AI phone agents can handle routine inbound scheduling, qualification, routing and questions when configured and tested. Reliability varies with integrations, speech conditions and escalation design; assess production performance before replacing existing call handling.
An AI phone agent is voice AI software that answers your business’s inbound phone calls in live conversation - not a menu of pressed options, but a model that listens, decides what to say next, and can book, route or escalate while no human ever joins the call.
Quick answer: Retell offers dashboard and developer workflows from $0.07/minute, with a published $0.07–$0.31 range and additional configurable costs. Ultravox is speech-native infrastructure from $0.05/minute plus applicable SIP and telephony. PolyAI provides custom enterprise deployments. These are use-case shortlists rather than independently tested rankings.
Evaluation methodology
This comparison covers nine shortlisted voice AI platforms, including complete-agent services and infrastructure requiring development. It considers documented inbound support, total operating cost, transfer, booking, language coverage, consent controls and failure handling. Prices are USD as of October 11, 2026; quote-only products are identified. No hands-on comparative testing is claimed.
Cresta, Observe.AI and SoundHound are outside this selected shortlist; missing self-serve pricing does not establish irrelevance. Cresta’s AI Receptionist supports inbound answering and scheduling, but a public per-minute rate was not established here. Salesforce acquired Tenyx on September 13, 2024, so it is not evaluated as an independent subscription. Sierra and Regal remain included with quote-based pricing.
The 9 phone agents at a glance
| Platform | Best for | Standout feature | Starting price | Free tier / trial |
|---|---|---|---|---|
| Retell AI | Inbound booking, qualification, overflow | Conversation Flow with node-level call control | $0.07–$0.31/min pay-as-you-go | $10 promotional credits |
| Bland AI | High-volume outbound and inbound at scale | Parallel and async calling modes | $0 monthly platform fee + $0.14/min Start; Build $299/mo + $0.12/min | Introductory credits advertised |
| Vapi | Developers building a custom agent | Provider costs passed through with no markup | $0.05/min hosting + model costs | $5 free credits |
| Ultravox (ex-Fixie.ai) | Speech-native developer infrastructure | Audio-native model, open weights | $0.05/min incl. TTS, up to 5 concurrent | 30 introductory minutes; $0 platform fee |
| PolyAI | Large enterprises, regulated verticals | Audio-native, TTS-optional Dialog-RSN-1 | Custom, per-minute | No public trial |
| Twilio Conversation Relay | Teams already living in Twilio | Native fit with Programmable Voice | $0.07/min + voice rates | $15 new-account credits |
| Parloa | EU enterprise contact centres | Consumption pricing tied to use-case complexity | Custom, per-minute | No public trial |
| Sierra | Outcome-priced enterprise agents | Agent OS with approval controls | No public pricing page | No public trial |
| Regal | Regulated outbound and contact-centre ops | Deep compliance and QA tooling | No public pricing page | No public trial |
Sierra, Regal, PolyAI and Parloa require quotes for a current usable rate. Vapi, Ultravox and Twilio Conversation Relay are infrastructure: implementation and additional service costs differ from turnkey call handling.
The 9 phone agents, reviewed
Retell AI
Retell AI is a developer-and-dashboard voice agent platform that terminates real phone calls and lets you build agents either as a single prompt or as a branching conversation flow.
Retell suits teams seeking booking, qualification and overflow with dashboard configuration or developer integration. Its single-prompt and branching-flow approaches have different control and maintenance requirements.
That dual mode is the genuinely useful thing. A hair salon wants free-form chit-chat and a calendar lookup; a mortgage lender wants a deterministic qualification sequence. Same platform, different risk profile.
Pricing. Pay-as-you-go runs from $0.07 to $0.31 per minute depending on the model stack you pick. Testing bills separately - text tests per message, voice tests at production per-minute rates.
Telephony and CRM. Retell-managed US and Canadian numbers are purchasable in-dashboard with no telephony account needed. You can also import numbers via SIP trunking, and the docs document specific paths for Twilio elastic SIP, Telnyx, Vonage, Avaya Aura, Genesys Cloud, Five9 and Amazon Connect. CRM integrations include HubSpot, Salesforce, Zendesk, Cal.com, Calendly, GoHighLevel, Zoho and Microsoft Dynamics 365.
Languages. A dedicated docs page maps each supported language to the ASR and TTS providers that carry it, which is the only honest way to evaluate multilingual support.
Transfer. The transfer tool supports cold, warm and agentic-warm modes, and a separate agent-transfer tool hands a call between two AI agents mid-conversation.
Limitation. The flexibility is a liability if nobody owns it. Conversation flows with global nodes, subflows and Flex Mode are powerful and genuinely easy to tangle. Retell also published a deprecation notice showing legacy monthly billing switching to prepaid credits on September 30, 2026 - migration is real work, not a settings toggle.
Onboarding and scale. Retell advertises $10 promotional credits and 20 concurrent calls on pay-as-you-go. Enterprise support and implementation options require sales confirmation; “Success Packages” is Vapi terminology, not the Retell plan structure.
Inbound configuration. Retell documents inbound webhooks, per-caller context, number binding and routing. Verify after-hours settings, calendar booking and transfer recovery using the actual integration stack; documented features do not establish superiority across all nine products.
Retell’s automated QA should supplement human review. Its optional AI quality assurance add-on is listed at $0.10/minute after an introductory allowance: reviewing 1,000 billed minutes would add $100 at that rate. The $0.07–$0.31 headline is not an all-inclusive bill; managed telephony, knowledge-base use, denoising, guardrails and other options can raise total cost.
Bland AI
Bland AI is a voice agent platform built for volume, offering parallel and asynchronous calling modes alongside conventional inbound and outbound agents.
Bland’s differentiator is throughput rather than conversational polish. Parallel mode lets multiple calls run at once for outbound campaigns, and async mode handles long or complex workflows that do not need a live conversation at all.
Pricing. Bland Start has no monthly platform fee and costs $0.14/minute, with introductory credits. Build adds $299/month and $0.12/minute usage. Ignoring other costs, the usage saving offsets the platform fee at 14,950 billed minutes. Start caps concurrency at 10 and daily calls at 100; Build lists 50 concurrent and 2,000 daily calls. Telephony and applicable transfer minutes are additional; operational limits can require Build before the cost crossover.
Architecture. August 2026 shipped adaptive resumption and node-scoped interruptibility, which is the vocabulary you want for handling a caller who interrupts mid-sentence. Interruptibility scoped per node matters more than it sounds: it lets you allow interruption during a greeting but not while confirming an appointment, which is exactly where a caller barging in causes real damage.
Limitation. Compare concurrency, daily call limits, implementation effort and fees on the intended tier. The Build platform fee is not a mandatory charge for every small Start deployment.
Fit. Low-volume users should compare Start against other configured deployments; high-volume campaigns need a closer review of concurrency, throughput and permissions.
Buying guidance. Model minutes and operational limits together. The 14,950-minute crossover compares only Start and Build platform/usage charges, not all purchasing considerations.
Vapi
Vapi is an open infrastructure layer for voice agents that passes through STT, LLM and TTS provider costs at cost with no platform markup.
Named here once as context for the buy-versus-build decision, not as a phone-agent product for a business that wants an answer on the phone by Friday. The economics that matter: $0.05/min Vapi hosting plus model costs, and the pricing calculator shows Deepgram transcription at $0.0095–$0.0099/min, OpenAI intelligence at $0.0077–$0.0452/min, and ElevenLabs voice at $0.0146–$0.0238/min. This selected component stack illustrates approximately $0.082–$0.129/minute. Actual cost varies with providers, model use, telephony, support and other charges; it is not a universal real-call rate.
Support packages start free, then $29/month Core, then Pro at $999/month minimum. Pro is priced at 10% of hosting fees with a $999/month minimum; include support costs in the model.
Limitation. With no Success Package you get four concurrent calls and 14-day raw data retention. Fine for a prototype. Not fine for a clinic.
Twilio alternative. Conversation Relay supplies communications infrastructure at $0.07/minute with Programmable Voice charged separately. External models and other services may create further costs; compare the complete architecture with Vapi.
Ultravox (formerly Fixie.ai)
Ultravox is speech-native infrastructure from $0.05/minute after introductory minutes, with five concurrent calls on pay-as-you-go and separate deployment costs.
Ultravox lists $0.05/minute after 30 introductory minutes and five concurrent calls on pay-as-you-go. Pro is $100/month at the displayed annual rate. SIP, telephony and separately metered services add to deployment cost; $0 platform fee does not mean free calls.
Architecture. Ultravox’s audio-native model can retain acoustic context and avoid some separate processing stages. Vendor benchmark results do not establish accuracy or latency for every phone deployment; measure the complete application.
Limitation. This is infrastructure. Built-in telephony integrations with major providers are listed, but you wire the CRM yourself. And Fixie.ai has rebranded to Ultravox, which means older blog posts and roundups referencing “Fixie” now describe a product name you will not find on the site.
Naming. Fixie rebranded to Ultravox. Use current documentation and pricing rather than assuming older Fixie offers remain available.
PolyAI
PolyAI is an enterprise voice agent platform that builds and operates agents for large contact centres, now shipping an audio-native Dialog-RSN-1 model that can run without a separate TTS step.
PolyAI positions itself as the enterprise end of the market - airlines, utilities, healthcare, financial services - and its pricing page confirms per-minute ongoing use bundled with proactive maintenance, 24/7 support and a 99.9% uptime SLA on phone lines.
Architecture. Dialog-RSN-1, released in late July 2026, is the notable move: audio-native and TTS-optional, cutting the latency penalty of a cascaded pipeline. Treat architecture and latency statements as vendor positioning; test the full deployment.
Limitation. No number is published. You cannot budget a per-minute cost from the public site, and there is no trial, so this is a conversation before it is a purchase. For most businesses outside the enterprise tier that ends the evaluation early - which is a fair outcome, not a flaw.
What you are actually buying. PolyAI’s positioning is not a voice agent you configure; it is a partner who will design, deploy and operate the agent for you, with proactive maintenance and 24/7 support priced in. That is a different kind of purchase, and for a large airline or utility it is the right one. For a business with one number and forty calls a week, the operational overhead dwarfs the benefit.
Twilio Conversation Relay
Twilio Conversation Relay is a per-minute voice AI relay that plugs AI voice agents into Twilio’s Programmable Voice stack, priced at $0.07/minute plus your normal voice costs.
Twilio Conversation Relay provides the real-time communications layer for voice AI applications, not an all-inclusive AI receptionist. Programmable Voice is billed separately, and the architecture may also require external language models, speech services and other components.
Pricing. $0.07/minute for the relay itself, with voice costs billed separately at Twilio’s standard Programmable Voice channel rates. Twilio’s pricing page also states the total works out to $0.07/min starting price with voice billed on top - so budget more than seven cents.
Context. Conversation Relay sits alongside Conversation Orchestrator, Conversation Memory and Conversation Intelligence, all metered separately. New accounts get $15 in credits.
Limitation. The $0.07 excludes voice minutes, and the broader Conversations stack meters independently on characters, profiles and recalls. Watch for the metering model, not the headline rate. There is also no free tier beyond those credits.
Parloa
Parloa is a European enterprise conversational-AI platform using consumption-based pricing, per minute for voice and per interaction for chat, with rates that rise with use-case complexity.
Parloa’s pricing philosophy is unusually explicit and worth quoting, because it explains why its number is not comparable to Retell’s. Simple interactions stay cheap; complex automation prices in “the reasoning, context and coordination required.” One commitment covers a volume, applied across every use case.
There is one line on that page that tells you Parloa’s actual bias: your billing model does not force the agent to protect a containment or resolution metric when a person should take over. That is a company choosing human escalation over its own metric.
Limitation. No public per-minute figure and no trial. The page is a description of a sales process - analyse volume, model the business case, design a custom model - rather than a price list.
Sierra
Sierra is an enterprise agent platform built around outcome-based reasoning, where agents take actions inside a company’s systems under approval controls rather than just answering.
Sierra is worth including for how it frames the problem: it markets agents that act and escalate under supervision, backed by a Ghostwriter tool that generates agents from natural-language instructions, plus an Agent Studio for grounding agents with knowledge, integrations and simulations.
Limitation. Sierra’s pricing requires vendor confirmation; no public per-minute rate is established here. Compare outcome definitions, usage, implementation and support rather than inferring quality from price transparency.
Regal
Regal is a contact-centre voice and compliance platform for regulated industries, pairing outbound and inbound agents with testing, QA and call-quality governance.
Regal’s differentiator is the unglamorous one: at scale in regulated verticals, the constraint is not whether the agent can book an appointment, it is whether you can prove what it did. That means QA tooling, audit trails and consent handling.
Limitation. Regal requires a quote for the required deployment. No public per-minute price or free allowance is established here; confirm support, QA and usage entitlements.
Routine calls and operational failure cases
AI can automate many routine calls effectively when workflows are bounded, integrations reliable and human escalation configured. Accuracy and resolution rates vary. Consequential, ambiguous and emotionally sensitive calls introduce additional risks.
Useful pilot scenarios include after-hours intake, booking, qualification, routing and bounded FAQs. Test noisy connections, interruptions, accents, language changes, overlapping speakers and repeated callers. A calendar failure must not produce an unconfirmed appointment; verify transfer recovery and safe fallback when tools fail.
Before committing, test the common call reasons using representative caller language and integrations. Measure completed tasks, failed bookings, incorrect answers, successful handoffs and cost. Vendor resolution percentages require cohort and methodology context; they are not a substitute for deployment-specific measurement.
Latency and conversational quality
End-to-end latency is one major factor alongside recognition accuracy, turn-taking, interruptions, prosody, model behaviour and network conditions. Cascaded systems offer modular control and provider substitution; audio-native systems can reduce some overhead and retain acoustic context. Actual latency and reliability depend on implementation.
Test on realistic phone connections, including interruptions and recovery. Model-level first-response figures do not establish the complete caller experience.
Disclosure, consent and call-handling obligations
Inbound answering has a different TCPA profile from automated outbound telemarketing. It can still implicate recording consent, AI disclosure, privacy, retention, authentication, security and sector-specific confidentiality. Payments, health information and other sensitive data need additional safeguards. Consult applicable recording laws.
The FCC treats AI-generated voices as artificial voices under the TCPA. Outbound requirements depend on call purpose, recipient and applicable consent rules; an inbound caller is not blanket permission for future automated marketing. Consult 47 CFR § 64.1200.
The supplied report identifies an October 1, 2026 FCC order and further proposed rulemaking on revocation. The exact primary order and applicable effective dates were not independently retrieved for this revision, so no proposed amendment is presented as an effective requirement. Verify final text, publication and effective dates before changing opt-out procedures.
Configure clear disclosure, recording permissions where required, opt-outs, limited data access and a tested human handoff. Technical vendor controls do not establish compliance for every deployment.
Which call types should go to a human first
A conservative escalation policy should route sensitive or consequential cases to a qualified person. Define the threshold for the actual business workflow.
- Medical, legal or financial advice. Clinical intake beyond scheduling, anything resembling a diagnosis, any legal opinion. The failure mode is not a bad answer - it is a confident, plausible, wrong one.
- The caller in distress. Self-harm, domestic violence, a safety-critical emergency. This needs a human who can act on what they hear, not a model optimising for containment.
- Escalating conflict. Third-party complaints, chargebacks, a customer who has already called twice. Set a rule: repeat caller, human.
- Complaints about money. Refunds, billing disputes, anything with a number attached that will end up in a dispute.
- Anything regulated or audited. Compliance-heavy verticals should treat AI as intake and routing only, never resolution.
The escalation rule matters more than the model. Retell supports cold, warm and agentic-warm transfers, and the agentic-warm mode exists because a blind cold transfer - dropping a caller into silence - is itself a failure. Configure a human handoff with context, a summary of what the caller wanted, and a stated wait expectation. If your configuration cannot do warm transfer, do not deploy it.
What transfer-to-human actually looks like in production:
- Warm transfer passes the caller and a summary, so the human opens the conversation mid-sentence. Best experience, most setup work.
- Cold transfer hands over a bare line. The caller has to repeat themselves. Some businesses cannot afford this; some can.
- Callback lets the caller choose to hold or be called back with a reference number. Under-rated, and the honest choice when nobody is free.
Two practical warnings. First, a transfer that arrives without context means your human agent starts the call already behind, having just interrupted a paying customer. Second, check what happens when every human is busy - if the fallback is an apology loop, you have built a worse experience than voicemail.
Set your threshold before you go live, not after your first complaint. A conservative escalation policy: transfer on any request for a refund, on any mention of a lawyer or regulator, on any repeat caller within 24 hours, and on any sentiment spike the platform flags. Write those four rules into the agent’s instructions and test that they actually fire.
Verdict
AI can handle routine inbound scheduling, qualification, routing and FAQs when properly configured and tested. Reliability varies by use case, integration and escalation design; measure production performance before replacing call-handling workflows.
Retell is a shortlist option for dashboard-led deployments, Ultravox for speech-native development and Twilio Conversation Relay for teams building on Twilio. PolyAI and Parloa support enterprise procurement. Compare complete costs and operational requirements; these are not independently tested reliability rankings.
Use a controlled pilot with human review and a safe fallback. Confirm appointments only after successful tool responses, and verify that escalation triggers work on actual phone connections.
Sources
- Retell AI - Pricing (accessed October 10, 2026) - https://www.retellai.com/pricing
- Retell AI - Documentation index,
llms.txt(accessed October 10, 2026) - https://docs.retellai.com/llms.txt - Retell AI - Transfer Call tool documentation (accessed October 10, 2026) - https://docs.retellai.com/build/single-multi-prompt/transfer-call.md
- Retell AI - Inbound call webhook documentation (accessed October 10, 2026) - https://docs.retellai.com/features/inbound-call-webhook.md
- Retell AI - Supported languages by provider (accessed October 10, 2026) - https://docs.retellai.com/build/language-support.md
- Retell AI - Handle background speech and noise (accessed October 10, 2026) - https://docs.retellai.com/build/handle-background-noise.md
- Retell AI - Do-not-call requests (accessed October 10, 2026) - https://docs.retellai.com/build/do-not-call.md
- Retell AI - Branded call application (accessed October 10, 2026) - https://docs.retellai.com/build/telephony/branded-call.md
- Retell AI - Testing pricing (accessed October 10, 2026) - https://docs.retellai.com/test/testing-pricing.md
- Retell AI - Custom telephony and SIP (accessed October 10, 2026) - https://docs.retellai.com/deploy/custom-telephony.md
- Retell AI - Deprecation notice: legacy billing ends September 30, 2026 (accessed October 10, 2026) - https://docs.retellai.com/deprecation-notice/2026/09-30_legacy_billing.md
- Retell AI - Agent Transfer tool (agent swap) (accessed October 10, 2026) - https://docs.retellai.com/build/single-multi-prompt/transfer-agent.md
- Bland AI - Pricing (accessed October 10, 2026) - https://www.bland.ai/pricing
- Bland AI - Change log, August 3, 2026: adaptive resumption and node-scoped interruptibility (August 3, 2026) - https://docs.bland.ai/changelog/08_03_2026
- Bland AI - Change log index (accessed October 10, 2026) - https://docs.bland.ai/changelog
- Vapi - Pricing (accessed October 10, 2026) - https://vapi.ai/pricing
- Ultravox (formerly Fixie.ai) - Pricing (accessed October 10, 2026) - https://www.ultravox.ai/pricing
- PolyAI - Pricing (accessed October 10, 2026) - https://www.poly.ai/pricing
- PolyAI - Pricing and integrations overview (accessed October 10, 2026) - https://www.poly.ai/pricing#integrations
- Twilio - Conversational AI solutions pricing, current as of August 2026 (accessed October 10, 2026) - https://www.twilio.com/en-us/products/conversational-ai/pricing
- Twilio - Conversation Relay product page (accessed October 10, 2026) - https://www.twilio.com/en-us/products/conversational-ai/conversationrelay
- Twilio - Programmable Voice pricing, United States (accessed October 10, 2026) - https://www.twilio.com/en-us/voice/pricing/us
- Parloa - Consumption-Based Pricing for AI Agents (accessed October 10, 2026) - https://www.parloa.com/pricing
- Regal - Official site, no public pricing page (accessed October 10, 2026) - https://reg.al
- SoundHound AI - OASYS platform overview (accessed October 10, 2026) - https://www.soundhound.com/oasys
- OpenAI - API pricing reference (accessed October 10, 2026) - https://platform.openai.com/docs/pricing
- Twilio - 2026 Gartner Magic Quadrant for Communications Platform as a Service, by Lisa Unden-Farboud et al. (May 18, 2026) - https://www.twilio.com/en-us/report/gartner-mq-cpaas-2026
- Retell AI - Cal.com agent functions, live booking during a call (accessed October 10, 2026) - https://docs.retellai.com/integrations/cal-com-functions.md
- Retell AI - Inbound calls: binding agents, routing and overflow (accessed October 10, 2026) - https://docs.retellai.com/deploy/inbound-call.md
- Retell AI - Conversation Flow overview (accessed October 10, 2026) - https://docs.retellai.com/build/conversation-flow/overview.md
- Retell AI - Flex Mode (accessed October 10, 2026) - https://docs.retellai.com/build/conversation-flow/flex-mode.md
- Retell AI - AI QA for call quality, hallucination and latency scoring (accessed October 10, 2026) - https://docs.retellai.com/ai-qa/overview.md
- PolyAI - Technology page (accessed October 10, 2026) - https://www.poly.ai/technology
- Cresta - AI Receptionist product page (accessed October 10, 2026) - https://www.cresta.com/ai-receptionist
- Additional primary reference (revision October 11, 2026) - https://www.salesforce.com/news/stories/salesforce-signs-definitive-agreement-to-acquire-tenyx/
- Additional primary reference (revision October 11, 2026) - https://www.rcfp.org/reporters-recording-guide/
