Bland AI
Enterprise Voice Agents, Self-Hosted
A voice agent platform built around what most tools skip: testing, simulation and post-call scoring. Pay as you go from about $0.07 a minute.

Retell AI is a platform for building AI voice agents that handle phone calls — inbound support, outbound follow-ups, appointment reminders, qualification. On the surface it belongs to the same category as Vapi and Bland. What separates it is where the product spends its attention: not on getting an agent talking, which everyone can now do, but on finding out whether the agent is any good before it talks to your customers.
That emphasis reflects the actual failure mode of this technology. Building a voice agent that works in a demo takes an afternoon. Discovering that it mishandles one caller in twenty, and that the twentieth caller is the one asking for a refund, takes three months of complaints. Retell ships simulation, A/B testing and automatic post-call scoring as first-class parts of the product rather than as an enterprise upsell.
The company was founded in 2023, went through Y Combinator's Winter 2024 batch, and raised a $4.6 million seed round led by Alt Capital. It now reports handling more than 50 million real-time calls a month and was named to Wing VC's Enterprise Tech 30 list for 2026 — unusually fast growth off an unusually small round.
A single-prompt agent is exactly what it sounds like: describe the job in plain language, attach some tools, take calls. It is quick and it is fine for narrow tasks where every path through the conversation looks roughly the same.
A conversation flow agent is a node-based diagram you assemble by dragging boxes — this question, then that branch, then this API call. It exists because prompts stop being reliable exactly when a call becomes valuable. Verifying identity before discussing an account, taking a payment, following a script a regulator will later read: these need a path the agent cannot improvise its way out of. Most teams end up with both kinds running side by side.
You can run an agent against simulated callers, replay awkward scenarios, and A/B two versions against each other rather than guessing which prompt is better. Considering how expensive a bad phone conversation is compared to a bad chat message, it is strange that this is a differentiator at all, but it is.
After the call, Retell scores it. Success and sentiment come built in, and you can define your own fields — whether the caller was offered the discount, whether the appointment was confirmed — and push them straight into your CRM. That turns a pile of recordings into something a manager can actually review on a Monday morning.
A knowledge base is built by crawling your website or uploading documents, and the agent retrieves from it mid-call rather than carrying everything in its prompt. Tools reach your own systems, and integrations cover Salesforce, HubSpot, Dynamics 365 and GoHighLevel out of the box. There is an API, SDKs and an MCP server, so the platform is reachable from an AI coding agent as well as from your application.
Calls arrive on Retell-managed US and Canadian numbers, through your own carrier over SIP, from a widget on your website, or through the web SDK inside your product.
Retell publishes its pricing as components rather than as a single number, which is more honest than a headline rate and slightly more work to read. The all-in range runs from about $0.07 to $0.31 a minute, and where you land inside that range is almost entirely a question of which language model you attach.
| Item | Price | Notes |
|---|---|---|
| All-in voice agent | $0.07–$0.31 per minute | The spread is the language model |
| Retell voice infrastructure | $0.055 per minute | The platform's own share |
| Text to speech | $0.015 per minute | Platform voices |
| Language model | $0.003–$0.16 per minute | Your choice, your bill |
| Telephony | about $0.015 per minute | US example; bring your own carrier if cheaper |
| Chat agents | from $0.002 per message | Same agent, text channel |
| Concurrency | 20 calls included | $8 per extra line per month |
| Free credits | $10 on signup | No card required |
| Knowledge base retrieval | $0.005 per minute | Charged on calls that use it |
| PII removal | $0.01 per minute | Add-on |
| AI quality assurance | $0.10 per minute | First 100 minutes free |
| Branded caller ID | $0.10 per outbound call | Displays your name on the handset |
| Enterprise | Custom | Dedicated servers, no concurrency cap, 24/7 support |
Two numbers deserve attention before you build a business case. Twenty concurrent calls sounds generous until a campaign goes out and every recipient calls back within the same twenty minutes. And quality assurance at $0.10 a minute can cost more than the conversation it is scoring, so it is a sampling tool, not something to leave on across all traffic.
If somebody senior will ask what percentage of calls went well and expect a real answer, the testing and scoring layer is the reason to choose this over a platform where you would have to build that yourself.
Reminders, follow-ups, renewals and win-backs are repetitive, measurable and easy to A/B. Branded caller ID matters more here than anywhere else, because an unknown number in 2026 goes unanswered.
Per-minute pricing with no platform fee, plus CRM integrations that clients already use, makes it practical to run many small agents rather than one large one. The GoHighLevel integration is a direct signal about who is already doing this.
If your answers already live in a help centre, the knowledge base turns that into an agent without rewriting anything into prompts. The quality of the result tracks the quality of the documentation more than the quality of the model.
No, but it starts with $10 in credits and no card, which at typical rates is a couple of hours of calls — enough to build something real and hear how it sounds. After that it is pay as you go with no monthly platform fee.
Less than with most alternatives. The flow builder is visual and the knowledge base takes a website URL. You will still want an engineer to connect tools to your own systems, which is where an agent stops being a phone answering machine.
Vapi is infrastructure that assumes you will build the surrounding tooling. Retell ships that tooling — flow builder, simulation, scoring, CRM sync — and gives up some of the flexibility in exchange. If you have an engineering team that wants control, Vapi. If you have an operations team that needs to run agents, Retell.
Yes — language support follows the speech and voice models you select rather than being a platform limit. Managed phone numbers are the constraint outside North America, not the conversation itself.
There is a PII removal add-on at $0.01 a minute that strips identifiers from transcripts and recordings. Whether that is sufficient depends on your regulator; for anything strict, the enterprise plan and a conversation with their team is the honest route.
Yes, via SIP trunking from your current provider. You can also buy managed numbers at $2 a month, or pay $10 a month for a verified number that shows your business name on the recipient's screen.
Retell AI is the voice platform that takes seriously the question of whether the agent is actually doing its job. Simulation before launch, scoring after every call, and a flow builder for the conversations that must not go off script — these are the things teams build themselves six months in, and here they arrive on day one.
The trade is flexibility for structure, and the cost model rewards attention. Read the component pricing properly, pick the smallest language model that can do the work, and treat quality assurance as a sample rather than a default. Do that and it is one of the more sensible ways to put an AI agent on a phone line that customers actually call.