Steve Hatch

Your AI Made the Call. You're the Caller of Record.

Your AI Made the Call. You’re the Caller of Record.

When an AI agent makes a phone call for you, the contract has already decided who’s on the line. You are. xAI put it in writing.

Quick roadmap:

  • What xAI’s Grok Bot Terms say about phone calls, quoted with section numbers
  • Why your bank is making the same move
  • The bug where my own system of record was wrong
  • What a call receipt has to contain if you’re the one on the hook

The argument ended before anyone had it

Grok Bot now places phone calls for people. According to The Rundown, it waits on hold, books appointments and negotiates bills. Teslarati called the first calls “wild.” The coverage is fun. The terms are more interesting.

I’m not a lawyer, and none of this is legal advice. I read contracts the way I read an IAM policy. Who is the principal, and what are they allowed to do? Here’s Section 2.7 of xAI’s Grok Bot Terms, last updated October 31, 2026:

“Grok Bot may place or receive phone calls on Customer’s behalf (“Calls”). Calls are Agentic Actions, and Customer is the caller of record and solely responsible for them, including compliance with all applicable cellular, landline, telemarketing, robocall, caller ID, and call-recording laws, such as the Telephone Consumer Protection Act.”

The same section adds an indemnity: “Customer will indemnify Provider for claims arising from its Calls or use of any temporary phone numbers.”

You might picture “Customer” as a company with a legal department. The definitions say otherwise: “If Customer is an individual using Grok Bot on their own behalf, Customer is also the User.” That’s a person with a phone bill.

Section 2.4 goes further:

“CUSTOMER IS SOLELY RESPONSIBLE FOR ALL AGENTIC ACTIONS TAKEN BY GROK BOT, WHETHER OR NOT SUCH ACTIONS WERE INTENDED, ANTICIPATED, OR AUTHORIZED BY CUSTOMER … PROVIDER IS NOT LIABLE … EXCEPT TO THE EXTENT DIRECTLY CAUSED BY PROVIDER’S GROSS NEGLIGENCE, OR WILLFUL MISCONDUCT.”

Translation: if the agent says something on the phone you never asked it to say, you own the sentence.

The approval controls don’t rescue you. Section 2.3 calls them “aids only” and says they “do not reverse completed actions.” A call happens in real time. There’s nothing to roll back.

The deal runs the other direction too. Under Section 2.7, “Customer, for itself and its authorized Users, agrees to receive Calls, including in an AI-generated voice, from Grok Bot on behalf of other users and from Provider and its affiliates about the Services.” Your agent might be talking to somebody else’s.

Your bank made the same move

Wells Fargo’s updated online access agreement takes effect November 24. Section 15 treats instructions from an AI agent as “authorized directly by you.” American Banker reports the bank also wants customers to prove when they revoked an agent. Different industry, same accounting. Whatever the agent does gets booked to the human who sent it.

I don’t object to the assignment. Somebody has to own the action, and the person who dispatched the agent is the obvious owner. My objection is to the step both contracts skip.

Liability moved. The evidence stayed with the vendor.

If you’re the caller of record, you need the record.

Here’s everything Section 2.7 says on the subject: “Calls identify themselves as made by a virtual assistant on Customer’s behalf and may be recorded. Customer may not alter or disable these disclosures.”

The call may be recorded. The terms don’t say you get the recording.

The product might show you a transcript. I haven’t checked. My point is about the contract. The clause making you liable is specific and printed in capital letters. The clause about recording is passive and never says who keeps the tape. When a contract puts every consequence on you, the evidence you’d need to defend yourself deserves more than a passive verb.

My own system of record lied to me

I built Banjo, an open-source agent that places real phone calls. One bug from the writeup still bothers me. When Banjo left a voicemail, the model would speak a short preamble and then pass the real message into a tool argument instead of saying it out loud. Twice, on real calls, the person heard a sentence and a half while every log in the stack insisted they’d gotten the whole message.

No component was lying on purpose. The database logged what the agent intended. Tool arguments are intent. The phone line is what happened. The fix forced the voice layer to speak the exact message before hanging up, so the recorded version and the spoken version became the same event.

That one bug is the whole receipt problem. If someone disputes a call you’re “solely responsible” for, you’d better know which of those two records you’re holding.

What goes on the receipt

My builder’s checklist reads like an audit log, because a receipt is one.

The transcript of what was spoken. Skip the prompt, the tool arguments and the model’s summary of how the call went. You want the words crossing the line, from both sides, with timestamps.

A recording, where the law allows one. Recording rules differ by state, and Section 2.7 makes compliance with “call-recording laws” your job. When you’re allowed to keep the audio, keep it. A transcript is a model’s reading of the call. The audio is the call.

The disclosure the agent spoke. The terms say calls announce themselves as a virtual assistant. Good. Banjo introduced itself as an AI when it called Xfinity for me, and I want proof of that line on every call. “The system is configured to disclose” is a setting. “Here’s the second it said so” is evidence.

The principal and the scope. Who authorized this call, and what was the agent allowed to do on it? In IAM terms, you need the principal plus the policy in effect when the action ran. Section 2.4 makes this record more important. If you’re liable for actions you never authorized, a record of what you did authorize is the only way to show where your instructions stopped and the agent’s improvising started.

Your own copy. This is the item everybody skips. A vendor’s log belongs to the vendor, lives under their retention policy and reaches you on their terms. Nobody runs an AWS account where the audit trail exists only where AWS can see it. A phone call you’re legally on the hook for deserves the same standard.

Ask for the receipt before the agent dials

The vendors settled who owns an AI phone call. You do, by contract, as caller of record, whether you meant the call to go sideways or not. I don’t expect the terms to change, and fighting it is wasted energy.

The open question is whether you get the evidence. Before any agent dials on your behalf, ask the vendor what you’ll be able to hold up when a call goes wrong.

If the answer is “we may record it,” that’s their copy. Ask where yours is.

Filed under AI, Technology.