Decagon's Alto tackles timing in duplex customer-service calls

By James Calloway

Decagon said Alto lifted the reported resolution rate—the share of calls counted as resolved—by more than five percentage points on live customer calls while delivering average responses in about 700 milliseconds. The company introduced Alto, a Decagon Labs model, on November 11, 2026.

Alto controls the conversational layer of Decagon’s duplex voice agent, an architecture that speaks with a caller while another model works in parallel. It decides whether an immediate answer is justified and what to say, extending the company’s model specialization beyond speech quality into conversational control.

That distinction matters in customer service because speed can create a reliability error. Decagon’s example is an agent telling a caller that a reservation has changed before the underlying system has confirmed the change.

The claim carries two stakes: voice quality concerns how calls sound, while timing concerns whether the agent accurately reports a completed action. Decagon attributes the measured improvement to the second problem, not a broader change in the agent.

Decagon said the rest of the agent’s logic remained effectively unchanged during the resolution comparison. The announcement did not identify the customers, sample size, baseline rates, or measurement period, limiting what buyers or independent researchers can infer from the result.

What problem does Alto solve?

A conventional cascade voice pipeline handles each turn in sequence: transcribe the caller, reason about a response, then generate speech. Decagon’s duplex system instead runs a fast conversational model in parallel with a more powerful reasoning model.

The conversational model listens and speaks. The reasoning model handles complex decisions and actions behind the conversation, allowing the agent to keep interacting while work continues.

That parallelism creates a judgment call on every turn. Waiting for the reasoning model can disrupt the flow, while answering too soon can state something the backend has not confirmed.

Decagon frames the central choice as the answer-or-wait decision. The speed advantage of a smaller conversational model therefore comes with a reliability problem that duplex architecture alone does not solve.

According to Decagon, Alto continuously tracks what the caller has said and what the rest of the agent is doing. It grounds its responses in a shared call record and waits for confirmation before announcing that an action is complete.

That shared state is the mechanism behind Decagon’s claimed combination of speed and reliability. It gives the conversational model a way to judge safety from information the agent has already confirmed.

During a lookup or tool call, meaning a backend search or action, Alto can request the next piece of information or check in with the caller. Decagon says it avoids generic filler while the agent waits.

If a result arrives as the caller asks something new, Alto can answer the question and supply the result together. That behavior aims to preserve momentum without mixing an unconfirmed answer with the completed action.

Takeaway: Alto treats conversational timing as a separate control problem, not merely a speech-quality feature.

How did Decagon train Alto?

Decagon started with a small open-weight model, whose trained parameters were available for further use, to keep responses fast. The company then fine-tuned, or further specialized, it on simulated calls informed by production experience.

During reinforcement learning, a method that scores candidate behavior to favor better decisions, Alto generated several possible responses at each conversation point. Decagon’s scoring process first prioritized whether Alto made the correct answer-or-wait choice.

The process then assessed whether its wording followed business policies. It also checked whether the response reflected confirmed information rather than an assumption.

Scoring in that order reflects the architecture’s main risk. A well-worded response can still be wrong if it arrives before confirmation, while a correct wait can still fail business policy if phrased improperly.

Decagon concentrated training on ambiguous cases after finding that Alto already handled most situations correctly. That result narrowed the hardest part of the task to uncertain moments rather than routine exchanges.

Training also varied policies and tools across simulated businesses. Decagon aimed to teach transferable judgment instead of a script memorized for one company.

Takeaway: The training strategy targeted uncertain turns and cross-company variation, though Decagon disclosed no separate performance breakdown for either.

How did Decagon test Alto before deployment?

Before any caller heard the model, Decagon reported testing Alto across more than 90 million simulated calls. That scale tested a broad set of generated conversations before live exposure.

Decagon then ran Alto in shadow mode, meaning it evaluated the model against live production traffic without speaking to callers. This allowed the company to inspect responses before they affected a real interaction.

The company subsequently reported the live-call result: a resolution-rate gain of more than five percentage points and average responses of about 700 milliseconds. Decagon’s announcement still provides no sample size, customer identities, baseline, or test duration for that phase.

The sequence suggests Decagon used simulation first, shadow traffic second, and live calls last. It does not establish how representative the simulated calls or shadow traffic were of the customers included in the reported comparison.

Neither simulation nor shadow mode proves live performance. Simulation can expose rare combinations, and shadow traffic can reveal differences between generated and real calls, but only the reported production comparison measures customer-call resolution.

Takeaway: Decagon describes a three-stage evaluation, but the disclosed evidence supports only the company’s own attribution of the final gain.

How do Alto, Voice 3 and Chord fit together?

Alto follows Voice 3 and Chord, both announced on October 1, 2026. Voice 3 combined Chord’s speech generation with the duplex architecture and support for more than 70 languages.

Chord shapes how Decagon’s agents sound, while Alto shapes how they conduct a conversation. Decagon previously credited Chord with an independent resolution increase of up to seven percentage points.

Together, the models divide voice performance into two training targets: speech naturalness and conversational judgment. Alto’s reported result links the second target to resolved customer calls, while Voice 3 and Chord address the first.

Both disclosed gains use resolution as the outcome. The figures do not measure speech naturalness directly, even though the models target that quality separately.

The architecture still hinges on the answer-or-wait decision. Alto must use the conversational model’s speed without letting a natural-sounding exchange get ahead of confirmed actions.

Takeaway: Decagon now separates voice realism from conversational control, but only the latter has the newly disclosed gain of more than five percentage points.

What can buyers conclude?

Alto’s mechanism is plausible within Decagon’s architecture: a fast model can maintain the conversation while a slower reasoning model completes complex decisions and actions. The claimed gain matters because an agent that speaks before confirmation can turn a quick response into an operational error.

Decagon’s 700-millisecond figure describes average responses, not an explicitly reported end-to-end task-completion time. Because the reasoning model can continue working in parallel, buyers should not treat response latency as total resolution time.

The announcement does not say whether the result held across every language, customer, or call type. It also omits the baseline against which five percentage points were measured, so the practical size of the gain remains uncertain.

For buyers evaluating customer-service agents, the relevant risk is not only whether an agent sounds natural. It is whether the agent confirms a reservation or another backend action before reporting completion, precisely the failure mode Alto was designed to control.

Until Decagon supplies the missing live-test details, buyers should treat the result as a company-reported outcome rather than an independently verifiable finding.

Takeaway: Customer-level live-test data would determine whether Alto’s timing gain holds beyond Decagon’s own evaluation.