September 2026 Summaries

ElevenLabs has launched operations in Belgium by opening a Brussels office, appointing Falke Van Onacker to lead its local presence, and planning to triple its Belgian workforce during the year. The company views Belgium’s multilingual environment, including its three official languages and extensive linguistic diversity in Brussels, as well suited to its AI voice, transcription, translation, and conversational-agent technologies. Belgian customers are using its products for multilingual applicant screening, medical documentation and patient-call transcription, audio publishing, and music-production tools, while the European Union has used the platform to dub diplomatic press conferences. ElevenLabs also reports that around 20 Belgian organizations hold grants for projects in education, arts, and accessibility, reflecting its stated aim to support the local ecosystem alongside commercial clients.

ElevenLabs announced a $300 million employee tender offer valuing the company at $22 billion, twice its February Series D valuation, with Wellington and T. Rowe Price leading the investment alongside several new and existing backers. Founded in 2022 around text-to-speech technology, the company has since expanded into an AI interaction platform, launching Eleven v4 and v4 Turbo voice models and supporting conversational agents through its ElevenAgents product. Enterprise customers now account for 55% of revenue, and the company says its technology is used by major technology, insurance, and telecommunications firms as well as governments. ElevenLabs reports that its agents handle more than 15 million weekly conversations across customer support, sales, healthcare, financial services, telecom, retail, and public services, while its agent business and annual recurring revenue have each more than tripled since February 2026. The company attributes its growth to an integrated approach spanning voice, listening, translation, product development, and customized deployment, and says it employs more than 800 people.

ElevenLabs’ Eleven v4 text-to-speech model uses bracketed natural-language Audio Tags to turn scripts into directed voice performances, allowing users to control emotion, delivery volume, pacing, reactions, accents, character voices, and sound effects inline. Tags such as [excited], [whispers], [long pause], or [laughs] affect the following speech and can be combined or customized with descriptive prompts like [whispering, fearful] or [a tired detective who has heard it all before]. The platform recommends pairing tags with punctuation, giving the model sufficient contextual scene detail, using distinct tags when a mood changes, and experimenting with neighboring emotional descriptions to refine results. Audio Tags work in Eleven v4, Eleven v4 Turbo, and Eleven v3 through the Text to Speech app and API, with prior v3 scripts remaining compatible, while the model supports more than 17,500 voices and positions itself as a highly ranked system for natural-sounding speech.

ElevenLabs announced that ElevenAgents is a launch partner in OpenAI’s new Marketplace for enterprise AI tools, allowing eligible OpenAI enterprise customers to register interest and eventually apply their OpenAI commitments toward purchases. ElevenAgents combines OpenAI’s reasoning models with ElevenLabs’ conversational AI platform, including real-time voice capabilities, integrations, knowledge bases, guardrails, and analytics, to support agents across voice, chat, and email in more than 90 languages. The company says organizations such as Stripe, Deutsche Telekom, and Klarna use the platform for customer support, lead qualification, cart recovery, and scheduling, while GPT models have performed well for complex reasoning and tool use in enterprise deployments.

NOS, a major Portuguese telecommunications and media company serving 1.5 million customers, is testing conversational AI voice agents powered by ElevenLabs Text to Speech to replace traditional IVR menus and provide immediate, 24/7 customer support in European Portuguese. The vendor-agnostic system integrates speech-to-text, language-model, and text-to-speech components through an open architecture, with ElevenLabs selected for its speech quality, pronunciation accuracy, expressiveness, and low latency. In an initial production rollout covering 3–5% of customers, the agents have handled about half of calls without follow-up contact, reduced time spent in automated service before human transfers by 15%, and produced double-digit increases in customer satisfaction scores. NOS is refining terminology pronunciation and exploring voice cloning while continuing to measure outcomes, improve handoffs to human agents, and eventually expand both customer-facing and internal voice automation.

ElevenLabs has introduced Eleven v4 and its low-latency counterpart, Eleven v4 Turbo, text-to-speech models designed to produce more emotionally expressive, context-aware, and natural-sounding speech while preserving speaker identity. The company states that the models improve tone, pacing, multi-speaker dialogue, audio fidelity, natural-language delivery controls, inline sound and emotion tags, and IPA-based pronunciation guidance. Eleven v4 Turbo targets real-time voice-agent applications with reported median inference latency near 100 milliseconds and time to first speech of about 150 milliseconds, and it is optimized for use with the ElevenAgents platform. Both models support more than 90 languages, with stronger cross-language accent adaptation and voice consistency for dubbing and localization, while updated voice-cloning capabilities can create high-fidelity instant clones from as little as 10 seconds of audio and add Professional Voice Clone support. ElevenLabs reports leading benchmark and listener-preference results, and the models are available through ElevenAgents, ElevenCreative, and the ElevenAPI for uses including audiobooks, character performances, voiceovers, conversational agents, and localization.

ElevenLabs has introduced Eleven v4 Turbo, a low-latency version of its Eleven v4 text-to-speech model, for use in ElevenAgents customer-facing voice agents. The model is designed to deliver expressive, context-aware speech that adapts tone, pacing, and emotion during live conversations, with approximately 100 ms median inference latency and reliability for longer calls. Supporting more than 90 languages, including improved Japanese, Brazilian Portuguese, Mandarin, and Cantonese, it can accommodate mid-call language changes while maintaining voice consistency or using native-optimized voices. Eleven v4 Turbo integrates with the company’s transcription and turn-taking models to enable faster responses and smoother interruptions across applications such as healthcare, retail, finance, government, and telecommunications. ElevenLabs says Eleven v4 ranked first on Artificial Analysis’s Provider Voice Arena leaderboard and was preferred by roughly 75% of listeners in blind comparisons with several competing models.

ElevenLabs presents Eleven v4 as an emotive text-to-speech model designed to turn written scripts into expressive audio performances using inline Audio Tags, punctuation, and contextual writing. Users can direct delivery with bracketed cues such as [furious], [sarcastic], [crying], or sound-effect descriptions, enabling changes in tone, pacing, emphasis, nonverbal sounds, and even emotion within a sentence. Examples demonstrate how the same dialogue can be interpreted as angry, excited, sarcastic, tearful, or worried, while a tagged horror scene adds effects such as creaking gates, footsteps, whispers, and laughter. The platform recommends iterating on tags rather than rewriting lines, generating longer passages for stronger contextual performances, and generally using one emotional direction per phrase to avoid conflicting instructions. Eleven v4 supports generations of up to 10,000 characters and is positioned for uses including advertising, storytelling, audiobooks, and multi-character audio scenes.

ElevenLabs has launched APIs for image generation, video generation, and reusable creative templates built in its Flows canvas, enabling teams to create visual assets through models from providers including Google, OpenAI, ByteDance, and Creatify. Users can generate, edit, and animate images into video at scale, while webhooks support production pipelines. Flows templates let creative teams define prompts, brand references, audio, output formats, and compositions once, then automatically generate new assets by supplying inputs such as a product image and name. The system is aimed at enterprise catalog and marketing workflows, allowing new products added to PIM, DAM, or storefront systems to trigger creation of pack shots, channel-specific advertisements, and product videos. These capabilities are available through the API and ElevenLabs MCP integration for tools such as Claude, ChatGPT, and Cursor, with the APIs available immediately.

ElevenLabs has launched ElevenLabs for Students, offering eligible university students aged 18 and older in the US, Canada, EU27, Australia, and the UK three months of free access to ElevenCreative, ElevenAgents, and ElevenAPI, plus one year of ElevenReader Ultra. The initiative is intended to expand access to professional AI voice, audio, agent, and development tools for creative projects, research, applications, and entrepreneurship. The company highlights student projects including a healthcare referral system, a cockpit communications safety agent, a misinformation-detection voice agent, and an interactive study assistant, while noting voice technology’s potential to support learners with dyslexia, visual impairments, and varied learning preferences. ElevenLabs also cites its Impact Program x Professors, through which more than 350 educators in nearly 50 countries have provided tools to over 1,500 students, and positions the new plan as support for students’ own ideas rather than a replacement for their creativity or critical thinking.

Modern AI voice agents combine real-time speech recognition, reasoning, live system access, and natural-language conversation to authenticate callers, resolve requests, and route complex cases without relying on traditional phone menus or queues. The guide describes ElevenAgents’ approach to maintaining in-call context and cross-call history through CRM integrations, webhooks, and dynamic variables; detecting caller frustration through sentiment analysis; using configurable de-escalation and escalation paths to humans or specialized agents; and grounding responses in company knowledge through retrieval-augmented generation from documents and web content. Agents can personalize recommendations by combining customer history, live conversation details, and current product data accessed through external tools. Privacy controls include configurable transcript and audio retention, redaction, zero-retention options, and HIPAA-eligible enterprise deployments when appropriately configured. Customer-satisfaction gains may result from faster, more consistent service and reduced wait times, although organizations should evaluate satisfaction alongside containment, handling time, and first-call resolution to ensure automation does not trap customers in unresolved interactions.

Zoom meetings can be transcribed through built-in live captions, cloud-recording transcripts, and AI Companion features, although many storage, download, and summary capabilities require paid Zoom Workplace plans. Transcripts can provide searchable, timestamped, speaker-labeled records that support accessibility, meeting reporting, knowledge sharing, training, and follow-up for absent participants. For post-meeting use, Zoom can generate downloadable VTT files from cloud recordings, while third-party platforms such as ElevenLabs offer batch transcription in more languages and formats, with features including word-level timestamps and speaker attribution. Developers needing lower-latency or more customized workflows can use APIs and WebSocket-based client-side or server-side streaming to route live meeting audio into applications, CRMs, or knowledge bases. The choice between native and external tools depends on budget, language coverage, accuracy, privacy requirements, automation needs, and whether transcription is required during or after a meeting.

A Python speech recognition tutorial explains how to build commercial-oriented transcription applications with ElevenLabs’ ElevenAPI, using Scribe v2 for batch audio processing and Scribe v2 Realtime for low-latency streaming through WebSockets. It outlines setup steps including creating an API key, storing it securely in environment variables, installing the ElevenLabs SDK and python-dotenv, and sending audio data to the API for transcription. The tutorial emphasizes enterprise-focused capabilities such as speaker diarization, word- or character-level timestamps, multilingual transcription, audio-event tagging, and keyterm prompting for specialized vocabulary. It also distinguishes batch transcription for prerecorded files from real-time transcription for applications such as meeting assistants, accessibility tools, and voice automation, noting that real-time results include partial and finalized transcripts. Additionally, it describes an optional ElevenLabs skill for AI coding assistants that can generate API integration code from natural-language requests, and notes that a free plan provides starter access to the Speech to Text API.

Human-in-the-loop (HITL) AI structures workflows so automated systems handle predictable, routine tasks while people review, approve, or take over decisions involving risk, ambiguity, regulation, specialized expertise, or customer sensitivity. It can also refer to human participation in machine-learning development through data labeling, output evaluation, corrections, and feedback. By combining automation with human judgment, HITL aims to improve accuracy, compliance, efficiency, and accountability in sectors such as banking, insurance, healthcare, fintech, and creative production. Effective systems rely on clearly defined handoff triggers, complete transfer of context between AI and employees, proportionate approval requirements, and regularly updated rules based on operational experience. The text distinguishes HITL, where human participation is required at designated checkpoints, from human-on-the-loop oversight, where AI operates independently while people monitor and intervene when necessary.

ElevenLabs has released Scribe v2 Medical, a generally available speech-to-text model fine-tuned for clinical audio and accessible through its API. Designed to improve transcription of medications, dosages, medical terminology, and multilingual clinical dictation, the model reportedly reduces clinical-audio word error rates by about 35% versus the base Scribe v2 model and achieved leading results in evaluations on Corti’s MedDictate and MedTerm benchmarks as well as the Eka Medical ASR dataset. The company reports particularly strong accuracy for annotated medical terms, though isolated drug names without contextual speech remain difficult for all evaluated systems; keyterm prompting can help bias transcription toward specified terms. ElevenLabs says the specialized model retains comparable performance on general non-medical speech, has already been adopted by clinical customers, and is HIPAA-eligible for enterprise users with a Business Associate Agreement and Zero Retention Mode, which deletes audio and text after API requests complete.

Studio 4.0 is a major update to ElevenCreative’s AI-native video editor that combines video, image, voiceover, music, sound-effect generation, editing, captioning, and exporting within a single project timeline. Users can create or refine assets directly in place, avoiding the need to move files among separate production tools, while the rebuilt timeline adds frame-level editing, clip snapping, synchronized layers, and editable captions. Its integrated Studio Agent can draft an initial edit, generate voiceovers, place clips, and synchronize effects based on frame-by-frame analysis, while allowing users to switch between AI-assisted and manual editing. The release also introduces clip-based collaboration tools and a redesigned playback engine intended to maintain responsive previewing and navigation in large, multilayered projects, targeting creators and marketers producing social content, product videos, explainers, and generated cinematic sequences.

Agentic workflows use AI agents to interpret goals, retrieve relevant information, decide on actions, adapt to changing circumstances, and execute multi-step tasks at runtime rather than following fixed, pre-scripted paths. They combine large language model reasoning with integrations, retrieval-augmented generation, memory, guardrails, deterministic approval steps, and human handoffs to handle variable or judgment-heavy work while retaining oversight for sensitive actions. Unlike traditional automation, which is suited to repetitive, low-risk processes, agentic systems can resolve complex customer support issues, qualify leads, support collections, coordinate healthcare intake, and assist with public-service requests across connected systems. Their potential benefits include higher resolution rates, faster service, greater consistency, and reduced routine workload for staff, but successful deployment requires careful testing, integrations, auditable logging, and clearly defined safeguards. The article presents ElevenAgents as a no-code platform for building and testing such workflows across channels including voice, chat, email, and WhatsApp.

A workshop led by ElevenLabs and ElevenCreative presented an AI-driven approach to marketing production that connects ideation, asset creation, localization, distribution, and performance analysis in a single workflow rather than relying on separate tools and handoffs. Using ElevenLabs’ Flows platform, the presenters demonstrated how two product photos and a prompt could generate an on-brand vertical advertisement with imagery, script, voiceover, music, and composition in minutes. The discussion emphasized that rising demand for ad variants and localized content makes AI-assisted production increasingly important, while dubbing that preserves the original performance can improve engagement and reduce advertising costs across languages. ElevenLabs argued that creative teams should prioritize AI producers with prompting skills and aesthetic judgment, while allowing agents to select models and automate technical tasks. Examples from Ramp and Clay illustrated uses in pre-production and multilingual campaign scaling, and ElevenLabs reported that localizing existing paid ads contributed to a 17% conversion increase and millions in incremental revenue. The company is developing a “Creative Engine” intended to automate the ongoing cycle of generating, publishing, localizing, and optimizing creative assets.

A context window is the token-limited active workspace an LLM uses for a single request, encompassing system instructions, prompts, conversation history, retrieved documents, tool outputs, and generated responses, but not its training data. Larger windows allow models to examine longer documents, codebases, and conversations together, yet they increase computation, memory use, latency, and cost while not ensuring that all included information is used effectively. Research such as Lost in the Middle and RULER indicates that models often retrieve information most reliably near the beginning or end of long inputs and may overlook relevant details buried in the middle. Because tokenization varies by language and model, developers should measure actual usage rather than rely on word-count estimates. Effective context management therefore emphasizes selective retrieval through RAG, summarization of older conversations, pruning irrelevant history, structured memory, and caching repeated information. Model selection should consider not only advertised context capacity but also practical performance on realistic workloads, including reasoning quality, retrieval accuracy, response speed, output limits, and cost.

Conversational AI for customer support can reduce operating costs and provide round-the-clock availability, but its value depends on reliably resolving customer issues while maintaining accuracy, brand alignment, natural interactions, and operational reliability. Effective evaluation combines quality metrics such as intent recognition, response accuracy, hallucination and word error rates; experience measures including customer satisfaction, voice naturalness, and latency; and operational indicators such as containment, escalation, uptime, and cost per conversation. The recommended approach uses real historical conversations to create ground-truth test sets, automated scoring validated by human reviewers, and per-intent analysis to identify weaknesses that aggregate scores may conceal. Changes should be A/B tested one variable at a time, such as greetings, voices, prompts, or model-routing strategies, and tests may need to run for weeks because conversational traffic is often limited. Before deployment, teams should conduct multi-turn simulations, response and tool-call tests, adversarial security testing, and phased rollouts with rollback plans, then continually monitor production transcripts, sentiment, user feedback, and model updates to detect regressions and feed failures into future improvements.

Cadence, a chronic disease care provider serving more than 100,000 patients across over 20 health systems, deployed ElevenAgents voice AI to accelerate outreach following abnormal vital readings amid rising alert volumes and a growing patient population. Using a HIPAA-compliant, low-latency platform integrated with its existing systems, Cadence built an alert-screening agent that conducts symptom assessments and transfers appropriate cases to nurses under embedded clinical protocols. After a human-supervised pilot and validation process, the agent reportedly adhered to clinical guardrails in all tested calls, reduced median patient-contact time from 1 hour and 48 minutes to 3.5 minutes, resolved 60% of screened alerts without clinician intervention, and increased answer rates by 47% relative to human RN outreach. Cadence handled 250,000 calls in its first seven months and plans to extend the agents to scheduling, device support, appointment preparation, and patient engagement, positioning the technology as a way to expand clinicians’ capacity rather than replace them.

ElevenAgents has introduced Reception, an AI receptionist designed for small businesses to answer calls, provide business information, and book appointments at any time. The service uses natural-sounding ElevenLabs voices and built-in agent orchestration, requiring no technical expertise and allowing businesses to forward an existing number or use a dedicated line. Reception can be configured to answer missed calls, after-hours calls, or all incoming calls, responding to customer questions in multiple languages about services, locations, and hours. It integrates with calendars to check availability and schedule appointments, while also offering public booking pages, call recordings, transcripts, and summaries for follow-up. The product is positioned for businesses such as home-service providers, professional services firms, real estate and property management companies, and salons that need to manage customer inquiries while staff are occupied or unavailable.

Real-time and batch speech-to-text transcription differ primarily in how they process audio, trading speed against accuracy, cost, and implementation complexity. Real-time, or streaming, models transcribe small audio segments with latency typically measured in milliseconds, making them suitable for live voice agents, captions, and meeting notes, but their limited conversational context can lead to errors. Batch models process completed recordings as a whole, using context from before and after ambiguous speech to improve word recognition, punctuation, formatting, and speaker labeling, which makes them better suited to call-center quality assurance, podcast archives, interviews, and bulk transcription despite slower turnaround times. Real-time systems generally cost more and require persistent connections and load balancing, while batch systems use simpler asynchronous workflows. Hybrid architectures can provide immediate live transcripts and later reprocess recordings with batch models for a more accurate archival version. ElevenAPI supports both approaches across more than 90 languages, offering Scribe v2 for batch transcription and Scribe v2 Realtime for low-latency applications.

Conversational AI uses voice and chat agents to automate repeatable customer and employee interactions such as answering questions, retrieving information, scheduling appointments, processing orders and returns, qualifying sales leads, and providing IT or internal knowledge support. Its effectiveness depends on narrowly defined tasks, access to approved business data and tools, and clear escalation paths for issues requiring human judgment. Applications vary across sectors, including healthcare scheduling and patient intake, banking account support, insurance claims, retail service, hospitality reservations, real estate lead qualification, education assistance, telecom troubleshooting, automotive bookings, and public-service information. Examples cited report faster response times, lower support costs, improved availability, and higher engagement, while emphasizing enterprise requirements such as system integration, compliance, governance, security, reliability, multilingual support, and testing before deployment.

Multi-agent systems coordinate multiple autonomous, LLM-powered AI agents to divide and complete complex workflows that would be difficult for a single agent, using communication protocols such as A2A for agent-to-agent exchange and MCP for access to external tools. Agents may be simple reflex, model-based, goal-based, utility-based, learning, hierarchical, or collaborative, and systems can use centralized, decentralized, hierarchical, or sequential architectures depending on needs for control, resilience, governance, scalability, or strict task dependencies. Their main advantages include specialized task handling, parallel processing, fault isolation, and easier scaling, but implementations can face coordination overhead, error propagation, high inference costs, and difficult debugging. Examples include Toyota subsidiary Woven’s code-compliance workflow, PGA TOUR’s content-production pipeline, CUADRA’s customer-support system, and Synergy Logistics’ warehouse-management tools. In voice AI, multi-agent designs can triage calls, retrieve information, and transfer customers between specialized agents while retaining context, and platforms such as ElevenAgents are presented as tools for creating and monitoring these voice and chat systems.

ElevenLabs has expanded its Model Context Protocol connector to let users manage agents and create multimedia content directly within assistants such as Claude, ChatGPT, Cursor, Grokbot, and Hermes. After a single OAuth-based installation, the connector provides access to more than 50 models for text-to-speech, transcription, dubbing, music, sound effects, image generation and editing, video creation, and lip-syncing, without requiring API keys or a separate server. Users can create connected assets from a single natural-language brief, including scripts, voiceovers, soundtracks, visual content, and finished dubbed videos, with outputs stored in the ElevenCreative workspace for further editing in Studio. The same MCP integration also supports ElevenLabs Agents, while workspace administrators can control available tools and select data residency settings during connection.

ElevenLabs has made Music v2.5 the default model for prompted and reference-based generation in ElevenMusic, describing it as producing more layered, coherent songs with more natural instruments than Music v2, particularly in vocal-led and acoustic-heavy genres such as R&B, soul, hip hop, rock, metal, orchestral, and cinematic music. The company reports that v2.5 was preferred more often in a blind comparison involving 47,885 paired generations, while Music v2 remains available. Lossless downloads are now offered across plans, with Free users receiving five per day and commercial-use rights when crediting ElevenMusic, while Pro users receive 400 per month; usage permissions remain attached to tracks even after cancellation, downgrading, or future terms changes. Downloads are not permitted for tracks referencing other artists’ songs, reflecting protections for rights holders. Music v2.5 is also available through ElevenCreative and the ElevenLabs API for uses including video scoring, Flows workflows, and product integrations, and ElevenLabs notes that its separate strategic agreement with Universal Music Group does not relate to this release.

The Model Context Protocol (MCP) is an open standard introduced by Anthropic in November 2024 and later donated to the Agentic AI Foundation, designed to let large language models connect securely with external data sources, tools, and services. It addresses the limits of models’ static training data and lack of native system access by allowing agents to discover available resources and functions dynamically, retrieve current information, and execute multi-step actions such as checking calendars, consulting CRMs, and sending emails. MCP uses a client-server architecture in which AI hosts use MCP clients to communicate with servers that expose tools and resources, reducing the scaling challenge of maintaining separate integrations for every model and service from an M×N problem to M+N connections. Unlike traditional APIs, which generally rely on hard-coded endpoints and parameters, MCP provides self-describing interfaces and runtime tool discovery tailored to contextual, agent-driven workflows. The protocol can improve automation, response accuracy, and complex agent behavior, with applications including conversational-agent management, context-aware coding assistants, and real-time customer-service voice agents; ElevenLabs is presented as a platform that can operate as both an MCP client and server.

ElevenLabs has entered a multi-year licensing agreement and strategic collaboration with Universal Music Group, marking its first partnership with a major record label and combining music licensing with AI product development. The companies plan to launch a separate AI-powered platform that will allow fans to create licensed remixes, mashups, alternate track interpretations, and personalized vocal experiences involving participating artists and songwriters. Both organizations emphasize responsible AI use, respect for human artistry, artist participation, rights management, and fair compensation, while also aiming to expand fan engagement, music discovery, and revenue opportunities. The partnership will additionally support future AI audio products and experiences, distinct from ElevenLabs’ existing Music API and ElevenMusic song-generation application.

AI orchestration coordinates multiple AI models, specialized agents, retrieval systems, and external tools into governed, multi-step workflows by routing tasks, maintaining context, executing API calls, and determining transitions between stages. It becomes important when processes such as customer support, loan servicing, healthcare scheduling, sales qualification, or e-commerce issue resolution require identity verification, data retrieval, policy application, and human escalation across connected systems. Effective platforms combine deterministic rules for high-stakes actions with model-based judgment for unstructured inputs, while providing multi-model support, context management, integrations, guardrails, audit logs, and human-in-the-loop controls. The piece argues that orchestration addresses the integration and governance gaps that limit many enterprise AI pilots, enabling organizations to scale specialized agents, improve consistency and oversight, reduce tool sprawl, and build new workflows more quickly. It highlights ElevenAgents as a no-code platform for designing such workflows, with configurable sub-agents, routing logic, approval gates, and deployment support.

Interactive voice response (IVR) is an automated telephone system that uses keypad tones or voice recognition to route callers, provide information, and complete routine tasks such as balance checks, appointment scheduling, payments, and order-status inquiries. Used since the 1970s across industries including banking, healthcare, retail, utilities, telecommunications, and government, IVR can provide around-the-clock service, reduce wait times and agent workloads, ensure consistent responses, and collect information before calls are transferred. Its effectiveness depends on clear menus, backend integrations, accurate speech processing, and an easy escalation path to a live agent, as rigid or poorly designed systems can frustrate callers and often fail to resolve requests independently. Unlike a basic auto attendant, which only routes calls, IVR can connect with CRM, billing, or database systems to retrieve information and process transactions. The text also describes conversational AI as an emerging alternative or supplement to menu-based IVR, using speech-to-text, language models, text-to-speech, and turn-taking detection to understand open-ended requests, access connected tools, take actions, and transfer callers with context when needed.

ElevenLabs has appointed former Adyen CFO Ethan Tandowsky as its chief financial officer to strengthen its financial operations during a period of rapid global growth. Tandowsky joined Adyen in 2016, helped guide its 2018 public listing, became CFO in 2023, and supported expansion from 400 to 5,000 employees alongside a 20-fold revenue increase. ElevenLabs, now an 800-person company serving millions of developers, creators, and businesses, offers voice and audio AI models in more than 90 languages and has deployed 10 million conversational agents. The company reports that large organizations generate 55% of its revenue, with customers including Nvidia, Stripe, Klarna, Deutsche Telekom, and several governments, while it also develops creator-focused revenue models such as a marketplace for voice intellectual property. Tandowsky will oversee the scaling of financial infrastructure and investment decisions related to research, products, and international expansion.

Sonic branding uses distinctive sounds, usually short audio cues, to make brands recognizable without visual elements, complementing the broader field of audio branding, which includes all customer-facing sounds such as notifications, ambient music, and service voices. Familiar examples include Netflix’s opening sound, Apple’s former Mac startup chime, Intel’s five-note signature, and Law & Order’s scene-transition effect, each illustrating how sound can become strongly associated with an organization or program. The source argues that growing audio consumption through podcasts, commuting, and multitasking makes sonic identity increasingly valuable, citing research that links it to improved recall, recognition, trust, and likability. It also emphasizes that consistent voices used in advertisements, support channels, and AI agents can function as part of a brand’s sonic identity, with Siri presented as a notable example, while promoting ElevenLabs’ tools for selecting and deploying voices consistently.

Vimeo integrated ElevenLabs Dubbing through ElevenAPI to provide automated video localization within its platform, combining transcription, translation, voice cloning, and speech synthesis in a single workflow that supports 35 source languages and 40 target languages. Customers can edit generated text tracks before publication, allowing them to review and refine translations, while Vimeo states that its AI policy prohibits training generative AI models on hosted videos without explicit creator consent. Since January 2026, Vimeo reports that users have dubbed 1.4 million minutes across more than 25,000 videos and completed over 134,000 multilingual dubbing jobs; 27% of enterprise accounts have tried the feature, with videos translated into an average of 3.3 languages and some into more than 30. The service is used for content such as training, product updates, and educational materials, enabling organizations to distribute videos internationally without maintaining dedicated dubbing teams in each market.

Retrieval-Augmented Generation (RAG) is an AI architecture that supplements large language models with relevant external information, allowing responses to reflect current, private, or organization-specific knowledge without retraining the model. It works by preparing documents as searchable chunks and embeddings, interpreting a user query, retrieving the most relevant passages, adding them to the model’s context, and generating an answer grounded in those sources. Unlike an LLM alone, which relies on training data and its immediate context window, a RAG-enabled system can draw from updated policies, product documents, databases, and large knowledge bases. RAG can reduce hallucinations and support applications such as customer service, internal HR or IT assistance, sales, and real-time voice agents, but its reliability depends on source quality, effective retrieval, context selection, and the model’s ability to interpret the retrieved content correctly. ElevenAgents offers built-in RAG capabilities for chat and voice agents, including query rewriting and parallel model processing intended to reduce retrieval latency for conversational use.

Havells has launched Agentic AI Voice Mode in its Havells One smart-home app, providing natural conversational control for connected appliances across eight languages, including English, Hindi, Tamil, Telugu, Bengali, Kannada, Gujarati, and Malayalam. Built with ElevenLabs’ real-time speech-to-text and text-to-speech technology alongside Havells’ proprietary device-control agent, the feature replaces rigid command-based interactions with context-aware, multilingual, and code-mixed conversations that can control individual devices, rooms, or multiple appliances at once. The system supports capabilities such as follow-up commands, ambiguity clarification, error-code explanations, device orchestration across connectivity protocols, and conversational feedback. Havells selected ElevenLabs after evaluating voice platforms for accuracy, multilingual performance, latency, and naturalness, then moved from prototype to production rollout in about six weeks with support for voice tuning in regional languages. Available to the app’s approximately 2.7 million-download user base, Voice Mode is intended as a foundation for future personalization, broader language support, automation, troubleshooting, scheduling, and expansion into the wider Havells ecosystem.

Intent recognition is a natural language processing technique that identifies the goal behind spoken or written user input so AI systems can provide appropriate responses, route requests, or perform actions. Rather than relying on keyword matching, it typically preprocesses input, converts it into numerical features or contextual embeddings, classifies it against known intents using confidence scores, and extracts relevant entities such as account details or order numbers. Common intent categories include informational, navigational, transactional, and conversational, while intent detection may refer more narrowly to determining whether an applicable intent exists at all. The approach supports applications in customer service, healthcare scheduling, and internal HR assistants, where accuracy can be especially important in high-risk settings. Key challenges include ambiguous language, questions outside a model’s scope, and requests containing multiple goals, which can be addressed through conversational context, edge-case training and guardrails, and multi-label classification.

ElevenLabs’ Impact Program highlights the story of Eliane, a Brazilian woman who lost her voice in childhood, and Alberto, who became blind at age 46, as they developed their own communication methods and fell in love after meeting in São Paulo in 2024. Using ElevenLabs technology, Eliane received a synthesized voice modeled on the likeness of her family, allowing her to communicate in a voice intended to reflect her identity; Alberto had not previously heard what her natural voice might have sounded like. The story is the program’s first featured initiative in Brazil and accompanies partnerships with the Associação Brasileira de Esclerose Lateral Amiotrófica, which will provide free voice-cloning access to people with ALS, and the Sociedade Brasileira de Fonoaudiologia, which will train speech-language pathologists to use the technology. These efforts support ElevenLabs’ 1 Million Voices initiative, a $1 billion in-kind commitment that the company says has reached more than 11,000 people globally, and Eliane and Alberto plan to marry in November.

Conversational AI and generative AI serve different but increasingly complementary purposes: conversational AI is designed for context-aware text or voice interactions, while generative AI creates original text, code, images, video, and audio from prompts. Conversational systems combine natural language understanding, dialogue management, natural language generation, company knowledge sources, and sometimes large language models to assist customers, complete workflows, and escalate issues when necessary. Generative systems typically use large language models, transformer architectures, and diffusion models trained on broad datasets to produce content, though they can introduce hallucinations or reflect biases in their training data. Both are generally considered weak, limited-memory AI, but conversational AI focuses on resolving requests and maintaining dialogue, whereas generative AI focuses on content creation. Businesses may choose conversational AI for customer support, appointment scheduling, lead qualification, and digital assistants, or generative AI for marketing, design, writing, coding, and summarization, while combining both to create experiences such as a support interaction followed by a personalized generated proposal or email.

ElevenLabs announced the appointment of Ashley Kramer as Chief Revenue Officer, bringing nearly two decades of leadership experience from high-growth technology companies including OpenAI, GitLab, Alteryx, and Tableau. Kramer will lead the company’s global revenue organization and partnership ecosystem as ElevenLabs expands its enterprise-focused AI interaction platform, which aims to provide natural, personalized voice and agent experiences across languages, cultures, and customer contexts. The company reports that its agents handle more than 15 million conversations weekly, with customers including Meta, Stripe, Klarna, Hasbro, and Deutsche Telekom, while enterprise customers now generate 55% of its revenue. ElevenLabs’ leadership describes the hire as part of its effort to scale deployments that support customer service, sales, marketing, product development, and business operations for large organizations.

Tool calling enables AI models to access external systems such as APIs, databases, software, and devices to retrieve current information, execute code, automate workflows, and complete actions beyond their training data. Rather than directly operating a tool, the model identifies when assistance is needed, selects an appropriate tool based on its schema, produces a structured request, and relies on an application layer to execute it and return results for a response or additional steps. This capability can reduce hallucinations, provide real-time data, support multi-step automation, and expand AI assistants from information providers into agents that can act across connected systems. Unlike retrieval-augmented generation, which mainly supplies static documents and knowledge-base context, tool calling typically accesses dynamic data and operational systems, though both approaches are often combined. Function calling is described as a narrower form of tool calling focused on predefined functions, while tool calling encompasses broader API, code, workflow, and device interactions. Reliability varies by model and workflow complexity, with benchmarks such as the Berkeley Function Calling Leaderboard evaluating accuracy, latency, cost, and multi-turn performance. ElevenAgents applies these capabilities to voice agents through webhooks, client and code tools, MCP servers, built-in system actions, and connectors for platforms including Zendesk, HubSpot, and Salesforce.

ElevenLabs reports that its AI sales development representative, Dom, has reduced inbound lead qualification from a two-day human median to minutes, generating more than $1 million in qualified pipeline in a month while saving its SDR team 180 hours. Dom operates outside standard business hours, supports 32 languages, earns an 8.8/10 customer satisfaction score, and achieved 92% qualification accuracy after improvements from an initial 78%. Its qualification process combines live-call responses with AI analysis of company websites and startup-grant eligibility, allowing inconclusive calls to remain eligible for conversion. Using enriched lead data and ElevenAgents Workflows, Dom routes prospects to tailored outcomes such as demos, self-service signup, or customer support, then sends relevant follow-up emails and records decisions and supporting data in the CRM. The company reviews all conversations daily and plans to make Dom the main entry point for sales leads while extending its interactions across voice, WhatsApp, and email to support both inbound and outbound sales activity.

A system prompt is a high-priority set of developer-defined instructions that establishes an AI agent’s role, context, tone, goals, operational boundaries, response format, and tool-use rules before a user begins a conversation. Unlike user prompts, which change with each request and specify individual tasks, system prompts provide persistent behavioral guidance and guardrails across interactions, although they do not govern underlying mechanics such as turn-taking or supported languages. The guidance recommends writing prompts as clear, structured specifications with distinct sections for non-negotiable rules, edge-case handling, and tool instructions, while matching their length and detail to the complexity and risk of the agent’s responsibilities. Examples involving technical support, refunds, and dental scheduling illustrate how identity verification, eligibility checks, tool-call sequences, escalation paths, and direct error handling can make agents more reliable. The material also notes that system prompts vary in configuration across platforms such as ChatGPT, Claude, and enterprise tools, and should be tested with candidate models and continuously refined using real customer conversations and simulations.