Voice AI Agents for Business: Sales, Support and Customer Operations

Voice AI agents are AI systems that hold real phone conversations: they answer and place calls, understand natural speech, work out what the caller needs, act in your systems — qualifying, booking, updating records — and hand over to a person when they should. In September 2026, platforms charge roughly 5 to 31 US cents a minute, before phone-line costs.

This guide explains how voice AI agents work, what they can do in sales, support and operations, what they cost, what the evidence says about results and which rules apply to AI calls in the UAE, the US and the EU. It ends with a deployment plan and the metrics to watch.

Key takeaways

• Voice agents are now a product category. OpenAI's Realtime API went generally available with phone calling in August 2025; Google, Salesforce and Microsoft have since shipped voice agents of their own.
• There are two designs. Speech-to-speech models sound most natural; chained pipelines let you check the text before the agent speaks — OpenAI recommends them where statements must be policy-checked.
• The per-minute price isn't the total. Model, platform, LLM and telephony are billed separately on most stacks, and calls to UAE mobiles can cost more than the AI itself.
• The evidence says hybrid, not replacement. Only 20% of service leaders in a Gartner survey had cut staff because of AI, and Gartner expects half of those who did to rehire by 2027.
• Outbound calling is the regulated part. UAE rules allow automated calls but set hours, numbers, consent and recording requirements, with fines up to AED 150,000.

What is a voice AI agent?

A voice AI agent is an AI system that holds a spoken conversation and takes actions to complete a task. It listens, understands intent in natural language, decides the next step, uses tools — a calendar, a CRM, an order system — and speaks the answer, all in real time. When the call needs judgement, empathy or authority it doesn't have, it hands over to a person with a summary.

Definition

A voice AI agent is a conversational AI that answers or places calls, understands free-form speech, reasons about what to do and acts in business systems within limits you set. It is the voice form of an AI agent: perception, reasoning and action, with a person accountable for the outcome.

It helps to separate three generations of phone automation, because vendors use "AI" for all of them:

Generation	How it works	Caller experience	Where it breaks
Phone menu (IVR)	Press 1, press 2; fixed decision tree	Slow, frustrating, predictable	Anything outside the menu
Voicebot	Recognises set phrases and intents; scripted answers	Better, but brittle	Unexpected wording, follow-up questions
Voice AI agent	Understands free speech, reasons, uses tools, acts	Close to a competent human on routine calls	Edge cases, dialects, tasks without clear rules

How do voice AI agents work?

A voice AI agent works in five steps on every turn of a call: it hears the caller, works out what they mean, decides what to do, takes the action in your systems and speaks the result — and hands over when a rule says a person should take it. Good agents do all five in well under a second, so the conversation doesn't stall.

One turn of a voice agent call

1. 01Hear
   • Speech recognition
   • Barge-in detection
   • Dialect

   Transcribes or listens to the caller directly

2. 02Understand
   • Intent
   • Details
   • Sentiment

   Works out what the caller actually wants

3. 03Decide
   • Policies
   • Knowledge base
   • Next step

   Chooses an answer or an action within limits

4. 04Act
   • CRM
   • Calendar
   • Order system

   Books, updates or looks up — with logging

5. 05Speak or hand over
   • Voice
   • Summary
   • Transfer

   Answers, or passes the call to a person with context

Under the hood there are two main designs. OpenAI's documentation describes both — plus a hybrid where a live voice model hands reasoning and tool use to a separate backend model:

Design	How it works	Strength	Weakness
Speech-to-speech	One model hears audio and speaks audio directly	Lowest latency; natural tone and interruptions	Harder to inspect what the agent will say before it says it
Chained pipeline	Speech-to-text, then a language model, then text-to-speech	Every answer exists as text first, so it can be checked, logged and filtered	More moving parts; slightly slower
Live voice plus backend	A voice model manages the conversation and delegates reasoning and tools	Natural conversation with stronger reasoning	More complex to build and monitor

The choice depends on what the agent is allowed to say. For a restaurant booking line, speech-to-speech is fine. For an agent that quotes prices, eligibility or contract terms, a chained design lets you check every sentence against policy first — which is exactly the case OpenAI recommends it for.

Which platforms offer voice agents in 2026?

OpenAI's Realtime API became generally available on 28 August 2025 with SIP phone calling, and OpenAI has released newer realtime and live voice models since. Google released its Gemini 3.8 Live models in September 2026, supporting more than 97 languages. Salesforce announced Agentforce Voice in October 2025, and Microsoft made real-time voice agents generally available in Dynamics 365 Contact Center in April 2026, starting in North America. Around them sits a layer of voice-agent platforms — ElevenLabs, Retell, Vapi, Bland, Deepgram, PolyAI and Sierra — that bundle telephony, tools and analytics.

Option	Examples	Best for	Trade-off
Speech model APIs	OpenAI Realtime and GPT-Live, Google Gemini Live	Teams with developers who want full control	You build telephony, tools, monitoring and guardrails
Voice agent platforms	ElevenAgents, Retell, Vapi, Bland, Deepgram Voice Agent	Launching quickly with numbers, tools and call analytics	Fees stack up per minute; check data handling and residency
Contact-centre suites	Salesforce Agentforce Voice, Microsoft Dynamics 365 Contact Center	Companies already running service on those platforms	Licensing; regional availability varies
Managed specialists	PolyAI, Sierra	Large service operations wanting a managed build	Custom pricing and longer projects

What can voice AI agents do for sales, support and operations?

Voice AI agents are best at high-volume, repeatable calls with clear rules: qualifying inbound enquiries, booking and rescheduling appointments, answering support questions from a knowledge base, giving order or case status, covering after-hours calls, and making consented follow-up calls. They are weakest where a call needs negotiation, authority or genuine empathy.

Use case	What the agent does	Autonomy	Watch out for
Inbound qualification	Answers enquiry calls, asks qualifying questions, scores and routes	High	Promises the business can't keep
Appointment booking	Checks calendars, books, confirms, reschedules and sends reminders	High	Double bookings and time zones
Support triage and resolution	Identifies the issue, answers from the knowledge base, resolves simple requests	Medium to high	Handing over complex or emotional cases quickly
Order and case status	Looks up records and reports progress	High	Verifying identity before sharing personal data
After-hours cover	Answers common questions, takes messages, books callbacks	High	Telling callers it's an AI and what happens next
Outbound follow-up	Calls people who asked to be contacted	Medium	Consent, calling hours, registries and recording
Agent assist and quality review	Listens to human calls, suggests answers, scores quality	Assistive	Telling staff and customers; data protection

In sales, the value is speed and consistency: every enquiry gets answered in seconds, asked the same qualifying questions and routed with a summary. In support, it is capacity: routine calls stop queuing behind complex ones. In operations, it is coverage: confirmations, reminders and status calls happen without anyone dialling. The executive guide to AI agents covers how to choose which workflows to automate first.

What does a voice AI agent cost?

A voice AI agent costs the sum of four things: the speech and language models, the agent platform, the phone line and the work to build and run it. Published platform prices ranged from about 5 to 31 US cents a minute in September 2026, but most exclude something — usually the language model or telephony — so compare like with like.

Platform or model	Published price (checked September 2026)	What's included
OpenAI GPT-Live 1	$0.05 a minute, billed per second	The live voice model; backend model and tool use billed separately
OpenAI gpt-realtime-2.1	$32 per million audio input tokens, $64 per million output	Model only; listening costs about 2 US cents a minute and speaking about 8, before the conversation history that is re-read each turn
Google Gemini 3.8 Live	$0.005 a minute of audio in, $0.018 out	Model only
ElevenLabs ElevenAgents	$0.08 a minute	Voice and platform; language model at cost; your own telephony
Deepgram Voice Agent API	$0.075 a minute pay-as-you-go	Speech-to-text, model orchestration and text-to-speech
Retell AI	$0.07 to $0.31 a minute	Voice infrastructure plus chosen voice, model and telephony
Vapi	$0.05 a minute hosting	Speech, model, voice and telephony billed at provider cost
Bland	$0.12 to $0.14 a minute	Model, speech recognition and voice included; telephony extra

The OpenAI per-minute figures are my arithmetic from its token rates (about one token per 100 milliseconds of caller audio and one per 50 milliseconds of agent audio), not published prices. Every price here changes often; check before you budget.

Telephony is the line item people forget. Twilio lists outbound calls to UAE mobiles at $0.2995 a minute and to landlines at $0.3635. A hypothetical outbound follow-up campaign of 1,000 calls averaging two minutes would cost about $160 on an $0.08-a-minute platform — and about $599 in call charges to mobiles, before language-model usage. For outbound work in the UAE, the phone line can cost more than the AI.

What the data shows

Gartner predicts that by 2030 the cost per resolution of generative AI in customer service will exceed $3 — more than many offshore human agents cost for business-to-consumer work. It's a forecast, not a measurement, but it's a useful warning: AI voice is cheap per minute, not necessarily cheap per resolved problem once models, integration and oversight are counted.

What does the evidence say about results?

The evidence says voice and chat AI can take a large share of routine contacts, but that companies keep people for the rest — and some that cut too far are reversing it. The strongest numbers come from chat and agent-assist deployments; independent data on fully autonomous voice agents is still thin.

• Prediction: Gartner expects agentic AI to resolve 80% of common customer service issues without human help by 2029, cutting operating costs by 30%.
• Survey: in an October 2025 Gartner survey of 321 customer service leaders, only 20% had reduced headcount because of AI; 55% kept staffing flat while handling more volume.
• Prediction: Gartner expects half of the companies that cut service staff because of AI to rehire for similar work, under new titles, by 2027.
• Company report: Klarna's AI assistant handled 80% of its customer chats in 2025 with no drop in satisfaction, according to its 2025 annual report. That is chat, not voice.
• Study: a study of 5,179 support agents by Brynjolfsson, Li and Raymond found AI assistance raised issues resolved per hour by 14% on average and 34% for novice agents. The AI helped human agents rather than replacing them.

What this means

Read together, the numbers suggest designing for a hybrid team from day one: the agent handles the predictable calls end to end, people take the rest with a summary in hand, and the system gets measured on resolved problems — not on how many calls it kept away from humans.

What rules apply to voice AI agents?

The rules that matter most are about outbound calling, consent, recording and disclosure. In the UAE, automated calls are allowed under the telemarketing rules but only within strict limits. In the US, AI-generated voices count as artificial voices that need consent. In the EU, people must be told they are talking to an AI from 2 August 2026.

UAE telemarketing rules

Cabinet Resolution No. 56 of 2024, in force since August 2024, covers marketing calls, marketing SMS and marketing messages on social-media apps by every UAE-licensed company, including free-zone companies. It expressly allows automated calling, provided the calls follow the same rules as people do:

Requirement	What it means for a voice agent
Prior approval and registered numbers	Call only from UAE numbers registered to the company's trade licence, with approval from the competent authority; individuals may not market from personal numbers
Do Not Call Registry	Screen every number against the national registry before calling
Calling hours	Marketing calls only between 9am and 6pm
Refusals and retries	No further calls after a refusal; after no answer, at most once a day and twice a week
Recording and identification	Record calls, say so at the start, identify the company and the purpose, and ask whether the person wants to continue
Data	No sharing or trading of personal data without consent; keep a log of marketing calls

Cabinet Resolution No. 57 of 2024 lists 18 violations, with fines for companies from AED 10,000 to AED 150,000 that rise for repeat offences. Individuals who market from personal numbers face AED 5,000 and suspended lines for a first offence, rising to AED 50,000 and a year without telecom services. Consumers can register or report numbers by texting DNCR to 2211 or 1012.

Two more UAE laws apply. The federal data protection law (PDPL) makes consent the default basis for processing personal data and lets people object to direct marketing and to decisions made purely by automated processing that affect them. And the UAE's cybercrime law makes recording conversations without consent a crime, which is another reason to announce recording at the start of every call.

US and EU rules

Jurisdiction	Rule	What it requires
United States	FCC Declaratory Ruling 24-17 (February 2024)	AI-generated voices count as "artificial" under the TCPA: prior express consent for calls, written consent for telemarketing, plus identification and opt-out
European Union	AI Act, Article 50 (applies from 2 August 2026)	People must be told they are interacting with an AI, at the latest at the first interaction, unless it's obvious

The EU's 2026 Digital Omnibus delayed several AI Act obligations but left the Article 50 disclosure duty on its original timetable. It applies to companies outside the EU whose AI talks to people in it — relevant for UAE businesses calling European customers. This section summarises the rules; it isn't legal advice.

Common misconception

"AI cold calling is illegal in the UAE." It isn't: Resolution No. 56 of 2024 expressly allows automated calling that follows the rules. What's illegal is calling outside 9am to 6pm, calling registry numbers, calling from unregistered numbers, ignoring a refusal or recording without notice. The compliant way to be fast is to answer inbound calls instantly and call back only people who asked.

How do you deploy a voice AI agent?

Deploy a voice AI agent one call type at a time: start with inbound calls that follow clear rules, write the escalation paths before the happy path, connect only the systems the agent needs, test with real callers' speech, and widen its scope only when the numbers justify it.

1. Pick one call type. Choose a high-volume, rule-based call — booking, status, first-line qualification — with a clear success measure.
2. Write the escalations first. List what the agent must hand over: complaints, pricing exceptions, legal questions, upset callers, anything it can't verify. Every call needs a route to a person.
3. Connect the minimum. Give the agent read access where it only needs to look things up, and write access only to the records it must change. Log every action it takes — the controls in the AI governance guide apply.
4. Choose the architecture. Speech-to-speech for natural conversation; a chained pipeline where statements must be checked; a platform if you need to launch in weeks.
5. Test the language. "Supports Arabic" often means Modern Standard or Saudi Arabic. Some platforms list only Saudi Arabic; Deepgram's Nova-3 Arabic recognises 17 variants including UAE Arabic. Test recognition and voice separately, with real callers who switch between Arabic and English mid-sentence.
6. Build in compliance. Disclosure at the start, recording notice, consent records, registry screening and calling-hour limits for anything outbound.
7. Launch inbound, measure, then expand. Run a pilot on a share of calls, review transcripts weekly, and add outbound follow-up only for people who asked to be contacted.

Metric	What it tells you
Resolution rate	Share of calls the agent completes correctly without a person
Transfer rate and reasons	Where the agent's limits are, and which to fix next
Booking or qualification rate	Whether sales calls produce the next step
Caller satisfaction	Whether callers accept the agent
Average handling time	Whether calls are efficient or looping
Cost per resolved call	The real economics, including models, telephony and oversight
Compliance incidents	Out-of-hours calls, missed disclosures, registry hits — the target is zero

For property businesses, voice AI for real estate applies the same pattern to enquiry lines and viewing bookings, feeding the AI real estate CRM that runs the rest of the pipeline.