Retell AI
Production-ready voice agents with a visual flow builder and low-latency calling
Retell AI is a platform for building and deploying AI voice agents that handle inbound and outbound phone calls. It combines a low-latency conversational engine with a visual conversation flow builder, telephony, function calling, testing tooling, and post-call analysis, positioned between raw developer APIs and no-code receptionist products.
Overview
Retell occupies a useful middle ground. It is API-first enough for engineers to build sophisticated agents, but its conversation flow builder means a technically comfortable operations person can construct and maintain an agent without writing the whole thing in code. That combination has made it popular with agencies and product teams deploying voice agents for real customers rather than experimenting.
The engineering emphasis is on the parts that determine whether a call feels acceptable: latency low enough that pauses do not signal a machine, interruption handling that does not talk over people, and voice quality good enough to hold attention. Around that sit the practical requirements of production deployment: batch outbound calling, warm transfer to humans, knowledge bases for question answering, and post-call analysis that extracts structured outcomes.
Pricing is per minute with the components broken out, which makes the economics legible and lets teams trade voice quality or model capability against cost. As with every platform in this category, the limiting factors are less technical than operational: conversation design determines quality, and the legal environment around automated calling and AI disclosure varies by market and continues to tighten.
Best for
Product teams and technically capable agencies deploying voice agents in production for booking, qualification, support triage, and outbound campaigns, who want flow-based control without building orchestration themselves.
Not the right fit for
- Completely non-technical buyers wanting a receptionist configured by a wizard.
- Organizations needing a full contact center with workforce management and omnichannel queues.
- Highly regulated scripts requiring guaranteed verbatim disclosure without extensive validation.
- Markets where automated outbound calling faces severe restrictions.
- Teams unwilling to invest in conversation design and iterative testing.
How it works
- 1
You build an agent either as a single prompt with tools, or as a conversation flow: a visual graph of nodes with instructions, conditions, and required data, which constrains behavior at each stage of the call.
- 2
Voice, language model, and speech recognition settings are configured per agent, along with turn-taking sensitivity, backchanneling, and interruption behavior that determine how natural the exchange feels.
- 3
The agent is attached to a phone number for inbound calls or invoked through the API and batch tools for outbound campaigns, with dynamic variables passing caller context at call time.
- 4
Functions let the agent act during the conversation, checking a calendar, retrieving a record, or transferring to a human, and after the call, transcripts, recordings, structured extractions, and success evaluation are delivered through webhooks and the dashboard.
Feature breakdown
20 features in 4 modulesAgent building
How behavior is defined and constrained.- Conversation flow builder
- A visual graph of nodes with per-node instructions, transitions, and data requirements, so agents follow a process rather than improvising.
- Single-prompt agents
- A simpler mode for straightforward use cases where a flow would be unnecessary structure.
- Knowledge base
- Documents attached to an agent for answering questions without hard-coding responses into the prompt.
- Dynamic variables
- Caller and account context injected at call time so one agent handles personalized conversations.
- Multi-language support
- Agents configured for different languages, with voice and recognition chosen appropriately per market.
Conversation quality
The engineering that decides whether calls feel acceptable.- Low-latency response
- Streaming across the pipeline so replies begin quickly enough that pauses do not signal automation.
- Interruption handling
- The agent yields when the caller speaks, with sensitivity tuned to avoid stopping for background noise.
- Backchanneling
- Natural acknowledgements during longer caller turns so the line does not feel dead.
- Voice selection
- A range of voices across providers with control over style, plus cloning options for brand consistency.
- Noise and line quality handling
- Configuration for real-world conditions rather than assuming clean studio audio.
Telephony and actions
Connecting to the phone network and to your systems.- Inbound and outbound calling
- Numbers provisioned in-platform or brought from an existing provider, with programmatic dialing.
- Batch calling
- Campaign tooling for calling lists with concurrency control, scheduling, and retry logic.
- Function calling
- Agents invoke your APIs mid-conversation to check availability, look up records, or write data.
- Call transfer
- Warm handoff to humans with context, plus configurable conditions for when to escalate.
- Voicemail detection
- Recognizing answering machines and behaving appropriately rather than talking to a recording.
Testing and analysis
Knowing whether the agent is actually working.- Simulated testing
- Automated test conversations against defined scenarios to catch regressions before customers hear them.
- Post-call analysis
- Structured extraction of outcomes and fields from each conversation, delivered to downstream systems.
- Success evaluation
- Automatic assessment of whether the call achieved its objective, which turns volume into a performance metric.
- Call logs with timing
- Transcripts and recordings with per-turn latency so failures can be diagnosed rather than guessed at.
- Webhooks and integrations
- Real-time and post-call delivery into CRM, scheduling, and workflow tools.
Use cases
4 documentedClinic network handling appointment demand
Reception misses a large share of calls at peak times, and each missed call is a lost appointment.
An inbound agent books and reschedules using calendar functions, escalating clinical questions to staff immediately.
Home services company qualifying enquiries overnight
Enquiries arrive at all hours and the first responder usually wins the job.
An agent answers around the clock, qualifies the job, and books a visit, with details written into the operations system.
Sales team running reactivation campaigns
A large dormant lead list is worth calling but nobody has the capacity.
Batch calling works the list with voicemail detection and retry logic, and interested prospects transfer straight to a representative.
Agency deploying agents for multiple clients
Each client needs a differently behaved agent, and rebuilding orchestration per client is unworkable.
Flow-based agents are configured per client on one platform, with per-client numbers, analysis, and reporting.
Pricing
from From roughly $0.07 per minute combined, with free credits to startUsage-based per minute with components broken out, voice, language model, transcription, and telephony, so teams can trade quality against cost. Volume discounts and enterprise arrangements available.
| Plan | Price | Includes |
|---|---|---|
| Pay as you go | From about $0.07 per minute |
|
| Volume | Reduced per-minute rates committed usage |
|
| Enterprise | Quoted annual |
|
Billing notes
- Component-level pricing makes the economics legible: a premium voice or a stronger model raises the per-minute rate visibly.
- Telephony rates vary by destination and are additional to platform and model costs.
- Outbound campaign cost should be modelled on connect rate and average duration rather than list size.
- Volume commitments materially reduce per-minute rates.
- Rates as published August 2026; pricing across this category has fallen steadily and should be reconfirmed.
Value assessment: Per-minute economics make the arithmetic against staffed calling straightforward, particularly for out-of-hours coverage where the alternative is missing the call entirely. Retell's specific value is reducing the engineering required to reach production: flow building, testing, batch calling, and post-call analysis are all present rather than being things you build around an API. The real cost remains conversation design and testing, which no platform removes.
Strengths & limitations
Strengths
- Flow builder gives structured control without requiring everything to be written in code.
- Strong focus on latency and interruption handling, which determine whether calls feel acceptable.
- Production tooling included: batch calling, voicemail detection, testing, and success evaluation.
- Function calling lets agents complete tasks rather than only gather information.
- Component-level pricing makes cost and quality trade-offs explicit.
- Good fit for agencies deploying differentiated agents across multiple clients.
Limitations
- Still requires technical capability; it is not a no-code product despite the visual builder.
- Agent quality depends heavily on flow and prompt design, which is genuine work.
- Telephony and provider costs stack on top of platform fees, complicating budgeting.
- Automated outbound calling faces varying legal restrictions by jurisdiction.
- AI disclosure expectations are tightening and differ by market.
- Open-ended conversations remain unreliable compared with bounded, structured ones.
Head-to-head comparisons
5 alternativesRetell AI vs Vapi
from From roughly $0.05 per minute platform fee, plus provider costsClose competitors for the same developer audience. Vapi goes further on composability, letting you choose and swap every provider in the pipeline. Retell is generally considered smoother out of the box, with a stronger flow builder and more production tooling included. Teams wanting maximum control choose Vapi; teams wanting to reach production faster often choose Retell.
Full Retell AI vs Vapi comparisonRetell AI vs Bland AI
from From roughly $0.09 per minute of call time, with free credits to startBoth target developers with production voice agents and pathway-style flow control. Bland self-hosts its entire stack for latency consistency, while Retell emphasizes tooling around the agent lifecycle including testing and evaluation. The practical approach is building the same agent on both and comparing real call recordings, since the differences are experiential rather than architectural on paper.
Full Retell AI vs Bland AI comparisonRetell AI vs Synthflow
from From roughly $29 per month with bundled minutesDifferent buyers. Synthflow is a no-code platform with templates and white labeling aimed at agencies serving small businesses. Retell assumes technical capability and offers more control in exchange. An agency selling packaged receptionists at volume will move faster on Synthflow; one building customized agents will prefer Retell.
Full Retell AI vs Synthflow comparisonRetell AI vs Phonely
from From roughly $0.10 per minute, with plans including bundled usageRetell expects you to design the conversation, with a visual flow builder, testing tooling, and batch outbound calling. Phonely is inbound-first and improves answers from historical call transcripts and human feedback rather than from flows you author. Support teams with a large back catalogue of calls and no appetite for flow design get more from Phonely; teams running outbound campaigns or needing explicit control at each stage get more from Retell.
Full Retell AI vs Phonely comparisonRetell AI vs Vocode
from Free and open source to self-host; hosted platform priced per minute of call timeRetell is a managed platform from about $0.07 per minute with the flow builder, batch calling, and post-call analysis included. Vocode's self-hosted path carries no license cost and keeps call audio in your own infrastructure, at the price of owning latency tuning, scaling, and provider upgrades yourself. Teams shipping production calling without an infrastructure team choose Retell; engineering teams with data residency constraints choose Vocode.
Full Retell AI vs Vocode comparisonImplementation & onboarding
- Setup time
- A working agent in a day. Production readiness takes weeks, dominated by flow design, testing against realistic audio, and defining escalation behavior.
- Learning curve
- Moderate. The flow builder is approachable, but designing conversations that hold up when callers say unexpected things requires iteration and listening to failures rather than reading documentation.
- Onboarding
- Self-serve with documentation, templates, and community resources, with support on larger arrangements.
- Migration notes
- Flows and prompts do not transfer between vendors. Keep conversation logic documented independently, and when migrating, rebuild and retest rather than translating, because differences in turn taking and interruption behavior change how the same script performs on a real call.
Platform, API & security
- Platforms
- REST API and SDKsWeb dashboard with flow builderTelephony integrationWeb calling
- API
- Full API for agents, calls, numbers, batch campaigns, and results, with function calling during conversations and webhook delivery of transcripts and structured data.
- Compliance
- GDPRCCPASOC 2HIPAA support on qualifying arrangementsTCPA considerations for outbound
- Data residency
- US-centric with enterprise arrangements for specific requirements.
- SSO
- Available on enterprise plans.
- Security notes
- Recordings and transcripts hold personal and often sensitive data; retention, access control, and disclosure that the caller is speaking with an AI should be configured deliberately given tightening requirements in several jurisdictions.
Support & resources
- Channels
- Documentation and developer communityEmail supportDedicated support on enterprise plans
- Documentation
- Practical documentation covering flow design, function calling, telephony, and testing, with examples for common deployments.
- Community
- Active developer and agency community, with substantial shared practice around flow design and evaluation.
Company
- Founded
- 2023
- Headquarters
- San Francisco, California, United States
- Ownership
- Private, venture-backed
- Employees
- ~50 (est. 2026)
- Funding
- Raised venture funding including a Series A round.
Timeline
- 2023Founded to make production-grade AI phone agents accessible to product teams.
- 2024Introduces the conversation flow builder, moving beyond prompt-only agent design.
- 2025Adds batch calling, testing, and success evaluation as deployments move into operations.
- 2026Established as a leading platform between raw voice APIs and no-code receptionist products.
Integrations
- Twilio
- Cal.com
- Google Calendar
- HubSpot
- Salesforce
- Make
- n8n
- Zapier
Frequently asked questions
10 questionsWhat is Retell AI?
Retell AI is a platform for building and running AI voice agents that handle inbound and outbound phone calls. It provides a low-latency conversation engine, a visual flow builder for controlling agent behavior, telephony, function calling, batch campaigns, testing tools, and post-call analysis.
How much does Retell AI cost?
Pricing is per minute with components broken out across voice, model, transcription, and telephony, starting from roughly $0.07 per minute combined and falling with volume commitments. Free credits are available for testing, and telephony rates vary by destination country.
Retell AI vs Vapi: which should I choose?
Vapi offers more composability, letting you select and swap every provider in the pipeline. Retell tends to be smoother out of the box with a stronger flow builder and more production tooling. If provider control matters most, choose Vapi; if reaching production quickly matters most, Retell usually gets there faster.
Is Retell AI no-code?
Not quite. The visual flow builder means much of the agent's behavior can be configured rather than coded, but production deployments involve function integrations, webhooks, and testing that assume technical capability. Genuinely non-technical buyers should look at no-code voice platforms instead.
What is a conversation flow?
A graph of nodes, each with its own instructions, required data, and permitted transitions, that constrains what the agent does at each stage of a call. It exists because a single prompt produces improvisation, and improvisation on a customer call is a liability rather than a feature.
Can the agent book appointments?
Yes, through function calling into calendar systems during the conversation, so availability is checked and the booking is made while the caller is on the line. This is one of the most common deployments, particularly for clinics, home services, and professional practices.
What happens when the agent cannot help?
It transfers to a human with context, provided you configure the escalation conditions. Designing those triggers well is one of the more important parts of a deployment, because an agent that persists when it should hand over is the fastest way to annoy a customer.
Does it detect voicemail on outbound calls?
Yes, which matters for campaign economics and for the caller experience. Without it, agents deliver conversations to answering machines, wasting minutes and producing confusing recordings for the recipient.
Do I have to tell callers they are speaking to an AI?
Increasingly yes. Disclosure requirements exist or are emerging in several jurisdictions, and expectations are tightening independently of law. Building disclosure into the opening is both the safer position and, in practice, less damaging to trust than being discovered.
How well do voice agents handle real calls?
Well for bounded conversations with a clear objective, such as booking, confirming, or qualifying. Less well with strong accents, poor line quality, or open-ended discussion. Evaluate using recordings of real calls under realistic conditions rather than a scripted demonstration, since demos never include the hard cases.
Editorial verdict
Retell AI has aimed at the gap that matters commercially: between a voice API that requires you to build everything around it and a no-code product that cannot be customized. The flow builder gives structure without demanding that every branch be written in code, and the surrounding tooling, batch calling, voicemail detection, simulated testing, success evaluation, is what separates a demo from something you can put in front of customers. Cost is legible and low enough that the arithmetic against staffed calling is easy. What remains hard is the part no vendor solves: designing conversations that survive real callers, deciding when to escalate, and navigating a legal environment around automated calling that is still moving. Approached with that discipline, it is one of the strongest choices in the category.
Written by the SaaSTracker editorial team. Awards, when shown, are judged against the published criteria in our methodology.