ASAPP's Consumer Agent Index, Vol. 1, measured consumer agents contacting brands by chat. Voice agents, including those from Muse and Grok, now reach support teams too. We already detect consumer agents in chat, and we are extending the same approach to voice. This poåst covers how it works and what we do with a detection.
Our position follows the Index: consumer agent traffic is a new channel to measure and set policy for. Blocking it by default is the wrong call. Customers send agents to handle tasks they would otherwise do themselves, such as changing a booking or negotiating a bill, though agents ask differently from people: they state exact limits and often hold off committing until their person approves. So our design records what we see and how confident we are, and each brand’s policy decides the response.
Extending detection from chat to voice
Chat detection works because a web request carries data an edge rule can inspect: the network it came from, the browser identity it presents, and a provider's classification of known AI assistants. A voice agent can reach a contact center through traditional telephony, through internet voice, or through an app or browser session. ASAPP's voice detection is designed to work across all of them, combining the transcript, the caller identifier, connection metadata, and the audio.
Three sources of evidence for voice
We are rolling out voice detection in phases. Each source adds to the picture, and none decides alone:
- The words. The voice transcript is checked in real time for language that signals a consumer agent. In the chat conversations the Index analyzed, fewer than 3 in 10 agents said “AI,” while two in three said they were acting for someone else and more than half called themselves an “assistant.” We expect voice agents to behave similarly. Scanning for one word would miss most agents, so recognition covers: acting-for-someone language, assistant labels, identity disclaimers, and agent names. The same logic runs on chat, so both channels share one definition of AI self-identification.
- The connection. Where a session provides them, the caller identifier (for a number, its line type and country code) and connection metadata help identify sessions that look unusual for a human consumer. We build the rule in three steps.
- First, we analyze existing traffic offline to see how many callers are unknown and how line type and country code are distributed among each brand's real customers.
- Second, if a combination is rare among real customers, we mark it as suspect.
- Third, we add the rule to the live voice path and record each match as an event, so it can be tracked.
A suspect session is not blocked. The conversational AI handles it with extra care and asks a confirmation question, such as "Are you a person or an automated assistant, and who are you representing?" Many real customers connect from internet-based numbers, so this signal never decides the outcome on its own.
- The sound. Synthetic speech detection runs on the audio and produces a likelihood that a voice is machine-generated. That score goes into the same decision record as the other sources, with its own confidence and the threshold it had to meet. It tells us a voice is synthetic, not why the agent is there. A synthetic voice may belong to a legitimate consumer agent, so the result goes to policy and does not trigger an automatic block. We are piloting specialist synthetic speech detectors on our own voice traffic.
How a decision is recorded
ASAPP saves each signal as its own record. It saves the decision about whether a consumer agent is present separately, linked to the signals behind it. Each decision includes a confidence score, the threshold it had to meet, and the agent's identity when known. A new decision is saved only when the answer changes, so brands can see when an agent joined a conversation or when a person took over. Handoffs like this might not be that common for all transactions, but the record captures them. Again, no single weak signal, such as an unusual number, is enough to label someone as an agent.
What chat looks like today
In chat, a rule at the CDN tags requests that come from certain cloud and content-delivery networks, present a browser identity seen from agent-run browsers, or match the provider's AI-assistant classification. The application stores the result with the session and flags each message and the conversation. This is live in production for select customers. Before using live traffic, we tested it by sending real assistants to chat, including desktop assistants and Muse, and checking that detection fires. Self-identification in the conversation is the next signal.
What we do with a detection
A detection only matters if it changes what happens next. For each session, ASAPP is designed to compute a set of permissions from the detection result, whether the agent identified itself, whether the customer is authenticated, and the channel. A detected agent might be allowed to look up status and policy information but not change an account, or be routed to a person for anything beyond that. This applies the Index's advice to decide policy deliberately and apply it early, before an agent reaches a representative.
The Index recommended that brands verify what a customer has authorized an agent to do. Most agents are ready for this: 87% said they were authorized to act, and 42% checked back with their person during the chat. In voice, a detected agent can be asked to confirm what its customer approved before it makes a change.
Where this is headed
Over time, agents should get their own way in, with their own protocols, on a dedicated connection. Today’s voice systems and chat widgets are built for people, but agents arrive with structured requests: the Index found a median opening message of 366 characters, compared with about 35 for a typical customer. In the future, an agent can use a direct, structured connection instead of talking to a voice bot, and the business spends less to serve it. An agent on a dedicated connection identifies itself, so there is no guesswork. Detection is different. It estimates how likely a visitor is to be an agent, and it covers the agents that still arrive through channels built for people.
As agents change, and our detection evolves, ASAPP can add new signals without redoing how decisions are made, because each signal is recorded on its own. We will share what we learn as voice detection rolls out, alongside the Consumer Agent Index.


