September 2026 Summaries

Vapi has launched Personal Agent Calling in beta, a service that lets AI personal assistants place phone calls on a user’s behalf to handle tasks requiring live conversations, such as checking product availability, scheduling appointments, or requesting airline upgrades. Users provide their assistant with the business, relevant details, and constraints, then authorize it through phone.vapi.ai; the assistant navigates phone menus, waits on hold, speaks with representatives, and reports the outcome through notifications or email. In a test case, Vapi’s Head of Revenue used an assistant to request a free Economy Plus aisle-seat upgrade, and the agent negotiated a waived $51.69 fee after a 14-minute hold. The service supports several assistants, including Claude, Cursor, Muse, OpenClaw, Hermes, Instinct, and GrokBot, and relies on Vapi’s existing enterprise voice infrastructure for telephony, speech recognition, language models, and text-to-speech. Vapi says the feature requires no API keys or additional charge beyond standard Vapi minutes.

Vapi for CX is presented as a voice-agent platform intended to shift customer experience operations away from traditional cost-focused metrics such as call containment and handle time toward resolving customer issues end to end. It argues that earlier contact-center technologies, including IVR, offshoring, self-service, and chatbots, reduced operating costs but did not improve satisfaction because they prioritized deflection rather than outcomes. The platform combines a natural-language agent builder for CX teams with APIs, CLI tools, and detailed configuration for engineers, while managing underlying telephony, interruption handling, scalability, fallbacks, and compliance. Its proposed lifecycle includes building configurable agents, deploying them alongside existing contact-center infrastructure, evaluating and monitoring calls through custom scorecards, and iteratively improving behavior through simulations, testing, and model comparisons. Vapi highlights functions such as specialized-agent routing, structured tool calling, warm transfers, versioning, compliance controls, and analytics, and cites customer examples including Amazon Ring and Kavak to support claims that greater control and continuous optimization can improve resolution, satisfaction, conversion, and revenue outcomes.

Vapi introduced three product releases aimed at helping builders select voice-agent models, test agents before deployment, and run outbound calling at scale. Model Intelligence provides presets and performance data for transcribers, LLMs, and voices, including cost, latency, transcription word error rate, and a proprietary voice Humanness Index, with metrics based on live-call performance. Simulations lets teams create customizable AI callers and test scenarios that generate pass/fail outcomes, transcripts, and recordings; it supports dashboard and API workflows, configurable LLM judges, and repeated test runs to account for non-deterministic results, while Vapi still recommends manual testing. Rebuilt Campaigns can call lists of up to 10,000 contacts with personalized calls, eligibility checks, scheduled dialing windows, and adjustable concurrency, though automatic concurrency management is planned rather than currently available. Vapi advises lower outbound concurrency to avoid poor answer-rate periods and recommends a fixed opening message to reduce first-turn latency.

TypeSafe AI’s early-access Jev “System One” model is being tested by Vapi as a fast, structured decision layer for voice AI, particularly to assess whether an assistant’s proposed response violates a defined rule before it is spoken. Unlike general-purpose LLMs that generate text, Jev returns focused classifications or scores quickly enough to operate alongside text-to-speech generation, providing an additional guardrail for sensitive uses such as medical or financial conversations. Early tests found that it caught many clear violations but also showed that concise, carefully worded rules and representative evaluation data are essential to limit false positives and missed issues. The approach is most effective for persistent, response-level rules where the response contains enough information for judgment, and it cannot independently verify tool use, external facts, compliance, or context-dependent rules without additional state. Vapi emphasizes that such models are not deterministic compliance guarantees and should supplement rather than replace prompts, testing, monitoring, privacy controls, and domain-specific safeguards. Potential future applications include voicemail detection and determining when a caller has finished speaking, reflecting a broader model in which general LLMs handle open-ended language, code handles exact rules, and fast decision models address bounded real-time judgments.

VapiCon 2026, scheduled for November 11–12 at Fort Mason in San Francisco, is a two-day voice AI summit organized by Vapi that will feature two parallel tracks for engineers and business or operations leaders, with more than 1,200 expected attendees. The event will focus on the growing production use of voice agents, including real-world performance, return on investment, reliability challenges, architecture choices, multilingual capabilities, enterprise data grounding, and deployments in areas such as healthcare, contact centers, and regulated industries. Confirmed speakers include leaders from Deepgram, Soniox, Gradium, Cartesia, Daily, Twilio, LangChain, Vercel, Coval, Arize AI, and Vapi, while additional customer and industry speakers are expected. Sessions will address speech-native versus cascaded systems, evaluation and observability, latency, fallback and human-handoff mechanisms, and the practical limitations revealed by large-scale calls involving interruptions, accents, noise, and changing user intent.

Voicemail detection, or answering machine detection, helps outbound voice agents distinguish between a live person and a recorded greeting within the first few seconds of a call, when audio characteristics alone are often insufficient because both may contain the same human voice. Effective systems combine timing and behavioral cues such as greeting length, cadence, post-greeting silence, ringing duration, keywords, beeps, call state, and machine-learning or transcript-based classification, while voice activity detection alone only identifies speech versus non-speech. The central operational tradeoff is between synchronous detection, which waits for a decision but can create noticeable dead air, and asynchronous detection, which allows an agent to begin speaking and then adjust if voicemail or a human pickup is identified. Accuracy should be evaluated through separate false-positive and false-negative rates, latency, beep timing, and performance across carriers and locales, since mobile voicemail, PBX systems, codecs, poor audio quality, and AI call screening can reduce reliability. Vapi presents its platform as a configurable, provider-agnostic approach that supports multiple detection providers, hybrid model and beep detection, continuous early-call monitoring, adjustable timing settings, and actions such as hanging up, leaving a message, or returning to conversation when a person answers.

Vapi has launched a limited private beta and public waitlist for GPT-Live, OpenAI’s full-duplex speech-to-speech model, with broader access planned in the coming months. Unlike traditional voice-agent architectures that separately use speech recognition, a language model, and text-to-speech, GPT-Live combines these functions in one model, enabling more natural interruptions, backchanneling, lower latency, richer interpretation of tone and emphasis, and improved handling of details such as names, numbers, and addresses. Vapi positions the model as particularly useful for human-sounding, high-stakes applications such as patient communication, identity verification, and customer support. Because Vapi’s platform is model-agnostic, beta users can deploy GPT-Live with their existing tools, phone numbers, testing systems, and observability stack without rebuilding agents, while retaining the option to use modular pipelines where independent component selection and tuning provide greater control.