Who is actually on the line when a calm, well-spoken voice asks your front desk whether Thursday at four is still open? Increasingly, the answer can be an assistant working for a person you have never met — and the phone network was never built to tell you that.
This note takes one call, a booking or availability or quote request placed by an AI assistant, and follows it one layer at a time. It covers what reaches the business, in the order it arrives, and then what the line cannot carry at all.
The Call, As It Actually Arrives
A business phone line receives a surprisingly small amount of structured information before anyone says hello. Everything after that is audio, and audio is a claim rather than a credential.
Here is the order in which the signals land:
- The calling number. A telephone number presented to caller ID, which may or may not belong to the party that placed the call.
- The carrier's attestation, if any. A verdict from the originating provider about how well it knows that number, carried alongside the call where the networks support it.
- The voice's self-description. Whatever the caller says about itself in the opening seconds — a name, a company, a purpose.
- The conversation itself. The request, the follow-up questions, and the answers your staff give back.
- The record after the call. A recording, a transcript, or a post-call review, where one exists.
Each layer tells you something different, and none of them tells you everything. The sections below take them one at a time.
Layer One: The Calling Number
The first thing to arrive is a number, and the number is the weakest signal on the list. The FCC's own Caller ID Spoofing guide (checked 2026-09-28) describes spoofing as a caller deliberately falsifying the information transmitted to your caller ID display to disguise their identity.
That same page notes that spoofing can also be used legitimately — for example, to display a business's toll-free number instead of the line the call physically leaves from. In other words, even an honest caller ID may show a number chosen for presentation rather than for traceability.
Keep in mind that an assistant placing a call on a person's behalf has three plausible numbers it could present: the vendor's own outbound number, a number provisioned per user, or the user's own phone number. The number alone does not tell your front desk which of these it is looking at.
Layer Two: Whether The Carrier Vouched For It
The second layer is caller ID authentication. According to the FCC's call authentication page (checked 2026-09-28), the Commission adopted rules in 2020 requiring voice service providers to implement the STIR/SHAKEN framework in the IP portions of their networks by June 30, 2021.
STIR/SHAKEN lets the originating carrier sign the call with a statement about how well it knows the caller and the number. The FCC describes the result as giving subscribers more confidence that the caller ID information they receive is accurate, and as supporting providers' blocking decisions and call labeling.
That said, the attestation is a statement about the number, not about the entity behind the voice. The attestation levels, the signed token, and how it survives non-IP segments of the network belong to a closer read — our standards hub covers the framework itself.
What matters on the channel is narrower. A well-attested call tells you a carrier stands behind the number; it says nothing about whether a person, a script, or a model is speaking.
Layer Three: What The Voice Says About Itself
Once the call connects, the most useful signal is also the least verifiable one — the caller's own description. Some assistant vendors now publish what their calling features do, and those published descriptions are the best available evidence of what a front desk should expect to hear.
Google is the clearest public example. Its Business Profile help page, About automated calls from Google on behalf of customers (checked 2026-09-28), states that Google may use automated calls to the number on a Business Profile to help customers with inquiries such as pricing, availability, and appointments.
The same page says these calls "are monitored and recorded for quality assurance," and that "Google will never ask for payment, credit card information, or account passwords." It also lists the opt-out routes: telling the caller to remove the business from its list, or switching off the relevant option under "Google automated calls and text messages" in the Business Profile advanced settings.
Google's consumer-facing guide to agentic calling (published 2025-11-25, updated 2026-07-24 per the page; checked 2026-09-28) describes the other side: a shopper searches with "near me," answers a few questions, and Google calls local stores about availability and discounts in categories the page names as toys, health and beauty, and electronics.
Note that neither page, as checked, publishes the exact opening script the caller uses or the number it presents to caller ID. Where the vendor's own documentation does not settle a point, we do not fill it in.
What A Self-Description Can And Cannot Establish
A voice that says it is an automated assistant calling for a customer is giving you useful routing information. However, it is a spoken assertion, and the business has no cryptographic way to check it in the moment.
This is the core asymmetry of the phone channel. On the web, an agent can now sign its requests — see what a signed agent request looks like and our note that agents have started signing their requests. A phone call has no equivalent field for the operator's identity beyond the number and its attestation.
Layer Four: What Only The Record Shows
Some facts about a call are simply not visible while it is happening. They surface only in a transcript, a recording, or a structured review after the line goes quiet.
Examples of what a post-call review can show include but are not limited to:
- Consistency of the self-description. Whether the caller identified itself the same way at the start and the end, and whether it answered "are you automated?" when asked.
- The shape of the request. Whether the request was narrow and fully specified — one service, one date range, one price question — which reads differently across a transcript than it does live.
- What was committed. Whether your staff confirmed a booking, quoted a price, or only answered a question, which matters when nobody shows up at four on Thursday.
- Whether a handoff occurred. Some calls begin with an automated voice and end with a person, or the reverse, and only the record shows where that transition happened.
All of these are after-the-fact observations. They improve the next call's handling; they do not change what the current one could prove.
What The Line Can Never Carry On Its Own
Several facts that a business most wants to know are not present anywhere in the call path unless the caller chooses to say them. Even then, they arrive as speech rather than as verifiable data.
Who The Principal Is
The person the assistant is working for — the principal — is invisible to the network. Caller ID identifies a number, and STIR/SHAKEN attests to the number; neither field names the human who asked for the booking.
If the assistant gives a name, that is useful for the reservation book. Keep in mind that it is still the assistant's claim about its principal, not an identity check.
Whether The Agent Is Authorised
A booking request implies that someone authorised the assistant to make it. The phone channel has no field that carries that authorisation, no mandate, no scope, no spending limit.
On other channels, open work on mandates and signed requests is starting to fill that gap. On a voice call, the authorisation lives entirely in the assistant vendor's systems and the principal's account, well out of the business's view.
Whether A Person Is Listening
A call can be fully automated, fully human, or a live person reading from a brief an assistant prepared. That third case — a human proxy relaying an agent's plan — sounds like an ordinary caller and carries none of the automated-call markers.
We treat that pattern separately in the third kind of caller. For the phone channel, the practical point is that "a human voice" and "no agent involved" are two different claims.
The Regulatory Backdrop, Briefly
The FCC has spoken to AI voices on calls, though not to the booking case specifically. In its February 8, 2024 announcement (checked 2026-09-28), the Commission said it had adopted a Declaratory Ruling recognising that calls made with AI-generated voices are "artificial" under the Telephone Consumer Protection Act.
As the FCC describes it, calls using such voices therefore fall under the TCPA's rules for artificial or prerecorded voices, including consent requirements for the calls those rules cover. How that ruling applies to an assistant calling a business at a consumer's request is a question the announcement does not settle, and this note does not attempt to.
Side By Side: What Each Layer Tells The Front Desk
The table below consolidates the four layers against the three questions a front desk usually cares about.
| Signal | Arrives | Tells you | Does not tell you | |---|---|---|---| | Calling number | Before answer | A presented number | Who placed the call | | STIR/SHAKEN attestation | Before answer, where supported | How well the carrier knows the number | Whether a model is speaking | | Self-description | First seconds | What the caller claims to be | Whether the claim is true | | Transcript or review | After the call | What was said and committed | Who authorised it |
Overall, the stack gets richer as the call goes on, but every layer is either about the number or about the caller's own words. Nothing in the path carries the principal's identity or authority as data.
What A Front Desk Can Do With A Call It Cannot Fully Identify
The answer is not to refuse the call. An assistant asking whether Thursday at four is open is, in the most common case, a customer who delegated a phone call — and the booking is real revenue.
Here is how we would structure the handling, in order:
- Answer it. Treat the request on its merits, the same way you would treat a caller whose name you cannot verify.
- Record the label. Write down what the caller said it was — automated, on behalf of a named customer, from a named vendor — alongside the number and time. That label is the receipt.
- Confirm through a second channel. For anything that commits inventory or money, confirm with the principal directly by text or email to the contact the assistant provided.
- Route it. Send assistant-originated requests down a path built for them, so they can be counted, reviewed, and followed up without clogging the queue for people who called in person.
Remember that the goal is recognition, routing, and a receipt, not certainty the channel cannot give. The channels hub compares what the phone line carries against web, chat, and forms, and the layers hub breaks down the reading, acting, human-proxy, and adversarial cases.
For how often these calls appear relative to other agent traffic, the current edition of the Aethelforge Index carries the first-party figures. If you are deciding how your own front desk should label and route these calls, our method lays out how we read each channel before we recommend anything.