At 4:17 on a Wednesday afternoon, two AI agents had a conversation about my hotel room. One was Hyatt's, the automated reservations line for Hyatt Union Square in New York. The other was mine. Hyatt's agent spelled my last name back one letter at a time and asked three separate times whether it had heard correctly. Mine answered every question in under a second, pointed out that the stay was months away and the free-cancellation deadline hadn't passed, and got off the call with a confirmation number. I wasn't on the line. I heard it the way you hear about any errand someone runs for you: afterward, in a recording.
That was the 71st call of the day, and the moment I understood what we had built.
There is a lot of talk right now about personal agents. Instinct, Muse, Dots: each promises an assistant that handles your life for you. None of them make phone calls. That is a strange gap, because the phone is still where the most tedious work in consumer life happens. Cancellations, billing disputes, rebookings, the call you put off for a week because you already know what the hold music sounds like. Most businesses have no API for "cancel my reservation." Every one of them has a phone number.
I work at Rime, where we just launched Prism, a speech-to-speech system for voice agents. Prism hears the caller's voice directly, where a typical voice stack only ever sees a transcript of it. I wanted a personal agent that could pick up the phone on my behalf, so I built one on Prism and gave it the most annoying errand I could design. Ankit on our team wrote the agent harness. The whole thing was a few days of work.
We booked 54 hotel rooms across New York, New Orleans, and Las Vegas: 41 properties and more than 20 brands under the Marriott and Hyatt umbrellas, from a $69 room at the Excalibur to an $849 room at the Ritz-Carlton New Orleans. Every booking was refundable and months out, so the only thing at stake was whether the agent could get the job done. It got what a human assistant would get: my name, the confirmation numbers and dates, and the two facts hotels use to check that you are who you say you are, the last four digits of the card and the email on file. Then it called every property and asked to cancel.
It cancelled 36 of the 54 in a single afternoon. Along the way it worked through touch-tone menus four levels deep, Marriott's and Hyatt's own voice bots, and live front-desk staff who put it on hold, transferred it, asked it to spell things, and once told it "no" after it read back a cancellation number wrong. It asked again and got the right one. Ten hotels refused to cancel because the caller was an AI or wasn't the guest. Most of the other eight were hotel phone systems that dropped the call mid-transfer, twice in a row.
The calls I keep replaying are the ones where my agent talked to theirs.
Hyatt's reservations line is an AI agent. On the six calls where mine reached it, I measured both sides from the recordings: how long each agent waited after the other stopped talking before it started its reply. Hyatt's agent took a median of 3.82 seconds. Prism took 0.88, over the same phone line. Hyatt's agent also confirmed every detail back as a yes-or-no question, so one cancellation took about three minutes of back and forth. Marriott's lines open with a voice bot that asks what you're calling about and whether you're a Bonvoy member, then plays a membership pitch before handing you to a person. Across 19 calls, that took a median of 76 seconds before a human said hello.
Every one of those seconds was designed to land on a customer. They landed on my agent instead, and that is the shift. Phone trees, hold queues, and slow bots are a cost the brand passes to the customer, and sometimes the cost is the point: a cancellation flow that depends on the caller giving up is a retention strategy. A personal agent doesn't give up. Mine sat through the Bonvoy pitch nineteen times, and on one call waited eight minutes on hold for a loyalty desk that never picked up.
Friction worked as a strategy because customers get tired. Agents don't.
That changes what a brand's own voice agent is for. When the caller is a person, a slow bot is an annoyance the brand can absorb. When the caller is another agent, every extra second of latency is just phone minutes and compute the brand is paying for, with nobody on the other end to impress. The bar for a brand's voice agent is now set by the agent calling it.
The refusals taught me more than the successes.
Every one of the ten refusals came from a human. Neither Hyatt's bot nor Marriott's ever declined to deal with my agent. The humans were also inconsistent with each other. A Marriott reservations associate cancelled my Element New Orleans booking before noon; two hours later, another associate on the same line refused the second booking at the same hotel because the caller was an assistant. A reservations agent for Le Pavillon in New Orleans said, "We don't speak to AIs." The Cosmopolitan cancelled one of my bookings and refused the other. There is no policy here. There is whoever picks up.
That is the decision brands need to make, and most haven't noticed they're making it. My agent opened every call with the same sentence: "Hi, this is an AI assistant calling on behalf of Nick Salzman." Ten hotels heard that and said no. If brands punish disclosure, they will teach agent builders to stop disclosing, and every agent that calls will sound like a person. I'd rather brands build a way to say yes: a verified handoff that proves an agent is acting for a real customer, with that customer's permission.
Right now the only check is the one built for humans. Hotels asked for my name, my email, and the last four digits of my card. My agent had all of it because I gave it all of it. Those questions can't tell a delegated agent from the guest, and they can't tell a delegated agent from a fraudster either. Agent identity is going to need its own front door.
Running this also showed me what a personal voice agent needs before people will trust it with anything that matters.
It has to check its work somewhere other than the call. Twice, a clerk said "yes, that's correct" to a cancellation number my agent read back, and the confirmation email that arrived later carried a different number. A person's "yes" on a phone line is a low bar for proof. My agent's real receipt was my inbox, and a good personal agent will reconcile against it on its own.
It has to carry my judgment. Hyatt's script says a one-night fee applies if you cancel after the deadline. My agent heard "fee," asked for a waiver it didn't need, and on one call declined to cancel at all. One added sentence in its instructions fixed it, and the next call went through cleanly. Every personal agent will need a version of that: which fees I'll accept, what I'll never agree to, and when to stop and come back to me.
And it has to have phone manners and a fuller picture of me. Three clerks said "hello?", heard a beat of silence, and hung up; on the phone you get a few seconds to prove someone is there. One hotel wouldn't cancel without my full legal first name, and my agent only knew "Nick."
Hotels were a test bed. The same pattern shows up anywhere a business still runs customer operations by phone: an airline rebooking a cancelled flight, an insurer working a claim, a clinic moving an appointment, a utility disputing a bill. Each has a phone number, a hold queue, and a script, and each is about to start getting calls from agents. The brands that decide now how to recognize, verify, and serve those agents get a faster and cheaper channel. The ones that leave it to whoever answers will end up with a different policy for every employee on shift.
I spent no time on hold for 54 hotel bookings. The four hours and forty-seven minutes still happened; my agent absorbed them. Ten hotels heard it introduce itself as an AI and hung up. The next agent that calls them might not introduce itself at all.
~1,500 words