AI voice agents
We put 6 AI voice agents through the same test. Retell AI came out on top with 3 out of 5. Here are all the results, what each tool costs and how we test them.
AI voice agents we tested
Best score first. Each one has a full review with the test results, prices and what users say.
Retell AI
3/5Read our review →PolyAI
3/5Read our review →Synthflow
3/5Read our review →Bland AI
2.5/5Read our review →Dapta
2.5/5Read our review →Vapi
2/5Read our review →Test results compared
Every tool got the same test, between October 20, 2026 to October 26, 2026.
| Tool | Score | Runs completed correctly out of 20 | Time to build each workflow from sign-up | How failures are handled (retries, alerts, human hand-off) | Cost per 100 runs at list prices | Integrations available for the workflow without code |
|---|---|---|---|---|---|---|
| Retell AI | 3/5 | First-pass 14/20 calls correct | 118 min/build; 1 workflow(s) | Retries recovered 4/6; alerts 7/10 | – | 7/10 required connections no-code |
| PolyAI | 3/5 | First-pass 15/20 calls correct | 104 min/build; 1 workflow(s) | Retries recovered 3/5; alerts 7/10 | – | 6/10 required connections no-code |
| Synthflow | 3/5 | First-pass 15/20 calls correct | 115 min/build; 1 workflow(s) | Retries recovered 4/5; alerts 7/10 | – | 6/10 required connections no-code |
| Bland AI | 2.5/5 | First-pass 14/20 calls correct | 118 min/build; 1 workflow(s) | Retries recovered 4/6; alerts 5/10 | – | 6/10 required connections no-code |
| Dapta | 2.5/5 | First-pass 13/20 calls correct | 128 min/build; 1 workflow(s) | Retries recovered 5/7; alerts 7/10 | – | 6/10 required connections no-code |
| Vapi | 2/5 | First-pass 12/20 calls correct | 128 min/build; 1 workflow(s) | Retries recovered 5/8; alerts 6/10 | – | 6/10 required connections no-code |
Want just the shortlist?
Our best-of list ranks every tool in this test, with what each one is best for.
What AI voice agents do
AI voice agents answer and place phone calls using speech recognition, a language model and synthetic voice, handling tasks such as booking, qualification and support without a human on the line. Platforms range from developer APIs to managed enterprise services.
Who uses them: Contact centres replacing IVR menus, Healthcare, insurance and finance firms with high call volume, Sales teams doing outbound qualification, Agencies and developers building voice products, Small businesses covering missed calls.
Price range: Self-serve platforms charge per minute with a free credit (Vapi 5 USD, Retell 10 USD, PolyAI model API 100 USD for a limited time). Published all-in rates run about 0.07 to 0.31 USD per minute (Retell), 0.12 to 0.14 (Bland), with Vapi's hosting at 0.05 plus provider costs. Synthflow and PolyAI's platform are quote-only. Paid tiers add concurrency and compliance: Vapi Core 29, Pro from 999, HIPAA 2,000 USD a month; Bland Build 299 a month.
What to check before you buy
Real per-minute cost. Headline rates often exclude speech, model, voice or telephony; compare the all-in cost per minute (Vapi calculator about 0.08 to 0.13 USD, Retell 0.07 to 0.31 USD, Bland 0.12 to 0.14 USD).
Latency and conversation quality. Delay and interruption handling decide whether callers hang up; vendor claims range from under 100 ms to about 600 ms and are unaudited.
Build style. API-first (Vapi), visual builder (Retell, Synthflow) or managed delivery (PolyAI, Bland enterprise) should match who maintains the agent.
Compliance. SOC 2, HIPAA with a BAA, PCI and data residency; check if the BAA is self-serve or enterprise only (Vapi 2,000 USD a month add-on, Bland Enterprise only).
Data use and retention. Check whether calls train models and how long data is kept; wording often differs between terms, privacy page and marketing.
How we test AI voice agents
The test for this category: Build a phone booking agent; 20 scripted test calls with interruptions and accents; success rate and latency.
Every tool builds the same three workflows (for example: answer support questions from a help centre, qualify a lead and book a call, pull data from a website into a sheet). Each workflow is run 20 times with the same inputs.
Each tool is tested on a paid plan, and its review shows the plan, the price and the test dates. The overall score is the average of the five parts on the right, rounded to the nearest half star. See the full method for AI agents and automation tools.
What we measure
- Runs completed correctly out of 20
- Time to build each workflow from sign-up
- How failures are handled (retries, alerts, human hand-off)
- Cost per 100 runs at list prices
- Integrations available for the workflow without code
The five parts of the score
Frequently asked questions
What is an AI voice agent?
Software that holds phone conversations using speech-to-text, a language model and text-to-speech, for inbound and outbound calls.
How much do AI voice agents cost per minute?
Published rates span about 0.07 to 0.31 USD per minute all in (Retell), 0.12 to 0.14 (Bland), Vapi 0.05 hosting plus about 0.03 to 0.08 in provider costs; Synthflow and PolyAI's platform are by quote.
Which platform is best for developers?
Vapi and Retell are the main developer-friendly self-serve options, with Vapi being API-first.
Are AI voice agents legal for cold calls?
Vendors place responsibility on the customer: Retell's terms require documented consent, AI disclosure, Do-Not-Call scrubbing and calling-hour limits; Vapi cites the TCPA. This is not legal advice.
Do I need HIPAA?
For health data you need a signed BAA: Vapi charges 2,000 USD a month for HIPAA mode, Bland offers BAAs on Enterprise, Retell is not a business associate unless a BAA is signed.
More in AI video and audio tools
Page updated November 4, 2026.
Make your case
Think we got a review wrong?
If you use a tool we reviewed, or work for the company, email us with what you saw. We'll look at the evidence and update the review if it changes our findings.
Want your company reviewed?
Send us your product, website and the category it fits. We'll test it the same way as every other tool. It's free, and we can't promise a date or a score.
