Decagon's Voice 3 and PACT Prepare Customer Support for Agents on the Other End

--- title: Decagon's Voice 3 and PACT Prepare Customer Support for Agents on the Other End meta_title: When the Customer Is an Agent, Support Has To Adapt meta_description: Decagon launched Voice 3 with its Chord speech model plus PACT, a protocol for delegated authorization when a person's AI agent contacts a business. ---

Most customer support software assumes a person is on the other end of the conversation. Decagon's October 1 release assumes that increasingly is not true, and builds for it in two directions at once.

The company introduced four products at its Decagon Dialogues event: Voice 3 with a new speech model called Chord, a Personal Agent Gateway, Agent Modules, and a Duet Apprentice beta.

Voice 3 and the Chord model

Voice 3 ships with Chord, the first voice model from Decagon Labs, post-trained specifically for customer conversations. Decagon says the model shapes speech phrase by phrase, slowing down for details like phone numbers and confirmation codes before returning to a conversational pace.

The architecture is duplex, with two layers running in parallel. A low-latency conversational model handles listening and speaking, while a more powerful model manages reasoning, tool calling, and guardrail enforcement behind the conversation. Because the agent processes incoming audio while speaking, it can acknowledge with a casual sound but yield when someone genuinely interrupts, rather than talking over them or going silent.

Decagon reported that in blind tests across three pairs of recordings, each pairing a real person with the same voice run through Chord, on average about 90% of listeners could not identify the human. The company says Chord was trained on licensed data and consented voice talent, never on customer-owned data, and supports more than 70 languages with automatic detection and mid-sentence switching.

The 90% figure is a company-reported blind test, and it cuts both ways. It is impressive as a quality signal and uncomfortable as a disclosure signal, since a caller who cannot tell they are speaking to a machine has no opportunity to ask.

The Personal Agent Gateway is the more consequential idea

Decagon's announcement notes that in the month before the event, Meta launched Muse, OpenAI introduced dots, and Instinct started placing phone calls for users. The company describes these as personal agents that book, buy, cancel, and negotiate on their owners' behalf, and says they are already contacting customer support.

That is the shift the gateway is designed for. It has two parts.

Personal agent detection flags likely personal agents across chat and voice, using signals from both the business and the platform. Each business decides what happens next and what those agents can access.

A dedicated personal agent channel sits alongside chat, email, and voice, with separate Agent Operating Procedures. So the same request can follow a different workflow depending on whether a person or an agent is asking.

Permissions live inside those procedures. Businesses define the scopes an agent can request and mark which sections require each one, and those sections are never exposed to an agent lacking the required scope.

Decagon's example is an airline. A traveler's personal agent requests permission to view and rebook flights; the traveler approves only viewing; the airline's agent shares earlier flight options but cannot make the change until rebooking is authorized separately.

PACT and why authorization needs a protocol

Decagon also introduced PACT, for Personal Agent Consent and Trust, a protocol that lets a person's agent act on their behalf under permissions the service defines and the person grants.

PACT builds on the Agent2Agent protocol, which covers how agents find each other and exchange messages, and adds delegated authorization built on OAuth 2.0. That lets an agent prove which person it represents and what that person allowed it to do.

The reasoning behind this is straightforward once stated. If a business receives a request from an agent, it needs to answer three questions before acting: which human is behind the agent, what did that human authorize, and can the agent prove it. Agent2Agent alone covers discovery and messaging but leaves authorization vague. Bolting OAuth 2.0 onto it gives the delegation a verifiable chain.

The specification is available, and Decagon says it is working with personal agent providers and enterprise customers to shape it. Whether PACT becomes a standard or stays one vendor's approach depends on whether other platforms adopt it, which is the usual test for a protocol with a single author.

Agent Modules and the apprentice model

Agent Modules combine infrastructure, analytics, and campaign orchestration for customer journeys beyond support, including lead qualification, onboarding, and collections, across financial services, travel, hospitality, healthcare, retail, and telecom.

The customer's team defines the business logic while Duet draws on the company's standard operating procedures, policies, integrations, and knowledge sources to translate processes into agent behavior. In collections, for instance, teams set disclosure language, approved payment plan tiers, and conditions that require a human handoff. Before launch, simulations test agent behavior against personas with varying intents.

Duet Apprentice is in beta, and the framing matters. It is positioned as an apprentice rather than an autonomous operator, which suggests a staged trust model: the agent learns the business's procedures with oversight before handling cases alone. That is a more honest description of how enterprise deployments actually progress than the usual launch language.

What this says about the direction of support

Two things stand out.

The first is that voice quality is no longer the differentiator. A speech model indistinguishable from a person in 90% of blind tests is a solved problem in the sense that matters, which means competition moves to what the agent knows, what it is allowed to do, and how it proves it. Christian Niedworok, who leads digital service communication at Deutsche Telekom, described Voice 3 in the announcement as sounding like it is actually listening and keeping the conversation moving instead of going quiet while it works. That is a comment about latency and turn-taking rather than timbre, which is where the remaining quality gap lives.

The second is that the counterparty is changing. Support systems have spent decades modeling human behavior: what confuses a person, what makes them hang up, how to route them. Now some fraction of incoming requests will be machines acting for people, and the system has to detect them, authenticate them, and scope their access. That is a different engineering program, and Decagon is the first major vendor to ship products aimed squarely at it.

The detection problem has a particular shape worth considering. A personal agent contacting support on someone's behalf is not necessarily malicious. It may be a legitimate delegation the customer authorized. Treating agent traffic as abuse would break a use case that is about to become common. Treating it as ordinary traffic opens the door to automated abuse at a scale human fraud never reached. The gateway approach, where detection informs routing rather than triggering a block, is a reasonable middle path, and it depends on detection being accurate.

What to watch

Whether PACT gets adopted beyond Decagon's own customers. If it does, delegated authorization becomes table stakes for any business that receives agent traffic. If it does not, each platform ends up with its own answer, and the parties actually harmed by the confusion will be the humans waiting behind their agents for an answer.

Whether detection accuracy holds up under adversarial pressure. Any signal that classifies traffic invites gaming, and personal agent detection is a signal worth gaming for anyone trying to automate a support workflow at scale.

Whether the multi-language claim survives layering. Voice 3 supports more than 70 languages with automatic detection and mid-sentence switching, and Decagon says every language is validated by native speakers before shipping. Switching languages mid-sentence while maintaining a consistent identity is a harder problem than switching them at a boundary, and independent testing would be useful.

The direction of this release is clear, and it is broader than one vendor's product line. Customer support assumed for decades that it was talking to people. That assumption is expiring, and the companies that ship the tooling for the transition will shape how the next few years of automated customer interaction look.