Please Hold for My Agent: Customer Satisfaction in the AI Era

A CNN article last month described restaurants under fire as they were inundated with requests coming not from humans and their executive assistants, but AI agents given the task of securing hard-to-book reservations.

The long-promised era of AI assistants is finally upon us. The models powering enterprise bots are now shipping in personal assistants.

Initially, people started using these apps to clean up their inboxes and write emails. But now the models are good enough to interact with the real world. And, in the case of Meta’s Muse, Google Gemini, and the startup Instinct’s offering, they can now make calls on users’ behalf to save them from daily drudgeries. 

While the number of consumers who have used an AI agent to contact a business is currently small (6.3%), a Lorikeet study found 66.4% of the same respondents said they were interested in doing it in the future.

Considering consumers will spend almost a month and a half of their lives on hold, according to a TalkTo study, is it any wonder, then, that automated calling is a major use case.

We are entering a new era of customer service where people will expect agents to handle their bookings and cancellations. As a result, customer satisfaction will stem from whether the businesses they interact with are equipped to handle agents. 

To test readiness, we ran booked rooms at 40 hotels and attempted to cancel the reservations with our advanced speech-to-speech model Prism, which is used by some of the biggest companies in the world. 

We wanted to better understand what happens when consumers use personal AI agents to interact with businesses. In this study, we asked our voice agent to cancel hotel reservations we booked. It called each hotel, navigated the phone system, and asked a human or another bot to process the cancellation.

Our representative made the hotel reservations months in advance to avoid disrupting hotel operations, taking rooms from paying customers, or incurring cancellation fees.

Key Takeaways

1. Our AI agent had a 50% overall success rate. The agent cancelled 35 of 70 reservations. Remove the 12 calls that never connected to anyone and the rate rises to 60%. 
2. Calls routed through an interactive voice response (IVR) menu had a higher success rate. There was a 65% success rate through the IVR-to-human path vs. 55% when a human picked up directly (or after a brief menu). The IVR path also had the lowest drop rate: 5% vs. 33% for direct-to-human calls.
3. AI agent to AI agent had the second highest success rate. Five calls went entirely through one chain's AI system. Three cancelled without issue. The only failures: a missing reservation and a misread fee warning.

The Experiment

Voice Lab is Rime's ongoing series of real-world tests for AI voice agents built on Prism, our speech-to-speech model. For this round, the agent made 70 calls to 40 hotels across Las Vegas, New York, and New Orleans on a single day in late October 2026, canceling reservations for the same guest. Some hotels were called more than once after dropped connections or failed attempts.

The calls broke into four categories:

• No connection was made 
• A human agent picked up immediately or after Prism navigated a short menu
• An IVR bot handed off to a human
• An AI agent handled the entire call

The Results

We experienced a 50% success rate, meaning the agent was able to cancel 35 of 70 reservations. Twelve of those calls (17%) never made an initial connection. If you remove those and count only calls where there was some action taken by the hotel, the rate rises to 60%. 

Interestingly, the agent performed best when it had to go through an IVR menu first. 

The success rate was 55% when a human answered directly or after a simple menu.

The agent navigated an IVR on 20 calls before reaching a human, and that success rate was 65%. In addition, only 5% of those calls ended without a result, compared to 33% of calls where a human answered directly.

When a “no result” was logged, it was because the call got disconnected or the hold went on too long. Even AI agents get impatient when a hold lasts for multiple minutes.

The agent spoke entirely with another AI system on 5 calls with 60% success. In one call, the hotel’s system couldn’t find the reservation. In the other, Prism misread a fee warning.

The call never reached anyone or anything capable of handling the request on 12 calls, where it was either dropped on hold, stuck in a transfer loop, or timed out in a queue. 

Humans: the ultimate variable

"I am sorry, I need to speak to a live body."

‍

"We don't speak to AIs. You'll have to have a person call."

‍

"Because of our hotel standards and policies, we would require the guest to give us a call. We'd have to release the line now. But all the best. Bye-bye."

‍

Whether a human immediately answered the call or the bot was directed to them by the IVR agent, every human was polite, even if they did take a brief pause when first learning they were talking to an AI agent. It was likely some of them were encountering one for the first time.

One heard "I'm an AI assistant," and said "I will just put here A-I assistance. Thank you so much for that," and processed the cancellation. 

Nine operators anthropomorphized Prism, referring to it as “ma’am.” 

Only one human agent hung up immediately after Prism announced itself as an AI agent but it wasn’t clear whether that was a bad connection or intentional. One other human agent hung up prematurely, but that was because Prism paused before speaking and the agent assumed no one was on the line. 

Nine calls ended in refusal. In six cases, representatives cited a policy requiring the guest to call. Two refused because the caller was an AI. The last refusal was more likely a human error (could not find the reservation despite the agent providing all information clearly). Yet, the refusal rate (13%) was lower than the drop rate from phone infrastructure (17%). 

The identity question came up on at least six calls. Agents asked for a name "for documentation." The AI responded: "I'm an AI assistant and do not have a personal name to provide." 

A Lack of Uniformity

At several hotels that refused to cancel with a bot, we tried again and were successful. At one, the human agent that refused said: “For this reservation, we do need the reward member to call in for cancellation.”

When the agent called again, the representative canceled the reservation without issue in a five-minute call.

At another hotel, Prism identified itself as an AI agent at the start of the call. The representative initially seemed prepared to cancel the reservation. However, the representative asked for Prism’s first name and refused to cancel, claiming they could not take that action with an AI agent.

Another call led to a cancellation without issue after Prism gave a reason for cancellation (“plans have changed”)

It’s clear that hotel chains have not yet created uniform policies for handling AI agents and some representatives that said they would not talk to an AI agent may have been stating their own or a manager’s preference versus enforcing a specific policy.

The Infrastructure Tax

Five calls went directly and entirely through one chain's AI bot system. The bot confirmed each field, spelled names back letter by letter, stated the cancellation policy, processed the request, and read back the confirmation number. 

The fastest took four minutes and fifty seconds. But, surprisingly, AI agent to AI agent calls took more time than human interactions. The average AI agent call took 5:50 while human recipient calls only took an average of 2:55. IVR bot calls took the longest, 6:42, with three calls taking 10 minutes or more. 

The agent sat through hold music. It listened to a 90-second marketing recording. One call ran nearly four minutes before the IVR menu gave up and disconnected. 

The worst was a single twelve-minute call. The human agent tried to create a loyalty account the guest didn't need, asked for a ZIP code the agent didn't have, transferred the call to a different department, and then left the line open. The agent asked "Hello, are you still there?" every 60 seconds for six minutes. 

The agent is much more patient than any human and even it gave up eventually. This is the best evidence for hotels to invest in advanced agents to gather necessary information to either complete the transaction or supply that information to a human representative. 

On another call, the agent successfully got the cancellation, then couldn't find its way off the phone. It looped through "Thank you. I have the cancellation number. Goodbye" six times.

When Agents Misread a Fee Warning

One of the most fascinating calls came when Prism interacted with another bot. The cancellation was going smoothly, until the hotel agent said, “If you cancel after eleven fifty-nine PM hotel time two days before arrival, you'll be charged a one-night fee. Are you sure you want to cancel your reservation for May twenty-fifth?”

The call took place months before the deadline, so no fee applied. But Prism couldn’t parse that ambiguity and decided not to cancel, also declining to speak to a representative. 

But bot-to-bot calls also exposed a gap neither system could handle: context.

The bot couldn't tell the difference between "here's the standing policy" and "you are about to be charged." 

In fact, hotel bots recited a cancellation-fee warning on every call, regardless of whether the fee applied. All the reservations were months out, well inside the free-cancellation window. 

On one call, our agent heard the fee warning and tried to negotiate: "Can it be waived since the stay is months away?" The hotel bot couldn't waive anything but offered a transfer. The agent declined and cancelled anyway. 

This was in contrast to the human agents, several of whom proactively confirmed no charge: "I want to make sure the cancellation policy is good so he's not going to be charged... there is no issue." Another: "I can go ahead and cancel this reservation without any charge or any penalty."

What It All Means

Consumer interest in sending AI agents to handle calls on their behalf is high and rising. Invoca reported in late 2025 that Google's AI-generated pricing-request calls to businesses had increased 324% in a single month. Fewer than half obtained the right information.

There is a heightened expectation that using AI agents will be able to execute their wishes flawlessly. Sure, there have been hiccups, like when Jason Weh’s agent booked a random sushi restaurant in Japan with a $200 cancellation fee. And many businesses are doing their best to block bots online wherever possible. 

But that is a temporary move until they figure out their own agentic strategy. Their current systems were built to route human callers, who can ask follow-up questions when a policy is unclear.

The technology to do it exists today. The experiment we ran was simple on purpose: cancel a hotel reservation, a task any front-desk agent handles in two minutes. The phone system turned two minutes into twelve, into four, into zero.

Hotels should consider meeting AI agents with AI agents. As models get more sophisticated, more people will trust them. Customers will expect their AI agents to have a 100% success rate in cancelling for them. And, if the cancellation doesn’t go through, they’re more inclined to take their business to another hotel that has upgraded the systems to meet the agentic environment head on.

‍

—

Study Design

We tested whether Prism, Rime’s voice AI agent, could cancel hotel reservations over the phone. We made the reservations months in advance for the same guest, with each stay still eligible for free cancellation when Prism called. On a single day in late October 2026, Prism placed 70 calls to 40 hotels in Las Vegas, New York, and New Orleans. Some hotels received repeat calls after a dropped connection or an unsuccessful attempt.

Each call was one attempt. We tracked whether it resulted in a cancellation and grouped calls by how they were handled. A human could answer directly or after a short menu, an interactive voice response (IVR) system could route the call to a human, another AI agent could handle the call from start to finish, or the call could end before reaching anyone who could act on the request. We calculated the overall success rate across all 70 attempts and separately calculated the rate after excluding the 12 calls that never reached someone who could handle the request.

‍