monday.com 推出 Harmony，押注專用模型打造語音 AI 銷售智能體

Sales teams can lose an interested customer in the gap between an inquiry and a return call. Harmony, a voice AI platform built by monday.com, is launching with a system designed to handle those conversations using a collection of smaller, task-specific language models.

Announced on October 19, the platform targets mid-market and enterprise sales teams in the United States, United Kingdom, Australia, and Singapore. It combines automated calling, qualification, follow-up, CRM updates, and transfers to human representatives. The technical bet is that specialized models and structured workflows can make sales calls faster and more controllable than relying on one general-purpose model.

Harmony reports approximately 200 milliseconds of audio-in to audio-out latency and contact rates of up to 70%. These are company claims, rather than results from an independent comparison. Their practical value will depend on the calling environment, the leads being contacted, and how the agent performs when a conversation departs from its expected path.

A collection of models for different parts of a call

According to the launch materials, Harmony uses proprietary models trained on customer conversations, with each assigned a specific function, such as qualifying a lead, capturing information, or handling an objection. A real-time predictive engine orchestrates their work. The compact models run largely on Harmony’s own infrastructure.

Specialization can make the system easier to optimize for a particular job. A model focused on extracting qualification details has a more bounded task than a general assistant expected to answer almost anything. But a collection of models also requires reliable coordination: information gathered in one part of a conversation must remain available and consistent in the next.

The latency claim is relevant because conversational pauses shape the experience of a phone call. Fast responses can reduce awkward turn-taking, but response time alone does not establish accuracy, useful interruption handling, or the ability to recover from a misunderstanding. The release does not disclose a full latency distribution, model parameter counts, or a reproducible testing methodology.

How the OneAI acquisition fits

Harmony incorporates voice technology acquired from OneAI. In its first-quarter 2026 results, monday.com disclosed an agreement to acquire OneAI to add voice-agent capabilities to its AI work platform. Harmony gives that strategy a specific commercial application: running customer conversations alongside the systems that store and manage sales work.

The platform is led by Vitali Margolin. His supplied biography says he co-founded monday CRM internally, helped grow it to $100 million in annual recurring revenue over three years, and later led product and engineering for monday.com’s new products and its OneAI acquisition. That background helps explain Harmony’s emphasis on the operational steps around a call, rather than speech generation alone.

The workflow around the voice

The launch release describes predictive and power dialing, branded caller ID, inbound routing, consent controls, and native CRM write-back, with telephony built on Twilio. Supported CRM integrations include HubSpot, Salesforce, Zoho, and Pipedrive. The agents can hand a qualified conversation to a live representative with a briefing.

Harmony’s platform overview adds an orchestration layer for tools and CRM, call analytics, and a process for inspecting calls, simulating changes, and testing alternatives before rollout. Its agent-builder page describes a visual canvas for call flows, branches, prompts, knowledge, and tools, alongside configurable variables and business rules.

Those controls are central to deployment. An illustrative lead-qualification agent needs a definition of a qualified prospect, a way to record the answer, and a clear transfer rule when the conversation needs human judgment. Building a plausible voice is only one part of making that workflow useful to the sales team.

What the developer documentation adds

Harmony’s public API documentation describes a narrower and more concrete integration surface than the broad launch language. Developers can create or update contacts, import batches of up to 1,000 contacts, optionally enroll contacts into workflows, retrieve call histories, and read transcripts.

The documentation says API keys act as the user who created them within that user’s account. It also states that building an integration for installation across many Harmony customers is not yet supported. That distinction matters for buyers planning a custom internal connection versus software vendors considering a distributed integration.

For a business connecting a CRM or data warehouse, the practical work includes maintaining contact identities, avoiding duplicate outreach, and deciding which changes should enroll someone into a calling workflow. API availability makes those connections possible; it does not remove the need to design them carefully.

Reading the early customer results

The launch includes a testimonial from Bright Car Finance, a UK car-finance broker. Director Richard Stubbs says the business generates around 2,500 approved applications a month, but had converted about 5% into agreements because leads cooled before representatives reached them. He reports that Harmony qualifies about half of its conversations, transfers 80% of those qualified conversations live to advisers, and helped increase deals closed by 9% in less than a year.

Those figures describe different stages of a funnel. Qualification is not a sale, and the reported 9% increase is not stated as a nine-percentage-point rise in conversion. The testimonial offers an early example of the intended workflow, but does not isolate Harmony’s contribution from other changes in the business.

Where human control still matters

Harmony’s website says agents introduce themselves as AI by default and offers controls for disclosure, consent capture, recording rules, transcripts, and audit trails. Its own FAQ acknowledges that compliance depends on how an agent is configured, rather than the technology alone.

The launch’s larger test will be whether teams can improve timely customer contact while keeping the resulting conversations accurate and appropriately bounded. That means measuring transfers, errors, corrections, and completed outcomes alongside speed. Harmony brings model specialization, telephony, workflow design, and CRM integration into one product; its usefulness will be determined by how well those pieces work together on real calls.