AI Voice Agents for Business: 2026 Costs, Rules and Risks

An AI voice agent can answer your phones for $0.07 to $0.31 a minute on published pricing. Here is what a full minute costs, how it compares with a receptionist, and which US and UAE rules apply before it calls anyone.

Head of AI

· 17 min read

A dark green rotary telephone against a warm red glow

Key takeaways

  • Retell's published pricing puts an AI voice agent at $0.07 to $0.31 a minute; on Vapi, the $0.05 platform fee plus speech-to-text, model, voice and a Twilio line comes to $0.09 to $0.14 (checked October 26, 2026).
  • A US receptionist earned a 2025 median of $18.27 an hour, per the BLS, about $0.30 per paid minute whether or not the phone rings; Ruby's live answering service costs $250 for 50 minutes and $1,725 for 500, or $5.00 and $3.45 a minute.
  • Since the FCC's February 2024 ruling, AI-generated voices count as artificial under the TCPA, so outbound AI calls need prior express consent, and people can sue for $500 per violation, tripled if willful.
  • UAE rules allow marketing calls only from 9 a.m. to 6 p.m., from local numbers registered to the company, with consent asked at the start for automated calls; The National reported more than AED 19 million in fines and 9,433 numbers disconnected since August 2024.
  • Deepgram's Nova-3 speech model lists 17 Arabic locale codes, including Emirati, Saudi and Egyptian Arabic, but its Flux conversational model supports neither Arabic nor Urdu as of October 26, 2026.

Yes: an AI voice agent can answer your business phone in 2026, and on Retell's published pricing a minute costs $0.07 to $0.31, depending on the model and voice. Answering is the low-risk part: in a 2024 proposal, the FCC noted that the TCPA's requirements "do not extend to technologies used to answer inbound calls." Since the FCC's February 2024 ruling, an AI voice counts as "artificial," so outbound AI calls need prior consent, and in the UAE an automated marketing call is allowed only if the consumer is asked for consent at the start of the call.

This guide is for owners and operations leads at clinics, real-estate agencies, home-services, logistics and relocation firms in the US and the UAE who miss calls or pay for after-hours answering. Inside: how a voice agent works, a dated price table, a comparison with human answering, US and UAE rules, languages and a launch checklist. We have not put a phone agent into production for a client yet. We have built AI call notes and an agent that answers customers in writing, and we say what carries over.

Can an AI receptionist answer your business phone?

Yes, for calls with a clear purpose. An AI receptionist answers questions from your information, books or moves appointments, takes messages and hands off to a person, around the clock and on several lines at once.

The bigger question is direction. Inbound calls come to you: after-hours and overflow answering, booking, new leads. Outbound calls go out: reminders, callbacks, follow-ups, sales. The US Telephone Consumer Protection Act (TCPA) restricts calls made with "an artificial or prerecorded voice," and UAE telemarketing rules govern marketing calls that companies make. Both target calls you place, not calls you answer.

Inbound and outbound AI calls and the rules that apply. FCC rules, US state laws and UAE telemarketing rules, checked October 26, 2026.

RuleInbound: the customer calls youOutbound: the agent calls
US consentThe FCC says the TCPA does not reach answeringPrior express consent; written consent for telemarketing
US hoursNoneSales calls 8 a.m. to 9 p.m., local time
UAEData law covers recordings and transcriptsPermit, registered numbers, 9 a.m. to 6 p.m., no DNCR numbers, consent at the start
RecordingTell the caller; 13 US states need everyone's consentThe same; UAE notice at the start
Main riskWrong answers, missed handoffs$500 per violation in US suits, UAE fines up to AED 150,000

Start with inbound: it carries the least legal risk, and your phone logs already show which calls go unanswered.

How does an AI voice agent work?

Calls reach the agent through a number from a carrier such as Twilio or a SIP trunk from your current phone system. Then one of two designs takes over.

The pipeline chains three models: speech-to-text (STT) transcribes the caller, a large language model (LLM) decides what to say and which tools to use, and text-to-speech (TTS) speaks. Each part can be tuned per language, and you get a transcript; the cost is delay.

Speech-to-speech models take audio in and return audio out. OpenAI pitches gpt-live-1 for "smooth interruption handling" and says it can "delegate reasoning and tool use to a backend agent." Google prices its Gemini Live models per minute of audio. The trade-off is control: fewer points to inspect and fix a bad answer.

Speed, interruptions and handoff

Callers expect quick replies. In a study of 10 languages, answers to yes-or-no questions started on average about 200 milliseconds after the question ended; the slowest language, Danish, averaged 469. A two-second pause after every sentence is far outside that. Tune turn detection, which decides when the caller is done, and barge-in, which lets them interrupt, on mobile networks and speakerphones in each language.

The value is in actions: book a slot, look up an order, create a lead in the CRM, text a confirmation, each with strict limits on what the agent may change. A warm transfer hands the caller to a person with a summary. After hours, the agent takes a message and gives a callback time your team can keep. After each call, the transcript, summary and actions go into the CRM.

We have built the after-call half. For Encore, a Dubai PR, media and events agency, we built a CRM and an AI call-notes tool that feeds it: Deepgram Nova-3 transcribes each call with speakers and timestamps, Claude writes the summary, decisions and tasks, and a follow-up email is drafted. Seventy client calls were transcribed and searchable within three weeks. It is not a phone agent, but the transcripts, summaries and CRM tasks carry over.

Encore logo
Case study · EncoreWebsite, CRM and AI call notes for a Dubai agency

How much does an AI phone agent cost per minute?

A full minute has five parts: platform fee, speech-to-text, language model, voice and phone line, and headline rates rarely cover all five. These are list prices, not our price list.

Published per-minute prices for AI phone agents and their parts. Vendors' pricing pages, checked October 26, 2026; full-minute totals are our math.

OptionPublished priceIncludedFull minute
Retell$0.07–$0.31Voice infrastructure $0.055 plus your chosen model and voice$0.11 with a $0.04 model; $0.125 with its line
Vapi$0.05 platform feeOther parts at provider cost$0.09–$0.14
Bland, Start plan$0.14Models and voice; phone line extraAbout $0.15 with a Twilio line
ElevenLabs Speech Engine$0.08Speech models; model and line not statedAsk the vendor
OpenAI gpt-live-1$0.05, billed per secondVoice layer; backend model and tools extra$0.05 plus extras
OpenAI gpt-realtime-2.1$32 per 1M audio tokens in, $64 outModel with speech in and outAbout $0.05 in new audio, rising with call length; plus line and tools (Retell lists it at $0.38)
Google Gemini 3.8 Live$0.005 audio in, $0.018 outModel with speech in and outAdd tools and phone line
Twilio US line$0.0085 inbound, $0.014 outboundLine only; number $1.15 a monthAdd where a platform excludes it

Vapi charges $0.05 and passes the rest through at cost: its calculator prices Deepgram transcription at $0.0095–$0.0099, an OpenAI model at $0.0077–$0.0452, an ElevenLabs voice at $0.0146–$0.0238 and a Twilio line at $0.008, so $0.09–$0.14 in all. Bland bills the phone line separately, so Twilio's $0.0085 inbound rate takes $0.14 to about $0.15. With a $0.04 model (GPT 5 or GPT 5.1 at Retell's standard rate), $0.055 for voice infrastructure and a $0.015 voice, a minute on your own line comes to $0.11; Retell's own line adds $0.015. ElevenLabs does not say whether its $0.08 includes the model or line, and Synthflow publishes no per-minute rate; it sells annual enterprise agreements.

OpenAI's realtime models bill tokens. OpenAI's cost guide counts a minute of caller audio as 600 tokens and a minute of agent audio as 1,200. At OpenAI's list prices for gpt-realtime-2.1, a call where each side talks half the time costs about $0.05 a minute in new audio. But "the entire conversation is sent to the model for each Response," so later turns cost more; the part of the history held in cache is billed at $0.40 per million tokens. The mini model costs about a third as much.

Check concurrency too: Retell includes 20 simultaneous calls, then $8 a month per extra line; Vapi's $29 Core plan includes 10; Bland's Start plan allows 10 at once and 100 calls a day. Fifteen callers at 9 a.m. on a Monday will not all get through a 10-line plan.

Usage is only the running cost; for the build and upkeep, see what an AI agent costs to build and run.

Is an AI answering service cheaper than a human receptionist?

Per minute, far cheaper. Per outcome, it depends on the call.

What a minute of phone answering costs: AI, a live answering service and an in-house receptionist. Pricing pages and BLS data, checked October 26, 2026.

OptionPublished figurePer minute
AI voice agentRetell's list range$0.07–$0.31 per talk minute
Ruby answering service, 50 minutes$250 a month$5.00 per talk minute
Ruby answering service, 500 minutes$1,725 a month$3.45 per talk minute
US receptionist, 2025 median$18.27 an hourAbout $0.30 per paid minute, busy or idle

Take a clinic with 500 minutes of calls a month, about 170 three-minute calls. AI usage at $0.09–$0.15 a minute comes to $45–$75. Ruby, with people answering 24/7 including holidays, charges $1,725. One receptionist at the BLS median of $38,010 a year costs about $3,170 a month before taxes and benefits, for 40 hours a week; the healthcare median is $19.00 an hour. Staffing all 168 hours of a week takes 4.2 full-time schedules.

The comparison flatters AI. A receptionist also greets patients and handles paperwork, and a person brings judgment to odd calls. The fair target is narrower: calls that today hit voicemail, a busy line or an after-hours service.

Should you use a voice AI platform or build your own?

Start on a platform. Retell, Vapi and Bland run the phone connection, turn-taking, interruptions and logs; you write the instructions and pick the voice and tools. A test line is configuration, not a build.

Build on realtime APIs when you need a model or data location the platform lacks, the agent sits inside a larger product, or volume makes the fee matter: at 100,000 minutes a month, Vapi's $0.05 fee alone is $5,000. Then you own the phone connection, turn detection, retries and monitoring. We compared the same trade-off for text agents in n8n, Make and Zapier vs a custom AI agent, and the questions in how to choose an AI automation agency apply to voice too.

Yes. Answering calls carries little legal risk; outbound calls and recordings need care.

Outbound AI calls need consent

The FCC's February 2024 ruling confirmed that AI-generated voices are "artificial" under the TCPA. The statute bars calls to mobile phones and home lines that use "an artificial or prerecorded voice" without "prior express consent," apart from emergencies and calls the FCC exempts. Telemarketing with such a voice needs prior express written consent. The message must name your business at the start, give a phone number and offer an automated opt-out by voice or keypress, and sales calls to homes are allowed only from 8 a.m. to 9 p.m. local time (47 CFR § 64.1200). People can sue for $500 per violation or their actual loss, whichever is greater, and a court can triple it for willful or knowing violations.

What changed in 2025 and 2026

  • AI disclosure: proposed. The FCC's 2024 proposal would require disclosing AI-generated calls at the start. As of October 26, 2026, we found no final rule. Disclose anyway.
  • Revoking consent: from November 25, 2026. Under FCC rules taking effect on that announced date, callers may designate how people revoke consent, such as a voice or keypress opt-out during the call, and must disclose it (Brownstein). The 10-business-day deadline to honor an opt-out stays, for now.
  • California. AB 2905, signed in September 2024 and in force since January 1, 2025, requires automated dialing calls to say when a message uses a voice "generated or significantly altered using artificial intelligence."
  • Utah. Since May 2025, businesses must disclose generative AI on a "clear and unambiguous request," and licensed professionals must disclose it up front in "high-risk" interactions, which collect sensitive data such as health or financial information and give personalized advice.

Our rule for every market: the agent says it is an AI in its first sentence.

Recordings and health data

Most states need one party's consent to record a call, but 13 require everyone's, including California, Florida, Illinois, Pennsylvania and Washington, per a 2024 survey by Kilpatrick Townsend. Across state lines, "the strictest standard generally applies." So announce recording and transcription in the greeting.

Clinics also need a signed business associate agreement (BAA): HHS treats a cloud vendor that stores or processes patient data as a business associate. Vapi sells a HIPAA add-on at $2,000 a month; Retell calls itself "HIPAA-ready," and its pricing table lists a HIPAA BAA on both pay-as-you-go and Enterprise, with a custom BAA on Enterprise only.

What do UAE rules say about AI phone calls?

Automated calls are allowed with conditions, and enforcement is active.

Cabinet resolutions in force since mid-August 2024 set the rules for marketing calls (Khaleej Times): 9 a.m. to 6 p.m. only; no second call after a refusal; retry an unanswered call at most once a day and twice a week; ask whether the person wants to continue before the pitch; use local numbers registered under your commercial license; skip numbers on the Do Not Call Registry (DNCR); announce recording at the start. They cover "all licensed companies, including those in free zones," and phone marketing needs prior approval (HiDubai).

Company fines reach AED 150,000 (Khaleej Times). The National reported in August 2026 more than AED 19 million in fines and 9,433 numbers disconnected since August 2024. Automated systems may be used only with consent asked at the start of the call, and personal numbers are banned.

Firms are not penalized for calls "initiated at the consumer's request," per Khaleej Times, so inbound calls and requested callbacks are the safer start. Mind the numbers too: under Twilio's terms for UAE numbers, a UAE geographic number may "only be used to receive inbound voice calls," and outbound use is "strictly prohibited," including over SIP.

Recordings and transcripts are personal data: the PDPL, the federal data law, lists voice among personal identifiers. Our WhatsApp AI agent guide covers its consent and transfer rules. As of June 2026, the law's implementing regulations "have still not been issued," and a new Federal Authority for Artificial Intelligence and Data absorbs the Emirates Data Office (Morgan Lewis). DIFC and ADGM companies follow their own data laws.

Clinics face one more rule. Federal Law No. 2 of 2019 bars moving health data about medical procedures in the UAE abroad, including storing or processing it there, unless a ministerial exception applies (CMS). Sending patient calls to a model hosted abroad needs legal review first.

Will an AI voice agent understand Arabic, Hindi, Urdu and Russian callers?

Often, but test first. Deepgram's Nova-3 lists 17 Arabic locale codes, including Emirati (ar-AE), Saudi (ar-SA) and Egyptian (ar-EG). It handles Hindi and Russian in a multilingual mode for callers who switch languages; Urdu works only on its own. Its conversational model, Flux, supports neither Arabic nor Urdu as of October 26, 2026. ElevenLabs says its platform supports more than 90 languages.

Supported does not mean accurate on your calls. Before launch, run 50 to 100 recorded calls per language, with consent, through the speech model and check names, numbers, dates and addresses by hand: one wrong digit breaks a booking. Have a native speaker judge the voice.

When should you not use an AI voice agent?

  • Emergencies. Clinics and home-repair lines need a fast route to a person.
  • Complaints. Hand upset callers to a person at once.
  • High-stakes negotiation, such as a property deal.
  • Card numbers and IDs. Text a payment link instead, and turn on PII removal (Retell charges $0.01 a minute).
  • Low volume. A few after-hours calls a week may need only voicemail and a quick callback.
  • Cold outbound calls in the UAE. A permit, registered numbers and consent make it a job for counsel.

A lesson from our text agent carries over. For a client under NDA that runs hundreds of Instagram accounts, the agent reads comments and DMs and hands anything sensitive to a person; if a run breaks or a DM needs judgment, a person gets a Telegram alert with the step, the account and the reason. A voice agent needs the same path to a human, not a caller stuck in a loop.

Instagram AI AgentClient under NDA

Case study · Under NDAHundreds of Instagram accounts, one AI agent

What should you check before launching an AI voice agent?

  • Calls to automate. A month of call logs: volume by hour, top reasons, missed calls. Start with one or two reasons.
  • Numbers. US: a local or toll-free number ($1.15 or $2.15 a month on Twilio) or your carrier over SIP. UAE: an inbound number; for outbound, numbers registered to your license.
  • Greeting. AI disclosure, your business name, a recording notice and how to reach a person.
  • Consent. For outbound, when and how each person agreed; opt-outs honored within 10 business days.
  • Handoff. Warm transfer with a summary, callback times after hours, alerts on failure.
  • Actions. Calendar, CRM and texts, each with limits.
  • Languages and speed. Recorded test calls per language, on mobile networks and speakerphones.
  • Data. Storage location, retention and access for recordings and transcripts; a BAA for US patient data; UAE health data kept in the country; vendor terms on training.
  • Capacity and budget. Lines for your busiest hour; a per-minute budget with line, models and add-ons.
  • Review. A weekly sample of calls, with resolution, transfers and complaints tracked.

If you want a partner to scope it, see our AI automation services, our page for UAE companies, or book a call.

Questions we get asked

Can an AI answer my business phone calls?

Yes, for calls with a clear purpose. An AI voice agent answers questions from your information, books or moves appointments, takes messages and transfers the caller to a person, around the clock and on several lines at once. Inbound calls are the safer start: in a 2024 proposal, the FCC noted that TCPA requirements do not extend to technologies used to answer inbound calls. Have the agent say it is an AI and that the call is recorded.

How much does an AI receptionist cost per minute?

On pricing pages checked October 26, 2026, Retell charges $0.07 to $0.31 a minute, depending on the model and voice. Vapi's $0.05 platform fee plus speech-to-text, model, voice and phone line comes to about $0.09 to $0.14, and Bland's Start plan costs $0.14 plus the line. Budget also for extra concurrent lines, a phone number from $1.15 a month in the US, and the cost of building and maintaining the integrations.

Is it legal to call customers with an AI voice in the US?

Only with consent. The FCC's February 2024 ruling made AI-generated voices artificial under the TCPA, so calls to mobile and home numbers need prior express consent, and telemarketing calls need written consent. The message must identify the business, give a phone number and offer an automated opt-out, and sales calls to homes are limited to 8 a.m. to 9 p.m. local time. People can sue for $500 per violation, and courts can triple it for willful violations.

Can a UAE business use an AI voice agent?

Yes, with conditions. Answering inbound calls and returning requested callbacks is the safer start, because firms are not penalized for calls made at the consumer's request. Marketing calls need prior approval, local numbers registered to the company's license, calling hours of 9 a.m. to 6 p.m., no numbers on the Do Not Call Registry and, for automated systems, consent asked at the start of the call. Recordings and transcripts are personal data under the PDPL.

Does an AI voice agent have to say it is an AI?

It depends on the market. As of October 26, 2026, the FCC's 2024 proposal to require AI disclosure at the start of calls was not final. California's AB 2905 requires automated calls to disclose an AI-generated voice, and Utah requires disclosure when a consumer clearly asks and, for licensed professionals, up front in high-risk interactions. UAE automated marketing calls must ask for consent at the start. Our rule: the agent says it is an AI in its first sentence, in every market.

Will an AI voice agent understand Arabic callers?

Often, but test it on your own calls. Deepgram's Nova-3 lists 17 Arabic locale codes, including Emirati, Saudi and Egyptian Arabic, while its Flux conversational model supports neither Arabic nor Urdu as of October 26, 2026. Nova-3 handles Hindi and Russian in a multilingual mode for callers who switch languages. Before launch, run 50 to 100 consented, recorded calls per language through the model and check names, numbers, dates and addresses by hand.

Sources

  1. Pricing · Retell AI, 2026
  2. Pricing · Vapi, 2026
  3. Pricing · Bland AI, 2026
  4. API Pricing · ElevenLabs, 2026
  5. Pricing · Synthflow, 2026
  6. API Pricing · OpenAI, 2026
  7. gpt-live-1 model · OpenAI, 2026
  8. Cost optimization · OpenAI, 2026
  9. Gemini Developer API pricing · Google, 2026
  10. Voice pricing: United States · Twilio, 2026
  11. Service country-specific terms: UAE phone numbers · Twilio, 2026
  12. Occupational Outlook Handbook: Receptionists · U.S. Bureau of Labor Statistics, 2026
  13. Pricing · Ruby, 2026
  14. Universals and cultural variation in turn-taking in conversation · PNAS (Stivers et al.), 2009
  15. Models & Languages Overview · Deepgram, 2026
  16. Declaratory Ruling on AI-generated voices in robocalls (FCC 24-17) · Federal Communications Commission, 2024
  17. Implications of Artificial Intelligence Technologies on Protecting Consumers From Unwanted Robocalls and Robotexts (proposed rule) · Federal Register, 2024
  18. 47 CFR § 64.1200, Delivery restrictions · Legal Information Institute, Cornell Law School, 2026
  19. 47 U.S. Code § 227, Restrictions on use of telephone equipment · Legal Information Institute, Cornell Law School
  20. New TCPA Revocation of Consent Rules Take Effect on November 25 · Brownstein Hyatt Farber Schreck, 2026
  21. AB 2905: Telecommunications: automatic dialing-announcing devices: artificial voices · CalMatters Digital Democracy, 2024
  22. Utah scales back reach of generative AI consumer protection law · Davis Polk, 2025
  23. Wiretap laws in the United States · Kilpatrick Townsend, 2024
  24. Guidance on HIPAA & Cloud Computing · U.S. Department of Health and Human Services
  25. New telemarketing rules in UAE: Timings, fines, exemptions explained · Khaleej Times, 2024
  26. Ministry of Economy, TDRA announce new resolutions on controls, procedures regulating telemarketing · HiDubai Focus, 2024
  27. Dh19 million in fines issued and 9,400 numbers disconnected for telemarketing violations · The National, 2026
  28. Federal Decree by Law No. (45) of 2021 Concerning the Protection of Personal Data · UAE Legislation, 2021
  29. UAE establishes Federal Authority for Artificial Intelligence and Data · Morgan Lewis, 2026
  30. UAE Health Data: new data transfer permissions · CMS, 2021

Share this article

About the author

Artem Zelenskii

Head of AI

Leads AI and ML: 5+ years in production. Apple Developer Academy graduate.

Meet the leadership

Keep reading

All posts

AI Automation

n8n, Make or Zapier vs a Custom AI Agent: When to Switch

Each platform bills a different unit, so plan prices don't compare. We count one workflow three ways, quote n8n's license and list the six signals that an agent has outgrown no-code.

Artem Zelenskii

· 14 min read

Let's talk

Got a project? Let’s scope it in 30 minutes.

Tell us what you're building. You'll get a straight answer on timeline, team and approach on the first call.

Reply the same dayTalk to a partnerNo sales pitch

Tell us what you're building.

A partner reads every request and replies the same working day, on the channel you pick.

Reply on
Reply on
What do you need? · optional

A few sentences: what it is, who it's for, the deadline.

We use these details only to reply. Privacy policy