In short: Since January 2025, four things have changed fundamentally for AI voice agents: the technology, the prices, the money in the market and the rules. OpenAI, Google, Amazon, Microsoft and xAI have made realtime voice models generally available that cope better with interruptions and pauses. Since August 2025, OpenAI connects its model directly to the phone network over SIP. At OpenAI and xAI, the voice layer launched at $0.05 per minute. The 21 funding rounds we could verify with a source add up to roughly $3.1 billion for voice and customer service AI. Since 2 August 2026, an AI agent in the EU must identify itself as AI at the start, unless that is obvious. And analysts have become more cautious: Gartner expects that by 2028 no Fortune 500 company will have fully eliminated human customer service.
The first version of this article from January 2026 promised up to 60 percent lower support costs and the recognition of emotional nuances, without evidence. In October 2026 we replaced it with a dated timeline. Every statement comes from the source itself: vendor changelogs, press releases, EUR-Lex and FCC documents. We checked Gartner press releases through verbatim reprints, because gartner.com blocks automated requests. All sources are listed at the end, retrieved on 6 October 2026.
The timeline at a glance
| Date | What happened | Why it matters |
|---|---|---|
| 5 March 2025 | Gartner predicts agentic AI will autonomously resolve 80 percent of common service issues by 2029 [33] | the starting point of many expectations |
| 8 April 2025 | Amazon introduces Nova Sonic, one model for speech understanding and speech generation [4] | speech in, speech out, without a detour through text |
| 14 May 2025 | Twilio makes ConversationRelay generally available [6] | telephony plus your own language model from one vendor |
| 28 August 2025 | OpenAI makes the Realtime API generally available, including SIP [1] [2] | a major voice model takes phone calls without a middle layer |
| 8 September 2025 | NiCE closes its acquisition of Cognigy, valued at about $955 million [24] | the big contact centre vendors buy AI |
| 2 October 2025 | Deepgram releases Flux, speech recognition that detects the end of a speaker's turn [5] | fewer cases where the agent cuts the caller off |
| 17 December 2025 | xAI launches the Grok Voice Agent API at $0.05 per minute [7] | a per-minute price instead of token billing |
| 15 January 2026 | Parloa from Berlin raises $350 million at a $3 billion valuation [12] | a German voice start-up becomes a billion-dollar company |
| 7 May 2026 | OpenAI introduces GPT-Realtime-2 with configurable reasoning [1] | thinking during a conversation becomes adjustable |
| 24 July 2026 | Regulation (EU) 2026/1744 is published, the disclosure duty stays unchanged [28] | no delay for the sentence "you are speaking with an AI" |
| 2 August 2026 | Article 50 of the EU AI Act applies [27] | AI agents in the EU must identify themselves |
| 4 August 2026 | Gartner: 87 percent of customers consider access to a human essential [33] | no trust without a handover |
| 10 September 2026 | OpenAI makes GPT-Live 1 generally available, a model that listens and speaks at the same time [1] [2] | conversations without rigid turn-taking |
| 15 September 2026 | Google makes Gemini 3.8 Live generally available [3] | built, according to Google, for voice agents without reasoning delays |
Technology: what got better on the phone
An AI voice agent does three jobs: listen, understand and answer. For a long time this ran as a chain. Speech recognition transcribes, a language model writes the answer as text, and speech synthesis reads it out. Every step adds time. Since 2025 there are two more designs. Speech-to-speech models listen and speak directly, without text in between. Full-duplex models go further and keep listening while they speak, like a person on the phone. All three designs are still being developed. In September 2026, Google released new speech synthesis models explicitly for the chain design. [3]
For callers, the designs matter less than four concrete problems voice agents used to have. All four have moved.
Interruptions and pauses. Callers who paused to think used to get cut off. In September 2025, Google released a model with better handling of interruptions. Deepgram calls Flux the first speech recognition model that knows when someone is actually done talking. In July 2026, OpenAI explicitly improved silence, noise and interruption handling with GPT-Realtime-2.1. [1] [3] [5]
Numbers and letters. Customer numbers, licence plates and email addresses are a well-known source of errors on the phone. According to the changelog, GPT-Realtime-2.1 recognises letter and number sequences better. [1] Test it yourself anyway, with your real number formats.
Telephony. In May 2025, Twilio made ConversationRelay generally available, a building block that lets companies connect their own language model to the phone. [6] Since August 2025, OpenAI's realtime model takes calls directly over SIP. Keypad tones followed in November 2025, dedicated IP ranges for SIP in January 2026. [1] In December 2025, Amazon connected Nova 2 Sonic with Amazon Connect, Vonage, Twilio and others, and in July 2026 xAI released a no-code builder that includes phone numbers and SIP. [4] [7]
Multi-step tasks. Moving an appointment, checking an order, then answering a follow-up question: the model has to think without leaving a gap. In April 2026, xAI introduced a voice model for complex multi-step workflows, and in May 2026 OpenAI released GPT-Realtime-2 with configurable reasoning. Since September 2026, GPT-Live 1 hands reasoning and tools to a second model in the background. [1] [7]
Open models are part of the picture too, two of them from Europe. The Paris lab Kyutai open-sourced its speech recognition, its speech synthesis and the Unmute system in June and July 2025. In July 2025, Mistral AI released Voxtral, open speech understanding models under the Apache 2.0 licence. [10] In January 2026, NVIDIA introduced PersonaPlex, a full-duplex model with published weights. [9]
One point is often overlooked: not every model speaks your callers' language on day one. At launch, Amazon Nova Sonic only understood American and British English. Microsoft's speech-to-speech model became generally available in July 2026 with 11 languages, and the documentation for Cartesia's Sonic 3.6 speech synthesis lists 44. [4] [8] [11] Ask every vendor which model your callers will hear in German, French or Spanish and whether the same model also handles Turkish or Polish.
Prices: what the voice layer costs today
Model providers publish their prices. They apply to developers who build their own agents and cover only the voice layer.
| Model | List price on 6 October 2026 |
|---|---|
| OpenAI gpt-4o-realtime-preview (predecessor) | Audio $40 input and $80 output per 1M tokens [2] |
| OpenAI gpt-realtime to GPT-Realtime-2.1 | Audio $32 and $64 per 1M tokens, our own calculation: 20 percent cheaper [2] |
| OpenAI GPT-Realtime-2.1 mini | Audio $10 and $20 per 1M tokens [2] |
| OpenAI GPT-Live 1 | $0.05 per minute, billed per second, background model and tools extra [2] |
| xAI Grok Voice Agent API | $0.05 per minute at launch in December 2025, the model overview now lists $0.08 for speech-to-speech [7] |
| Google Gemini 3.8 Live | Audio input $0.005, audio output $0.018 per minute [3] |
| Deepgram Voice Agent API | $4.50 per hour, which is $0.075 per minute (our own calculation) [5] |
At OpenAI, the audio price is 20 percent below the predecessor model, and GPT-Live 1, xAI and Google offer per-minute prices instead of token billing. For a business that buys a ready-made phone agent, that is only part of the bill. On top come the language model for knowledge and tools, the phone number and telephony, storage and transcripts, calendar and CRM integrations, operations and support. Our AI phone agent guide breaks down how the price of a finished agent is built and what vendors charge. EchoCall costs €0.30 per minute in a package, excluding VAT. [35]
Money: who was funded and acquired in 2025 and 2026
We could verify these 21 funding rounds of voice and customer service AI companies with a source. The list is a selection, not a complete market overview.
| Date | Company | Round | Amount in USD | Valuation per source |
|---|---|---|---|---|
| Feb 2025 | Bland | Series B | $40M | not stated |
| Mar 2025 | Cartesia | Series A | $64M | not stated |
| May 2025 | Parloa, Berlin | Series C | $120M | $1B |
| Jun 2025 | Decagon | Series C | $131M | $1.5B |
| Jun 2025 | Synthflow AI, Berlin | Series A | $20M | not stated |
| Sep 2025 | HappyRobot | Series B | $44M | not stated |
| Sep 2025 | Sierra | Funding | $350M | $10B |
| Sep 2025 | Assort Health | Series B | $76M | not stated |
| Oct 2025 | Sesame | Series B | $250M | not stated |
| Nov 2025 | Giga | Series A | $61M | not stated |
| Nov 2025 | Wonderful | Series A | $100M | not stated |
| Dec 2025 | PolyAI | Series D | $86M | not stated |
| Jan 2026 | Deepgram | Series C | $130M | $1.3B |
| Jan 2026 | Parloa, Berlin | Series D | $350M | $3B |
| Jan 2026 | Decagon | Series D | $250M | $4.5B |
| Mar 2026 | Wonderful | Series B | $150M | not stated |
| May 2026 | Vapi | Series B | $50M | not stated |
| Jun 2026 | Bland | Series C | $50M | not stated |
| Jun 2026 | Assort Health | Series C | $120M | $1.2B |
| Aug 2026 | HappyRobot | Series C | $150M | $1.2B |
| Sep 2026 | Wonderful | Series C | $550M | $5B |
Sources: [12] to [23]. Together that is roughly $3.1 billion (our own calculation). Not included is the $950 million round Sierra announced in May 2026 at a valuation of over $15 billion. The source does not confirm that it closed. [14]
Acquisitions ran in parallel. NiCE bought Cognigy for about $955 million, Zendesk bought Local Measure and Forethought, Meta acquired PlayAI, and SoundHound AI bought Interactions for about $60 million in cash. Salesforce and ServiceNow committed $1.5 billion to Genesys, and Google licensed technology from Hume AI. [24] [25] [26]
Three things stand out. First, two of the funded companies are based in Berlin. Second, most of the money goes to platforms aimed at large enterprises. Third, established contact centre and service vendors such as NiCE and Zendesk are buying AI companies. Our comparisons with Parloa and Cognigy show how two of these platforms fit small and mid-sized businesses, and our Vapi alternative article covers developer platforms.
Law: what has applied since 2025
European Union
Article 50 of the EU AI Act has applied since 2 August 2026. [27] Paragraph 1 requires that people are informed when they interact with an AI system, unless this is obvious. The European Commission published guidelines on 20 July 2026. According to its questions and answers, the notice must come from the start of the first interaction, and the "obvious" exception must be interpreted restrictively. [29] Under Article 99(4), violations can cost up to €15 million or 3 percent of worldwide annual turnover. [27]
The so-called Digital Omnibus did not postpone any of this. Regulation (EU) 2026/1744 of 8 July 2026 replaces only paragraph 7 in Article 50. It gives systems that were on the market before 2 August 2026 until 2 December 2026 for the machine-readable marking under paragraph 2. The disclosure duty on the phone under paragraph 1 stays as it is. [28] On 31 July 2026, the Commission announced it would enforce the rules from 2 August. [29] Our EU AI Act article explains in detail what this means for chat and phone.
United States
The US has no federal duty to disclose an AI voice on the phone. In February 2024, the FCC ruled that AI-generated voices count as an artificial voice under the Telephone Consumer Protection Act. Outbound calls with an AI voice therefore need the consent the act requires. [30] In August 2024, the FCC proposed that callers disclose the AI voice at the beginning of each call. Inbound calls are explicitly not covered. We found no final rule as of October 2026. [30] In December 2025, the President signed an executive order to establish a national framework intended to preempt conflicting state laws. [31]
Individual states set their own rules. Since 2025, Maine has prohibited misleading consumers through text or voice into believing they are talking to a human without a clear notice. Since May 2025, Utah requires the notice only on a clear request or in high-risk interactions in regulated professions, but protects businesses that disclose from the start. Since January 2026, Texas requires the notice for government agencies and health care services, not for businesses in general. In September 2026, California signed a law that requires a disclosure for AI-generated performers in video and audio ads. Pure customer service bots are explicitly excluded from the companion chatbot rules in California and New York. [32]
Analysts: from "80 percent automated" to "humans stay"
Put the Gartner and Forrester press releases since March 2025 side by side and the change in tone is clear.
- March 2025: Gartner predicts that agentic AI will autonomously resolve 80 percent of common customer service issues by 2029 and cut operational costs by 30 percent. [33]
- June 2025: Gartner expects that by 2027 half of the organisations that planned to significantly reduce their service workforce because of AI will drop these plans. According to Reuters, Gartner also expects more than 40 percent of agentic AI projects to be cancelled by the end of 2027 because of rising costs and unclear business value. [33]
- September 2025: Gartner predicts that by 2028 no Fortune 500 company will have fully eliminated human customer service. [33]
- October 2025: Forrester expects a third of companies to harm customer experiences with frustrating AI self-service in 2026. [34]
- December 2025: Only 20 percent of surveyed service leaders have reduced staff because of AI. 55 percent keep staffing stable while handling more customer volume. [33]
- January 2026: Gartner expects the cost per resolution for generative AI to exceed $3 by 2030, more than many offshore human agents. By 2028, regulation will increase the volume handled by humans by 30 percent. [33]
- February 2026: By 2027, half of the companies that cut jobs because of AI will rehire staff for similar functions under different job titles, Gartner says. 91 percent of service leaders report pressure from executive leadership to implement AI. [33]
- April 2026: 54 percent of customers trust human agents more than AI for recommendations, 32 percent trust AI more. [33]
- May 2026: Forrester calculates that by 2030 AI will cause 49 percent of today's customer service jobs to disappear. The basis is a model with assumptions about staffing ratios and containment rates. [34]
- July 2026: According to Gartner, customers are about three times more likely to use third-party AI tools than company chatbots to resolve a service issue. Forrester reports that US customer service job postings are roughly 10 percent below pre-pandemic levels. [33] [34]
- August 2026: 87 percent of customers consider it essential to be able to reach a human when a company uses generative AI for service. AI spending in customer service rises by 38 percent, the overall budget by 2 percent. Surveyed service leaders name generative AI chatbots as the most valuable future channel, followed by live chat and generative AI voicebots. [33]
- September 2026: Only 27 percent of customers would try a chatbot again after a bad experience. [33]
The forecasts partly contradict each other. Forrester sees almost half of all service jobs disappearing, Gartner sees companies rehiring. Both agree on two points: AI takes over simple, recurring requests, and customers rarely forgive bad automation. More figures on market size, adoption and costs are in our AI voice agent statistics.
What this means for businesses in Germany and the EU
- The technology handles routine calls. A well-configured agent can take over appointments, opening hours, callback requests and status questions in 2026. Advice, complaints and special cases still belong with people.
- The AI notice is mandatory. One sentence in the greeting is enough. It has to come at the start, not only when someone asks.
- The way to a human is part of the product. 87 percent of customers expect it. An agent should be able to transfer live or promise a callback instead of keeping the caller in a loop.
- The first impression counts twice. If only about a quarter of customers give a chatbot a second chance after a bad experience, it pays to test with real calls before launch and to read the transcripts during the first weeks.
- The differences are no longer in the model. When the voice layer costs a few cents per minute, telephony, integrations, data protection, language quality and support decide which vendor you choose.
That is where EchoCall comes in. The agent speaks more than 30 languages on the phone and recognises the caller's language. It transfers live to your team and stores every conversation as a transcript in the EchoHub. Hosting is with IONOS in Frankfurt and Karlsruhe, with a data processing agreement under Art. 28 GDPR. Documents on the agreement and sub-processors are on our GDPR page. The trial runs for 14 days with 30 voice minutes and 30 chat conversations. You add a card for it, which is verified with €1. Plans and prices are on the pricing page, all features on the AI voice agent page.
Frequently asked questions
What is an AI voice agent?
An AI voice agent is software that holds spoken conversations, on the phone or in the browser. It understands full sentences, answers in natural language and can complete tasks, such as booking an appointment or taking a callback request. Unlike a phone menu, it does not need keypad input.
What has changed for AI voice agents in 2025 and 2026?
The major model providers have made realtime voice models generally available that handle interruptions, pauses and number sequences better. OpenAI has connected its model directly over SIP since August 2025. In 2026, models arrived that listen and speak at the same time, from September also from OpenAI. Funding rounds worth several billion dollars followed, and in the EU the disclosure duty under Article 50 of the AI Act.
How much does an AI voice model cost per minute?
The voice layer alone costs $0.05 per minute with OpenAI GPT-Live 1, and xAI's model overview lists $0.08. Google charges $0.005 per minute of audio input and $0.018 per minute of audio output for Gemini 3.8 Live. A ready-made phone agent costs more, because telephony, knowledge, integrations and operations come on top. EchoCall charges €0.30 per minute in a package, excluding VAT.
Which AI voice start-ups raised money in 2026?
Among voice and customer service AI companies: Parloa from Berlin with $350 million at a $3 billion valuation, Decagon with $250 million, Deepgram with $130 million, Wonderful with $150 million and $550 million, HappyRobot with $150 million, Assort Health with $120 million, and Vapi and Bland with $50 million each. Sierra announced a round of $950 million.
Does an AI voice agent have to say that it is an AI?
In the EU, yes. Since 2 August 2026, Article 50(1) of the AI Act requires that people are informed when they talk to an AI, according to the European Commission from the start of the first interaction. Only cases where this is obvious are exempt. The Digital Omnibus did not postpone this. The US has no federal duty, but individual states have rules.
Will AI voice agents replace call centre agents?
Partly. At the end of 2025, only 20 percent of service leaders had reduced staff because of AI, according to Gartner. Gartner expects half of the companies that cut jobs to rehire by 2027. Forrester, by contrast, expects 49 percent of today's customer service jobs to disappear by 2030.
What do Gartner and Forrester say about AI in customer service?
In March 2025, Gartner expected AI to resolve 80 percent of common issues on its own by 2029. Since then, Gartner has stressed the limits: 87 percent of customers want to be able to reach a human, and only 27 percent would give a chatbot a second chance after a bad experience. Forrester warns that in 2026 a third of companies will harm their customer experience with poor AI self-service.
Sources
All sources are in English unless marked otherwise.
- OpenAI, API changelog: developers.openai.com
- OpenAI, model pages and pricing: gpt-realtime, gpt-4o-realtime-preview, gpt-live-1, pricing
- Google, Gemini API changelog, pricing and Vertex AI release notes: ai.google.dev, ai.google.dev, cloud.google.com
- Amazon, Nova Sonic and Nova 2 Sonic: aws.amazon.com, aws.amazon.com
- Deepgram, Voice Agent API and Flux: deepgram.com, developers.deepgram.com
- Twilio, ConversationRelay: twilio.com
- xAI, Grok Voice Agent API, Grok Voice Think Fast, Voice Agent Builder and model overview: x.ai, x.ai, x.ai, docs.x.ai
- Microsoft, Azure Speech release notes: learn.microsoft.com
- NVIDIA, PersonaPlex: research.nvidia.com
- Kyutai, blog and Unmute; Mistral AI, Voxtral: kyutai.org, mistral.ai
- Cartesia, Sonic 3.6 and model overview: cartesia.ai, docs.cartesia.ai
- Parloa, Series C and Series D: parloa.com, parloa.com
- Decagon, Series C and Series D: decagon.ai, decagon.ai
- Sierra, 2025 and 2026: sierra.ai, sierra.ai
- Wonderful, Series A, B and C: techcrunch.com, prnewswire.com, prnewswire.com
- Deepgram, Series C: deepgram.com
- PolyAI, Series D: prnewswire.com
- Vapi, Series B: vapi.ai
- Bland, Series B and Series C: bland.ai, prnewswire.com
- Synthflow AI, Series A: synthflow.ai
- HappyRobot, Series B and Series C: happyrobot.ai, happyrobot.ai
- Assort Health, Series B and Series C: prnewswire.com, prnewswire.com
- Cartesia, Series A; Giga, Series A; Sesame, Series B: cartesia.ai, giga.ai, techcrunch.com
- NiCE and Cognigy, announcement and closing: nice.com, cognigy.com
- Zendesk, Local Measure and Forethought: prnewswire.com, prnewswire.com
- Meta and PlayAI; SoundHound AI and Interactions; Genesys; Hume AI: techcrunch.com, sec.gov, silicon.co.uk, prnewswire.com
- Regulation (EU) 2024/1689 on artificial intelligence, Articles 50, 99 and 113: eur-lex.europa.eu
- Regulation (EU) 2026/1744: eur-lex.europa.eu
- European Commission, Article 50 guidelines, questions and answers, start of enforcement: digital-strategy.ec.europa.eu, digital-strategy.ec.europa.eu, digital-strategy.ec.europa.eu
- FCC, Declaratory Ruling FCC 24-17 and proposal FCC 24-84: docs.fcc.gov, docs.fcc.gov
- The White House, Executive Order 14365: whitehouse.gov
- Maine, Utah, Texas, California: legislature.maine.gov, le.utah.gov, capitol.texas.gov, gov.ca.gov
- Gartner press releases, checked through verbatim reprints: cmswire.com, stocktitan.net, Reuters via Yahoo, it-online.co.za, mhlnews.com, it-online.co.za, it-online.co.za, stocktitan.net, it-online.co.za, it-online.co.za, it-online.co.za, it-online.co.za
- Forrester, 2026 predictions and blog posts: forrester.com, forrester.com, forrester.com
- EchoCall, pricing: echocall.de/en/pricing
All links were checked on 6 October 2026. Third-party prices may change. The provider's current price list applies. Analyst forecasts are not measurements. This article is not legal advice.
Read next
- AI Voice Agent Statistics 2026: Market Size, Adoption, Cost per Minute, Highest-Value Niches and the Law, Every Number Sourced
- AI Phone Agent 2026: What It Costs, How It Works and How to Roll It Out
- EU AI Act 2026: What the AI Disclosure Duty Means for Voice and Chat Systems
- Vapi Alternative 2026: EU Region, GDPR, Real Costs and When Vapi Is the Better Choice
- Retell AI Alternative 2026: Costs, GDPR Without EU Operations, White Label and When Retell Is the Better Choice
