Voice AI Agents Training with ElevenLabs

ElevenLabs' Agents Platform combines a system prompt and knowledge base with a voice. The result is an agent that can hold a live conversation on the phone or the web. A polished five-minute demo still has to survive production. This training walks your team through an end-to-end build. You choose or design a voice, ground it in your own knowledge base and tools, tune turn-taking and latency, and test with simulated calls before the agent reaches a customer. It also covers the parts most teams skip. That includes multilingual and telephony deployment, voice-cloning consent and disclosure of an AI voice to callers. By the end, your team owns a working agent and a repeatable process for shipping the next one.

modules
6
hours
13
Contact us

Why this training

Why train on this tool?

System prompts and voice briefs. Knowledge-base entries and simulated call scripts. Consent records sit beside escalation rules and post-call transcripts. A voice-agent build produces all of it on ElevenLabs' Agents Platform. A demo that nails three questions in five minutes proves little once a caller talks over the agent or switches languages mid-sentence. An out-of-scope question can break it too.

None of that list is exotic once you separate it from the parts that are genuinely new. A knowledge base and a system prompt are familiar problems with a voice wrapped around them. Teams can learn directly how to ground answers in approved documents and write for a persona. What's new is the audio layer: turn-taking, latency, interruption handling, and a caller who never sees a transcript to fall back on. This training treats that layer as a specific, teachable skill beyond a general chatbot background.

Your team designs or clones a voice with documented consent. It connects a knowledge base and tools, then tunes turn-taking until the conversation feels natural. One module covers multilingual and telephony deployment through a web widget or native phone number, including a SIP trunk. Another covers what keeps the agent trustworthy after launch. Every caller gets disclosure, consent stays on file and monitoring continues. Teams that need broader architecture across tools should start with our AI training hub and AI Agents training.

Does the demo survive an interrupted caller?

Three test questions answered smoothly doesn't prove much. A real caller talks over the agent, has an accent the model wasn't tuned for, or asks something out of scope, and the polished demo cracks. Most voice-agent projects stall in that gap between demo and production. Closing it is a specific skill a team can learn.

Latency has no chatbot precedent

A half-second pause or talking over a caller can break a phone conversation in ways a text interface never will, and a voice agent built like a chatbot with a microphone will frustrate callers.

Every call discloses the AI, every clone has signed consent

Cloning a voice without documented consent, or letting an agent go unannounced on a call, creates legal and reputational exposure that lands on whoever shipped it. Consent and disclosure get built into the agent from day one and enforced in the call itself.

Made once, expensively

Whether an agent lives on a web widget, a phone number, or a SIP trunk, and how it handles a caller who switches languages mid-sentence, shapes the whole build. Getting this wrong later means rebuilding.

The platform ships faster than any wiki keeps up

ElevenLabs adds new turn-taking controls and testing tools on a fast cycle. Integrations change quickly too. A team that learns the underlying workflow keeps up without retraining from scratch.

Syllabus

Training syllabus

  1. Voice agents on ElevenLabs and the program boundary

    90 minBeginner

    The group starts with the parts of a voice AI agent: a voice, a knowledge base, tools, and a conversation loop running on ElevenLabs' Agents Platform. It compares that setup with a text chatbot and an IVR menu, then with a general-purpose AI agent. It also sees where this program stops, handing off to contact-center strategy or general agent architecture.

    • What builds a voice agent: a voice, a system prompt, knowledge base, and tools
    • A voice agent versus a text chatbot and IVR tree, then a general-purpose AI agent
    • Good first use cases, FAQ and triage, booking, reminders, multilingual front desk
    • What this program will not touch, identity checks and safety-critical instructions
    • Where this program stops

  2. Designing the voice and the persona

    120 minBeginner

    The group compares a voice from the Voice Library with one generated through Voice Design or cloned through Instant or Professional Voice Cloning. Every path includes its required consent step. Participants then write a persona and system prompt built for speech.

    • Voice Library versus Voice Design versus voice cloning
    • Voice Design controls: age, gender, accent, and tone for a voice that does not mimic anyone
    • Instant Voice Cloning vs Professional Voice Cloning, and the consent each one requires
    • Writing a system prompt for speech, with pacing and backchanneling built in
    • One consistent voice and persona across every language

  3. Knowledge and tools in the conversation flow

    150 minIntermediate

    Ground the agent in a knowledge base so its answers come from approved documents, then wire tools for bookings or CRM lookups. You also tune its eagerness and silence handling. Interruption settings complete the work so turn-taking feels natural.

    • Grounding answers in a knowledge base instead of the model's memory
    • Wiring tools and function calls for bookings and lookups, including CRM actions
    • A narrow sub-agent with its own tools and knowledge
    • Turn-taking tuned by eagerness and silence timeout, plus interruption handling
    • Clean human handoffs

  4. Testing and evaluation before launch

    150 minIntermediate

    A voice agent earns production trust through evidence. Live and simulated conversations become a test suite. The team defines success criteria up front, and a review habit catches hallucinated answers and dropped conversations before a caller does.

    • Turning underperforming live conversations into simulation test cases
    • Simulation and next-reply tests, plus tool-call tests
    • Mocking tools safely
    • Success criteria and structured data collection, defined before launch
    • A hallucination check against your own knowledge base
    • Reviewing transcripts as a standing habit

  5. Multilingual and channel deployment

    120 minIntermediate

    The agent moves from one test call to a web widget, phone number, or SIP trunk, including callers who switch languages mid-conversation, and those deployment decisions are expensive to change after launch.

    • Web widget and phone deployment, including SIP
    • What automatic language switching covers, and what it doesn't
    • Native phone-number integration versus SIP trunking for scale
    • Encrypted signaling and static IPs for compliance
    • One human-handoff route across every deployed channel

  6. Consent and disclosure while running the agent

    120 minAdvanced

    The operating habits that keep a voice agent trustworthy after launch. Documented voice-cloning consent comes first. The agent tells people they're talking to AI, while a monitoring routine draws on the platform's own analytics. The module closes with a rollout plan for one governed agent.

    • Documented consent and rights before any voice clone goes live
    • Disclosing an AI voice to callers and users, plainly
    • Detection tooling that catches unauthorized use of a cloned voice
    • Live monitoring through post-call analytics and evaluation trends
    • Choosing a pilot use case and champion, with adoption metrics beyond call volume
    • One tested, disclosed, monitored voice agent, shipped as the capstone

Outcomes

Outcomes & audience

What you will learn

  • Build a working ElevenLabs agent, voice, knowledge base, tools, and conversation flow, end to end
  • Choose the Voice Library, Voice Design, or cloning, with documented consent before any clone
  • Ground answers in approved knowledge
  • Tune turns to sound human
  • A simulation test suite and launch-ready success criteria
  • Deploy to a widget or phone number, including a SIP trunk for multilingual callers
  • Every caller told it's an AI, every clone consented in writing
  • Monitor live agents and run a pilot-to-rollout plan

Who should attend

  • Product and platform owners building a voice agent for support or sales, including booking
  • CX and contact-center leaders scoping where this fits their operation
  • Growth and localization teams shipping a multilingual voice front line
  • Engineers and no-code builders configuring agents on ElevenLabs itself
  • Founders and product managers piloting their first voice agent
  • Whoever on the trust and safety side signs off on consent and disclosure

Format

Training format

Voice-agent teams spend two days building, with most of it spent configuring, testing, and iterating on a working agent rather than watching slides. Onsite and live-online cover the same syllabus, and small groups mean everyone builds instead of observing.

Format
Onsite or live online
Duration
2 days (about 12 hours), can be split into half-day sessions
Group size
Up to 12 participants per group (hands-on building)
Materials
Voice and persona brief, test-suite template, and a consent and disclosure checklist
Language
English or Turkish
Certificate
Certificate of completion

About Zeo

Why Zeo

Zeo started in 2011 and now works out of San Francisco, Istanbul, Ankara, and Lisbon. We run Copilot Academy and organize Digitalzone, an international digital marketing conference. This program draws on the 10+ years of consulting and training work behind that, applied to corporate AI adoption.

  • 2011founded in Istanbul
  • 10+years of consulting and training experience
  • 3offices: San Francisco, Istanbul, Ankara, Lisbon

Keep exploring

Video Production with Sora6 modules · 12 hoursSora and other text-to-video tools can produce a striking clip in seconds. Turning that into a repeatable production stage, from brief and storyboard through motion prompting, continuity, review, and a provenance-tagged handoff, is what this training builds.View trainingAI Agents6 modules · 13 hoursNo-code AI agents training for business and operations teams: what an agent really is, when to use one, and how to build and govern one on your own workflows, from Copilot Studio to custom GPTs.View trainingAI in Customer Support & CX6 modules · 13 hoursAI training for customer support and CX teams, built around the actual queue: faster response drafting, clearer knowledge content, consistent quality, and escalation that stays visible instead of buried.View trainingAI Automation with n8n6 modules · 13 hoursAI automation training for ops and marketing teams: build, test, and govern workflows with n8n, keep a person in the loop where it matters, and hold clear data boundaries throughout.View training

Next step

Let's build your first governed voice agent

Share the use case and the languages you need, then tell us how you plan to deploy it. From there, we shape the syllabus and the labs around what you're actually shipping.
Contact us
Illustrated figure working on a laptop surrounded by floating tool windows

Questions

Frequently asked questions