Your AI agent just dialed the cable company. It sat through 40 minutes of hold music, negotiated $20 off your monthly bill, and hung up. Congratulations: you may have just become party to a legal arrangement you never thought about.
This week, SpaceXSI switched on phone calls inside Grok Bot for paid users. The agent can book appointments, wait on hold, and haggle over bills on U.S. numbers. If the representative asks something the bot can’t handle, you get a mid-call notification, you type a reply, the bot relays it, and you can follow the whole thing on a live transcript. Early users have already used it to negotiate a Comcast bill and order a pizza.
The demo is fun. The terms of service are the part worth reading twice.
You are the caller of record
Buried in SpaceXSI’s terms is a phrase that should stop you before you hand over your to-do list: “caller of record.” It means you, the human who told the bot to dial, are legally the person making that call. Every call. If the bot misrepresents something, records a call in a two-party consent state without you realizing that matters, or violates a telephone marketing rule, the liability sits with you. The company positioned itself as a tool provider. You are the caller.
There’s a disclosure mechanic, at least. The bot tells whoever picks up that it’s a “virtual assistant.” So the person on the other end knows they’re talking to software. What they are not told, and what the terms make clear, is that a specific human has authorized this specific call and stands behind it. That human is you.
It’s a clever structure, and worth understanding before you delegate anything sensitive. The feature that makes consumer agents genuinely useful, “go do this phone chore for me,” is also the feature that makes accountability murky. SpaceXSI resolved the murk by assigning all of it to the user. Source: The Rundown AI
There’s one more wrinkle. Responding to complaints about usage limits, Elon Musk said the platform’s cost is “unsustainably high relative to revenue.” Read that as: expect tighter caps. This is a paid feature running on expensive real-time voice infrastructure, and the economics of letting millions of people outsource hold music are not settled. Grok Voice reportedly already handles 20,000 Starlink support calls a day, so the plumbing exists at scale. What’s unresolved is who pays for it.
United Airlines just answered the same question differently
The same week Grok Bot learned to dial, United Airlines told AI booking agents to stay away. The airline will block personal AI assistants from booking tickets, citing data-privacy risks and possible FAA and DOT rule violations if a bot sits between the customer and the carrier. One executive put it plainly: “I don’t want to create another middleman.”
Two companies, same seven days, opposite conclusions. One says our bot can act for you, but you own the consequences. The other says your bot cannot act here at all. Both are answers to the same underlying question: when software acts on a human’s behalf, who is accountable for what happens?
The travel industry spent two decades fighting intermediaries over who owns the customer relationship. It recognizes an unaccountable middleman arriving at speed, and it is drawing lines before the holiday booking rush. United is the first big airline to do it. It will not be the last.
What this means if you run a business that answers phones
If you run a business, you’ve probably been thinking about AI agents as tools you point at your own work. In 2027, they’ll also be arriving on your phone lines as customers.
A bot calling to book an appointment, dispute a charge, or ask about inventory is a customer with infinite patience and a perfect memory of every prior call, owned by a human who may be half-paying-attention. United’s response, walling bots off, works for an airline with pricing power. It works less well for a restaurant, a clinic, or a two-person agency that genuinely wants the booking.
Three things are worth deciding before the calls start arriving:
- Pick your disclosure question now. When a caller identifies as a virtual assistant, what does your staff do with that information? Some businesses will hang up. Some will see a patient, tireless customer and lean in. Either is defensible. “We’ll figure it out when it happens” is not a policy.
- Decide what your recorded-call rules are. If you operate in a two-party consent jurisdiction and a bot records your representative without notice, that’s your problem to catch, not the calling platform’s. Front-desk scripts should treat “this call may be processed by software” the way they treat “this call may be recorded.”
- Know who you’d hold responsible. If a bot books a table for eight and nobody shows up, who ate the cost? The caller of record framework says: the human who sent the bot. That’s worth knowing before you argue about it with someone who didn’t read the terms.
The pattern to watch
The caller-of-record question will not stay confined to phone calls. An agent that files your expenses or messages your clients creates the identical structure: a principal (you), an agent (the software), and a third party (everyone else) who has no easy way to know how much authority the agent actually has.
Contract law has handled principals and agents for centuries. What’s new is the volume, and the fact that the “agent” never signs anything and cannot be sued. Every platform building these features faces the same choice SpaceXSI made: absorb the liability themselves (nobody will), push it to users (cheap, for now), or wait for regulators to impose something (coming, eventually).
For now the answer is you. The bot makes the call. You are the caller. And somewhere in Illinois, a Comcast representative is about to spend 40 minutes negotiating with software that has no idea what a late fee is, while the human who sent it watches the transcript from the couch.
That’s the milestone, really. Forget the hold music. “Who is responsible” now has a checkbox, it gets checked by default, and it’s yours.
