Sierra sells phone-answering AI agents. As of today it also sells the thing that catches them calling you.
fleming-1, launched October 7, is a model that sits on the receiving end of a phone call and scores the caller's audio for signs it was generated by a machine. Not how the voice sounds to a human ear. The model looks past that, into artifacts of synthesis, while the call is still happening. When it decides the caller is likely an agent, it flags the call and leaves the decision to the business. A bank might add a verification step. A high-volume call center might just start counting.
The company says the question changed. It used to be who is calling. Now it is what is calling. Sierra ran into the shift itself: in 2025, its own agents started calling its own agents inside healthcare deployments, each side unaware of the other. With personal agents spreading, Sierra expects a large share of incoming business calls to come from software acting on behalf of consumers.
Read the two launches together and the strategy shows. On October 6, Sierra announced the Personal Agent Protocol, an open standard built with Meta and a roster of commerce and communications companies, for agents to identify themselves and name who they act for. On October 7 it launched the detector for the calls where that doesn't happen. The protocol is the polite path; fleming-1 is the audit trail. Sierra's own wording gives away the split: some agents, it said, may not respect the protocol.
The numbers, read carefully
fleming-1 is "tuned to be conservative by default," which is a polite way of saying it is designed to miss agents rather than mislabel people. Sierra published no accuracy figures, no false-positive rates, no benchmark. And the company admits the underlying job is getting harder: bad connections and background noise hide the clues while synthetic voices keep improving. A detector tuned conservatively against an improving adversary is a product with a built-in decay curve.
There is also a narrower scope than the announcement's tone suggests. fleming-1 "works with any voice agent built on Sierra." That is not a universal AI-call detector; it is a feature of Sierra's own stack, covering the calls its own ecosystem generates and receives. The Caller ID analogy Sierra reaches for, BellSouth's 1988 launch in Memphis, is instructive in the other direction too: Caller ID became universal infrastructure. fleming-1 is, for now, a vendor feature.
What this means in practice
If you run a call center, agent-to-agent calls are about to become a line item in your fraud and ops reviews, and "measure how often it happens" is Sierra's own recommended first step. If you are a consumer who hands chores to a personal agent, expect the other end to start treating your agent's calls as a distinct category, with verification steps that human callers don't get. The era of the phone call as a human-to-human channel ended quietly. The industry is now building the caller ID for whatever replaced it.
Sources
- [1] Unite.AI, "Sierra Launches Fleming-1 to Detect AI Agents Calling by Phone" (Oct 7, 2026)Read source