AI on both sides of the conversation
First, Hertz put AI in front of its own customer care. Now customers bring their own AI assistants, and Hertz has to be there too.
Built with DecagonAt a glance
- My role
- Owner of AI chat and voice for Hertz customer care. Owner of Hertz’s strategy for AI assistants and agents.
- Team
- Customer care operations, engineering, workforce management and Decagon’s team
- When
- Vendor evaluation in 2024. AI chat live in February 2025. The Hertz app in ChatGPT live in June 2026.
- What changed
- Chats handed to a live agent fell from 14% to under 6%, and travelers can now compare Hertz cars inside ChatGPT.
There are now two kinds of AI in a customer conversation with Hertz. One is ours: the AI agent that answers when a customer chats or calls. The other is theirs: the assistant a traveler already uses to plan a trip.
I lead product for both. This is the story of each, and of the one lesson they share: define what success means before you celebrate it.
Part 1: AI in front of Hertz’s customer care
The problem
Many customers contacted care for routine tasks: extending a rental, changing a booking, asking about a policy. Self-service should have handled these. Each contact cost the customer time and cost Hertz an agent’s time.
Context
Hertz compared “legacy players, tech giants, and startups” before choosing Decagon. AI chat went live in February 2025. The AI voice agent, Hailey, followed in 2025 and reached every primary phone line in about eight months, one line at a time.
My role
I owned AI chat for customer care, working closely with the customer experience and Decagon teams “from vendor evaluation to implementation and going live”. I have owned AI voice since its launch. I run the partnership with Decagon: weekly operating reviews, quarterly business reviews, the joint roadmap and incident response.
I owned Hertz’s proactive outreach with Decagon as well. Before a rental ends, the AI agent contacts the renter about their return plans and updates the reservation, so the car is ready for the next driver.
In 2026, Hailey came to the Hertz iOS app, where customers can now book a car by chat. I partnered on that launch and owned its evaluation.
I also wrote the specs for Hertz’s AI call analytics and built the drilldown app that leadership uses. Engineering built the transcription pipeline.
Approach
- Launch narrow, with a person one step away. Then widen. Every new line or brand went live in small ramps behind a one-click off switch.
- Read the calls, not just the dashboard. A “deflected” call can be a customer who gave up. I read AI-handled conversations line by line.
- One honest definition. I defined “effective deflection”: a contact counts as resolved only if the customer doesn’t come back, give up or get escalated.
- Hold the vendor to the same standard. When something broke, we ran root-cause reviews and added safeguards.
- Never promise what the system can’t do. One rule from the incident runbook I wrote: the AI must never promise a transfer the phone system can’t complete.
Key decisions
- Choosing the vendor. We compared vendors on channels, fit with our customer systems, engineering effort, model flexibility and price.
- Reporting voice honestly. I explained to leadership that voice resolves fewer contacts by design on complex lines, and reported it that way instead of averaging it away.
- Rebuilding the voice agent. I coordinated Decagon’s six-week rebuild of Hailey. The new design verifies the caller once, offers only the actions the booking system can complete, and asks a clarifying question before handing off to a person.
- Reusing what worked. In 2026 we brought AI chat and voice to Dollar and Thrifty with the same setup. Voice reached every call within days, with each step signed off by operations.
Results
- AI now serves 72% of Hertz’s US inbound chats.
- The share of chats that ended without a handoff to a live agent rose from about half to about three in four within a year.
- Chats handed to a live agent fell from 14% to under 6%.
- The rebuilt voice agent raised resolution without an agent by 5 points, from 30% to 35%, while customer satisfaction held steady.
- More than 2 million customer calls have been analyzed with AI. About 500,000 a month are sorted into a five-level tree of reasons, on a fixed codebook so months stay comparable.
- Across the whole program, Hertz’s human contact rate fell about 11% from 2024 to 2025. AI was one lever among several, so I don’t credit it alone.
Part 2: Hertz inside customers’ AI assistants
The problem
Travelers started planning trips in AI assistants. Those assistants could show rental cars, but often couldn’t book them, and some of our competitors were already there. If a traveler’s assistant can’t find or book Hertz, the conversation never reaches us.
Approach
I lead Hertz’s product strategy for AI assistants and agents. It rests on four steps an assistant has to take for Hertz: find us, choose us, book us, and bring the traveler back. There are two ways to win them:
- Be the answer. Show up accurately when an assistant answers a travel question.
- Be bookable. Let travelers compare and choose a Hertz car inside the assistant.
Underneath both is one Hertz service, built once so any assistant can plug into it. Checkout, payment and loyalty stay on Hertz.com.
Results
- In June 2026 we launched the Hertz app in ChatGPT, where travelers compare cars and get recommendations before booking on Hertz.com. Hertz pointed travelers to it in August 2026.
- I ran my own test bookings in it and turned the gaps into the next round of work.
- Every assistant gets a scorecard that we review with Finance.
What the data can’t show yet is whether these bookings are new to Hertz or would have happened anyway. I say that plainly when I report it.
What I learned
- The entry point matters as much as the model. Putting the voice agent at the front of the phone line, where callers simply say what they need, mattered as much as the model behind it.
- Reported deflection and real resolution can be far apart. Define the metric before you celebrate it.
- Separate referrals from bookings, and count the customers who are new to the brand.
- Build once for every assistant, instead of a new integration each time.
“Our goal wasn’t just to ‘add AI to voice.’ It was to remove friction from one of the most common customer needs.”