal-to-terminal in under 24 hours; an agent would handle tracking by voice, WhatsApp, or web.

That last one closes a loop Juan Pablo opened three years ago. The day ChatGPT launched, he told the founders of the group’s e-commerce platform that travel booking was about to stop being a form to fill in and start being a conversation — you would simply say where you needed to be and when, and get an answer.

“Nobody has delivered that yet, not even the global travel platforms,” says Juan Pablo. “What excites me about working with HappyRobot is that it finally looks like we’ll get there.”