
Voiceflow lets support, design and engineering teams build, test and ship chat and voice agents in one workspace, with playbooks, visual workflows, test simulations, environments and credit-based pricing.
Used this tool? Rate it
Used this tool? Rate it
AI assistant for writing, analysis and tasks
DeepSeek AI models for chat and API
One workspace for every AI model
Chat Your Way to a Finished Video
Plan, build, and refine work with an AI agent
The agentic customer platform for go-to-market teams
AI image editing from a reference photo
Multi-model image studio, free to start
Voiceflow is an AI agent platform for customer experience automation: teams build, test and ship chat and voice agents in one workspace, then keep improving them from real conversations. Work happens in the browser app, from the API and CLI, through an MCP server, or by asking Atlas to make changes.
Support, design, product, engineering and QA teams share the same workspaces, where editors build agents and viewers review projects. Voiceflow reports 3,000+ customers, 10,000+ live agents and more than 1 billion messages processed.
Trade press reported in August 2023 that a $15 million round led by OpenView brought total funding to $35 million at a $105 million post-money valuation.
An agent brings a global prompt, instructions, playbooks, workflows, knowledge and tools together, and the agent studio is the collaborative visual workspace where those parts are edited and reviewed.
Two kinds of logic coexist. Playbooks are goal-specific conversations that adapt to what the customer says and what tools return, while workflows are defined steps and branches that can call a playbook wherever flexible reasoning is useful.
| Need | Choose | Example |
|---|---|---|
| Questions and actions must adapt to the customer's situation | Playbook | Clarify an order-status request and explain the lookup result |
| A defined process with explicit branches | Workflow | Collect an order ID, check eligibility, then branch |
| A controlled process with one flexible part | Both | Use a playbook to clarify a return reason inside a returns workflow |
The distinction is about control rather than about whether AI is involved: a playbook can call tools, and a workflow can contain AI reasoning.
Answers are grounded in imported content. Documents can be imported as .pdf, .txt or .docx files up to 10 MB, and only text content is imported. Tables arrive as .csv or .xlsx files up to 10 MB, with each row treated as a chunk and column headers becoming field names. Web pages, sitemaps, pasted text and connected help-centre sources such as Zendesk articles are the other import routes.
Tools are how an agent acts: integrations, API calls, MCP servers and JavaScript functions, so a refund, a booking change or an account lookup runs inside the conversation instead of being handed to a person. Human hand-off routes a conversation to Salesforce, Zendesk, Intercom, Gorgias, Kustomer, Freshdesk, Genesys, Amazon Connect, webhooks or another helpdesk. The platform also runs its own MCP server, and a connection can be narrowed to selected tool groups, read-only access or a single project.
The model is chosen per agent, with per-playbook overrides, and transcript evaluations can use a separate model. Voice pickers load the complete ElevenLabs and Cartesia catalogs, and built-in voices show languages, accents, ages and use cases. Voice configuration itself controls speech recognition, spoken replies and phone-session behaviour for an environment.
Tests specify user messages and checks on agent turns, while simulations describe a scenario and an outcome and let AI generate that part of the conversation. A run executes the authored conversation, including its scripted turns, simulations and checks. Starting a run is explicit: saving a test does not start it.
Analytics brings conversation volume, outcomes and operating cost into one view. Transcripts set customer and agent messages beside the execution logs that produced them, and call recordings sit next to the transcript of the voice conversation. Translating a transcript into the reader's language is available from the transcript view and spends the workspace's credits.
Atlas is the AI copilot for building and improving agents. It investigates conversations, proposes instruction edits and creates tests from real cases. Its edit mode switches between Manual, which asks before applying changes, and Auto. Routines let Atlas run on a schedule, such as a weekly review of last week's conversations.
Each environment is a separate copy of an agent's configuration with its own draft, published version and version history, and every project starts with Main. Publishing promotes the draft to the live version through a side-by-side diff, and if the environment is protected only admins and owners can publish it. A share of incoming traffic can be assigned to different environments to compare versions or roll out gradually, and a user stays on one environment across sessions. Editors build agents, viewers review projects and leave comments, and workspace roles and project permissions control access.
The documented loop repeats for every change: build in a separate environment, prove the change with tests, then publish and watch production.
A published version can be reverted, and the revert is recorded as a new version noting which version it came from. Reverting restores content, not credentials: the environment keeps its current secret values.
The documentation's guides map seven concrete builds: inbound calls that answer policy questions and transfer callers, outbound calls such as a pickup reminder, chat support grounded in policy, website lead generation that creates a HubSpot contact, customer copilots embedded in a product, agent assist for support staff, and internal workflows such as IT intake that files a Zendesk request.
The support agent is positioned for high-volume requests with known steps: order status, refunds inside policy, resets, plan changes and billing questions. Everything else routes to the team with the case assembled, and hand-off attaches the transcript and account context in the helpdesk the team already uses.
Voiceflow reports a 70% AI resolution rate for Trilogy, and its own FAQ says the figure comes from customer deployments rather than from a benchmark: a team's rate depends on its ticket mix and on how much of its policy is automated.
Voiceflow fits organisations where several roles touch the same agent and the agent must act in business systems, not only answer questions.
Self-serve builders are served too: Pro is designed for individual builders, includes all LLM models and a higher agent limit than Free.
Four channels are documented: a website chat widget, phone numbers for calls, SMS through Twilio, and custom apps on the conversation APIs.
| Channel | How an agent reaches customers | Notes |
|---|---|---|
| Website | Embedded chat widget, customisable and extensible | Styling, modality and file upload settings are configurable |
| Phone | A connected number for inbound or outbound calls | Call transfers go to a phone number or SIP destination |
| SMS | A Twilio number for inbound and outbound text | Handoff replies relay back to the customer's phone |
| Custom app | Conversation APIs for your own interface | Chat and voice clients are built on the same runtime |
A solutions page also lists WhatsApp and routes it to a demo request, so the marketing surface is broader than the documented guides.
On the developer side, coding agents such as Claude Code, Codex or Cursor connect through a setup prompt, the CLI or an MCP server. Infrastructure runs on AWS and GCP with isolated environments, and the voice page states 99.95% agent uptime, 50 ms Voiceflow latency and 300K messages per minute scale. The security page lists SOC 2 Type II, ISO 27001, GDPR and HIPAA as the frameworks it reports on.
Voiceflow offers four plans: Free, Pro, Business and Enterprise.
| Plan | Designed for | Credit basis | Notable extras |
|---|---|---|---|
| Free | Trying out Voiceflow | One-time grant that does not renew | Limited features, not suitable for production |
| Pro | Individual builders | Monthly allotment | All LLM models, higher agent limit than Free |
| Business | Growing teams | Monthly allotment | LLM fallback models, priority support |
| Enterprise | Large-scale workloads and advanced security | Custom credit allotments | Unlimited product usage, single sign-on, private cloud hosting |
Billing moved to credits on 29 April 2025, replacing token-based pricing. Credits cover automated phone calls, LLM responses, text-to-speech minutes and messages, so usage is measured in one unit. Every interaction consumes credits depending on the vendor and model used: the published table prices chat usage at $0.005 per message and phone usage at $0.05 per minute, with LLM and speech vendors metered separately.
Editor seats are an add-on, every paid plan includes some free seats, and viewers are free and unlimited. One phone number per workspace is included; extra numbers and a higher concurrent-call limit are add-ons, and that limit does not apply to chat or API sessions. Credits are shared across all workspaces in an organisation and reset at the start of each billing cycle, and auto top-ups buy more when the balance reaches zero.
The public pricing page asks businesses to book a demo or request pricing, while the agency track advertises a free trial with no credit card and usage-based billing with optional auto top-ups. When the credit system launched in April 2025 the published list showed Pro at $60/mo for 10,000 credits a month, and additional editor seats at $50 per month. Current prices live in the in-app Plans and Billing tab, so treat those as the last published list rather than today's quote.
Voiceflow is a builder, not a managed agent service, and it does not replace the contact-centre platform: the agent runs on top of the platform you have, answers and resolves calls before they reach a queue, and hands the rest to the team through existing routing.
For buyers who need published self-serve security rather than a sales conversation, single sign-on and private cloud hosting appear in the Enterprise plan description rather than in the Free, Pro or Business tiers.
Trade press coverage in 2023 listed Google Dialogflow, IBM Watson, AWS Lex and Microsoft Bot Framework among Voiceflow's rivals, and those platform-native options matter when a company has already standardised on one cloud.
The Free plan has a one-time credit grant that does not renew and is not suitable for production usage. Credits expire at the end of the subscription period: monthly credits do not roll over, while annual subscribers receive all their credits at once.
Testing is metered too. Tests and simulations consume credits when they use LLM-powered features, and there is no single fixed price documented for every test. Transcript translation is metered the same way and is refused when credits run out.
The agent model has a documented scope: when a playbook is active, agent-level routing instructions are no longer in context, so the guidance a playbook needs must sit in its own instructions.
Knowledge base document contents are shared across environments, so editing a shared document can affect the published agent even while the work sits in a separate environment. Imported documents and tables are capped in size, only text is read from documents, and every page of a sitemap counts toward the plan's document limit.
A tool that creates a support ticket does not, by itself, connect the customer to a live human conversation, and the hand-off step is separate from transferring a phone call.
PII redaction is a paid add-on: it must be subscribed to by contacting the sales team. Redaction applies only to production conversations, and development and test conversations are not redacted. Enabling redaction does not affect existing transcripts, and disabling it does not unredact earlier ones.
All payment plans are paid in advance and are non-refundable, and there are no refunds or credits for partial months or for months unused with an open account. Prices can change on 30 days' notice posted on the sites or in the service. Users retain ownership of their user content, but Voiceflow has no obligation to store, maintain or provide a copy of it.
The security page states that Voiceflow does not use customer data to train any machine learning or AI models, and that model providers integrated through the platform, such as OpenAI, Anthropic and Google, are configured with zero data retention. That is a vendor statement rather than an independent audit finding. Personal information is retained for as long as needed to provide the service, or as required or permitted by law.
Voiceflow is based in Canada, and the privacy policy says data may be hosted, transferred and processed in Canada and in other countries whose data protection laws differ from the user's own. The terms make the service available only to individuals aged 13 or older.
Teams build, test, deploy and improve chat and voice agents in one workspace, connecting business knowledge and tools so an agent can answer questions, complete tasks and hand over when needed.
No. Instructions, visual workflows and pre-built integrations cover most builds, and developers can add custom functions, call APIs and build their own interfaces. The platform also provides APIs, a CLI and MCP for teams that want to automate their development workflow.
The documented channels are a website chat widget, phone numbers for inbound and outbound calls, SMS through Twilio, and custom apps built on the conversation APIs. WhatsApp appears on a solutions page as a demo-request channel.
Yes. The model is chosen per agent with overrides per playbook, and a separate model can be used for transcript evaluations, which lets a team trade quality, speed and cost per task.
A plan fee, optional add-ons such as editor seats and extra phone numbers, and credit-based usage. Chat usage is priced at $0.005 per message and phone usage at $0.05 per minute on the published credit table, with LLM and speech vendors metered separately. Credits are shared across the organisation and do not roll over.
No. The agent runs on top of the platform you already have, resolves calls before they reach a queue and hands the rest to your team through existing routing, with the transcript and account context attached.