AethelforgeSTAGING
Read the Index

Research · The layers

The third kind of caller: a person on the line, an agent on the brief

Muse's first calls were placed by people reading an assistant's script. A phone label that stops at bot or human cannot describe that caller, so ours does not stop there.

Published 24 Sep 20264 min readAethelforge editors

In one paragraph

What is a human-for-agent call?

A call placed by a live person because of someone else's AI assistant. The voice is human; the brief, the intent and the follow-up belong to the agent. The label records the principal, the intermediary and the intent separately instead of guessing from the voice.

Are you sure the last call your receptionist screened was a person? It sounded like one. It probably was one. And that is precisely why the phone is the hardest channel to label, because the question that used to settle the matter, human or machine, has quietly stopped being the whole question.

What Meta's tests actually did

When Meta began testing a calling feature for Muse, its personal assistant, the calls were not placed by a synthetic voice. According to reporting by 404 Media and Tekedia in September 2026, they were placed by trained contractors in a call centre. The contractor received the assistant's brief, made the call, spoke to the business as any caller would, and then reported the transcript and a summary back to the assistant that sent them. Meta has said the feature will roll out only when it is ready and with the proper disclosures, which suggests the company knows exactly how this looks from the other end of the line.

From the receptionist's chair, though, that call is indistinguishable from a person calling for themselves. It attests correctly on the network. There is breathing, hesitation, a real accent. Every screening layer built to catch robocalls waves it through, and every label built on the assumption that a human voice means a human customer records it as one.

Three parts, not two

The way out is to stop labelling the voice and start labelling the call. Every inbound contact has three parts worth recording separately, and on the phone they can all be different:

  • The principal. The person or business the call is ultimately for. In Meta's tests, the Muse user who asked for an appointment.
  • The intermediary. Whoever or whatever is actually on the line. A synthetic voice, a contractor, the principal themselves.
  • The intent. What the call is trying to get done, and whether it is one of several parallel attempts.

A robocall has a machine intermediary and no real principal. An ordinary customer has a human intermediary who is also the principal. A synthetic assistant has a machine intermediary and a human principal. The contractor call has a human intermediary, a human principal, and an agent standing between them holding the brief. Two labels cannot express that; three fields can.

Why the network cannot help here

It is worth being precise about what the phone network does and does not tell a business. STIR/SHAKEN, the attestation framework carriers now apply to calls, says how confident the originating carrier is that the caller is entitled to the number it is calling from. Full attestation is a strong signal against spoofing, and the receptionist reads it first for exactly that reason. However, attestation is about the number, not the speaker, and never about the reason for the call. A contractor placing a call for an assistant will attest perfectly well, because nothing about the call is spoofed.

Web traffic has an answer to the identity question in Web Bot Auth, where an agent signs its requests. The phone has nothing equivalent. No standard says "this call is placed by an agent on behalf of this person," and none is on the horizon. Until one exists, the only place the information can come from is the conversation itself.

How the label is applied

The receptionist that answers the monitored line therefore does two things in order. During the call it screens as it always has: attestation level, number reputation, whether anyone speaks at all, and a gatekeeping greeting for calls that look suspect. After every call with speech, an analysis of the transcript returns a caller kind — human, AI agent, human for an agent, recording, or unknown — together with whether the caller declared itself, whom it said it was calling for, which assistant vendor it named, and up to three short quotes that support the verdict.

Two rules keep the label honest. First, the analysis never defaults to human; a call with too little evidence is unknown. Second, calls logged before the label existed are recorded as unknown or recording, never quietly upgraded to people. The current edition of the Index reports the phone channel on exactly those terms, and the layers page sets out the full label every channel shares.

What a business should listen for

Remember that the tell is rarely the voice. It is the shape of the request. A caller who has never been to the practice, wants availability rather than an appointment, asks the same two questions in the same order, and offers a callback number that belongs to someone else is describing an agent's brief, whoever is reading it. Disclosure rules, when they arrive, will make this easier; Google's Duplex demonstration in 2018 produced exactly such a norm, with the assistant announcing itself at the start of the call. Until then, the label has to be earned from what is said, and the count has to include the calls where it could not be.

FAQ

Questions this note answers.

  1. 01Did Muse really use humans to make calls?

    Reporting in September 2026 by 404 Media and Tekedia described Meta testing Muse's calling feature through trained contractors who placed the call from the assistant's brief and reported the transcript and a summary back. Meta says the feature will launch only with proper disclosures.

  2. 02Why can't the phone network tell you who is calling?

    STIR/SHAKEN attestation vouches for the calling number — that the carrier believes the caller is entitled to use it. It says nothing about who is speaking or on whose behalf, and a contractor placing a call for an assistant will usually attest perfectly well.

  3. 03How does the receptionist label a human-for-agent call?

    After every call with speech, an analysis of the transcript returns a caller kind — human, AI agent, human for an agent, recording or unknown — plus whether the caller declared itself, whom it said it was calling for, which vendor it named, and up to three supporting quotes.

  4. 04Is there a precedent for disclosure rules on assistant calls?

    Yes. When Google demonstrated Duplex placing restaurant bookings in 2018, the backlash produced a disclosure norm: the assistant announces itself at the start of the call. Meta's statement that calling will ship only with proper disclosures follows the same pattern.