Synthflow, Vapi, Retell, Bland: what the voice AI platforms get right, and when to build instead

We ran a production voice AI product for long enough to learn where every platform's demo stops and the real work begins. If you are searching for a Synthflow alternative, or comparing Vapi with Retell, or wondering why your Bland agent behaves in front of customers, this note is the conversation we would have with you on a call.

The first thing to understand is that the alternatives question is rarely about which platform is best. It is about which layer of the problem you want to own. Voice AI in 2026 is three markets wearing one name, and most disappointment comes from buying in the wrong one.

The three layers people are actually comparing

• Hosted agent platforms. Synthflow, Vapi, Retell, Bland, Air and a long tail of others. You write a prompt, pick a voice, connect a number, and you have an agent in an afternoon. You pay per minute, and the platform owns the stack.
• Contact centre suites. Amazon Connect, Twilio Flex, Five9, Genesys, NICE. These are staffing and routing systems for teams of human agents that have been growing AI features. Comparing Amazon Connect with Twilio Flex is really a question about your contact centre, not about AI.
• Building on primitives. A telephony layer such as Twilio, a realtime voice model, and your own logic in the middle. Slower to first call, and the only option where you fully own the behaviour, the data and the economics.

What the hosted platforms genuinely do well

Credit where it is due. For a single use case, answering after hours, qualifying inbound enquiries, confirming appointments, a hosted platform is the fastest route to a working agent, and the per minute pricing is fair at low volume. If your requirement fits inside a prompt and a couple of integrations from their gallery, use one, and spend your energy on the script and the failure cases rather than on infrastructure.

Where they run out of road

The pattern in almost every alternatives search is the same. The demo went well, the pilot went fine, and then one of four walls appeared.

• The integration ceiling. The agent needs to check a real diary, write to your CRM the way your team does, or follow a business rule with exceptions. Prompt level control stops being enough, and webhook gymnastics begin.
• Per minute economics. Platform pricing stacks the model, the voice, the telephony and the margin into one rate. At meaningful call volume you are paying several times the cost of the underlying parts.
• Behaviour you cannot reach. Interruption handling, when to hand off to a human, what happens on a bad line, how the agent behaves when the caller is angry. On a hosted platform these are settings. On your own stack they are code you control.
• Compliance. If callers discuss health information, the question is not whether the platform mentions HIPAA on its pricing page. It is whether every layer, telephony, transcription, model and storage, will sign a BAA and where recordings actually live. Ask for the documents before the pilot, not after.

What building on primitives looks like now

This changed in the last two years, and it is why we now steer more projects this way. The realtime voice models have become good enough to hold a natural conversation with proper interruption handling out of the box. Our current favourite is Grok voice, which is the most natural we have deployed to date, though the model layer moves quickly, so we architect every build so the model can be swapped without touching the rest of the stack.

A build on primitives typically means your own Twilio number and call routing, a realtime model carrying the conversation, and your logic in between: the diary it books into, the CRM it writes to, the rules for when it hands a caller to a person, and transcripts landing where you can actually read them. You own the number, the data and the behaviour, and the per minute cost is the sum of the parts rather than a platform's bundle.

The honest downside: it is a build, not a signup. There is real engineering in barge in behaviour, latency, failure handling and the unglamorous telephony edges. That is exactly the experience we took from running our own voice product, and it is why the second system is much cheaper to get right than the first.

Questions to ask whichever way you go

• What happens when the agent does not know? The transfer path to a human matters more than the happy path.
• Where do recordings and transcripts live, who can read them, and can you delete them?
• What does a call actually cost at your expected volume, all layers included?
• Can you take your number and your call history with you if you leave?
• If callers discuss health or payment information, who signs what?

If you are still at the stage of deciding whether any of this is worth building, our note on whether your business actually needs a custom AI solution applies to voice as much as to anything else. Repeated calls, a person in the loop for the hard cases, and a structured place for the outcome to land: the same three tests.