AethelforgeSTAGING
Read the Index

The layers

Four kinds of arrival.

Inbound to a business is not bot or human. It is a reading layer, an acting layer, a human-proxy layer and an adversarial layer — and one label that every channel shares.

Every channel a business runs can already see crawlers and the people they send. What none of them carried was a label for the arrivals in between.

An assistant that calls five clinics for one patient is not a robocall, and it is not a person. A contractor placing that call from the assistant’s script is a person, but the intent belongs to the software. A crawler re-indexing a site the day after a launch creates no work at all; the same company’s assistant filling a quote form creates a follow-up.

The four layers give each of those arrivals a name, so the count means something and the routing can differ.

In one paragraph

What are the four layers of inbound?

The reading layer is crawlers and answer-engine fetchers that read a site. The acting layer is AI agents that call, chat or submit for a person. The human-proxy layer is a live person working from an agent’s brief. The adversarial layer is robocalls, script bots and spam. People are the baseline every layer is measured against.

Layer 01

The reading layer

Crawlers and answer-engine fetchers. They read a site so a model can be trained on it or so an assistant can quote it back to a person.

Looks like

A training crawler re-indexing every page the day after a product launch. An assistant fetching one page because someone just asked it a question.

The signal we read

Declared user agents, the crawler category behind each one, and the share of all bot traffic that is AI rather than search or SEO.

At the door
  • WEBREADINGMETA-EXTERNALAGENT · TRAINING CRAWLER
  • WEBREADINGCHATGPT-USER · FETCH FOR ONE QUESTION
On each channel

Web: the declared crawler or fetcher user agent, and its Web Bot Auth signature when it signs. Forms: a person who arrived from an AI answer. Phone and chat: not applicable — nothing reads a phone line.

Layer 02

The acting layer

AI agents that call, chat, book, quote or submit on a person’s behalf. Declared when the agent says so or signs its requests; undeclared when it arrives looking like a browser or a caller.

Looks like

An assistant calling five clinics to compare prices for one patient. An agent submitting a quote form with a real person’s details and a schedule it negotiated itself.

The signal we read

An agent’s Web Bot Auth signature on the request, a self-declaration in the conversation, or post-contact analysis of who was actually speaking and for whom.

At the door
  • PHONEAGENTVOICE · UNDECLARED · PRICE CHECK
  • FORMSAGENTQUOTE REQUEST · UNDECLARED
  • WEBAGENTSIGNATURE-AGENT · SIGNED
On each channel

Web: an agent’s request signed under Web Bot Auth; a crawler that signs is declared, but it is still reading. Phone: a caller that says it is an assistant, or a post-call analysis that concludes it was. Chat and forms: a self-declaration in the text or a signed request behind it.

Layer 03

The human-proxy layer

A live person acting because of someone else’s AI assistant. The caller is human; the brief, the intent and the follow-up belong to an agent.

Looks like

A contractor placing a call from a script an assistant wrote, then reporting the transcript back to the assistant that sent them.

The signal we read

Post-call analysis that labels the principal, the intermediary and the intent — not just the voice on the line.

At the door
  • PHONEHUMAN PROXYPERSON ON AN AGENT BRIEF · SCRIPTED
On each channel

Phone, mostly. A human voice, a scripted brief, a transcript that gets reported back to the assistant that sent the caller. The label names the principal and the intermediary separately.

Layer 04

The adversarial layer

Robocalls, voicemail drops, script bots and form spam. The noise that already dominates the phone and a third of every contact form, and that an agent lane has to be separated from.

Looks like

Sixty robocalls for every five real callers. A quote form filled by a script that never read the page.

The signal we read

Call attestation, number and IP reputation, spam scoring and behavioural tempo — the gates a business already runs, kept beside the agent lane rather than replaced by it.

At the door
  • PHONEAUTOMATEDROBOCALL · ATTEST C
  • FORMSAUTOMATEDSPAM · SCRIPTED FILL
On each channel

Phone: attestation level, number reputation, dead air and agent-only transcripts. Forms: the spam score. Chat: automation user agents and tempo. Web: unidentified bots, kept out of the reading count.

The label

One field. Every channel.

The phone, the web, the chat widget and the forms all write the same field with the same five values, so the channels add up to one index instead of four incompatible reports.

Companion fields carry the evidence: whether the contact declared itself, who it said it was acting for, which vendor it named, and the quotes that support the verdict.

How the Index is built
inbound_kindMeaning
humanA person speaking or typing for themselves or their organisation.
ai_agentAn AI assistant or agent acting for a person or business — declared, or evident from behaviour.
human_for_agentA live person acting because of someone else’s AI assistant or automated request.
automatedA robocall, IVR, voicemail drop, script bot or form spam — the adversarial layer.
unknownToo little to tell. Dead air, a one-word chat, a form with no text.
Companion fieldCarries
agent_declaredWhether the contact said it was an AI, an assistant or an agent.
on_behalf_ofThe principal, as stated by the contact.
agent_vendormuse · chatgpt · gemini · claude · copilot · alexa · siri · other · unknown
agent_evidenceUp to three short quotes or observations that support the label.
channelphone · web · chat · form

Rules

What the label never does.

  1. Never default to human. Missing evidence is unknown.
  2. Label the principal, the intermediary and the intent — not the voice.
  3. Keep the agent lane beside the spam gates, never in place of them.
  4. Publish aggregates only. No names, no numbers, no transcripts.

FAQ

The layers, answered.

  1. 01Why not just label everything bot or human?

    Because the interesting cases sit in between. A person calling from an assistant’s script is a human voice with an agent’s intent. An assistant fetching one page for a person is a bot doing a human errand. A two-value label hides exactly the arrivals a business needs to route differently.

  2. 02What is the difference between the reading layer and the acting layer?

    Reading changes nothing at the business — a crawler indexes a page, an assistant quotes it. Acting creates work: a call to answer, a chat to staff, a form to follow up. The Index keeps them apart because a surge in reading costs bandwidth, while a surge in acting costs people.

  3. 03Is the adversarial layer the same as bot traffic?

    No. Adversarial means robocalls, voicemail drops, script bots and form spam — contacts nobody wants. Most bot traffic on the web is the reading layer, which a business generally does want. The two are labelled separately so a spam gate never swallows a legitimate agent, and an agent lane never lets spam through.

  4. 04What happens when the evidence is thin?

    The contact is labelled unknown. Dead air on a call, a one-word chat, a form with no message. Unknown is a real value in the Index, reported as such, and never rolled into the human count.

Research

Research on the layers.

All notes on the layers
  1. Layers24 Sep 2026The third kind of caller: a person on the line, an agent on the briefMuse'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.4 min