AI voice agents for businesses in 2026: where they work and where they fail

This fortnight's launches show where voice AI is going

ElevenLabs launched Eleven Realtime on 20 October in two versions: v1 Mini for natural, low-cost conversation and v1 Max for customer support, lead qualification and multi-step tasks that need tool calls. Both support more than 90 languages, and the API is compatible with OpenAI's Realtime API, so developers can switch by changing an endpoint and a model name. Convoso, an outbound sales platform for contact centres, made its no-code Voice AI Agent generally available on 2 November. It qualifies leads, answers routine inbound questions, books appointments and routes callers to the right team, using the same campaign, reporting and compliance controls as the human agents.

Two launches focus on language. Vaani AI Research opened its self-serve developer platform on 29 October, letting teams build agents for phone calls, websites, apps and WhatsApp, with support for 22 Indian languages plus English. Its chief executive, Tushar Shinde, put the problem plainly: “Seventy-five percent of the world does not primarily speak English, and most voice AI still treats that as an edge case.” AethexAI released Kora 2.0 the same day, adding more than 25 languages and dialects such as Yoruba, Hausa, Swahili and Amharic, built to handle accents and code-switching, where a caller moves between languages mid-sentence. As iAfrica noted, the company has not published accuracy benchmarks or the full language list, which is a good reminder to test any agent in your own callers' languages.

Where voice agents work: the calls nobody answers

The strongest case for a voice agent is not replacing a receptionist. It is answering the calls that currently go to voicemail. Upfirst's State of Voice AI for Small Businesses 2026, published on 14 October, combined surveys of 270 US consumers and 262 business owners with 677,562 real calls. It found that 58% of businesses miss at least one potential customer's call in a typical week, rising to 91% for firms with 10 to 49 employees, and that 78% of consumers have bought from a competitor because a business did not answer or call back. Only 24% of callers left a voicemail the last time a local business did not pick up.

Callers also accept a narrow, sensible set of jobs. In the same survey, 70% would let an AI answer questions about hours, services and pricing, 62% would let it take a message, and 45% would let it book or change an appointment. Language is another clear win: among Upfirst customers with at least 50 calls, 90% took at least one call in Spanish. Convoso's early customers point the same way. Josh Zakarya, president of Ace Building Materials, said staff had been juggling in-store customers and ringing phones, and the agent now routes callers “to the right department.”

Where they fail: money, complaints and misunderstandings

The same research shows the limits. Only 16% of consumers in the Upfirst survey would let an AI take a payment and 15% would let it handle a complaint. An AnswerConnect study run with OnePoll in May 2026 across 6,000 adults found that 85% would rather speak to a real person when contacting a business, 82% have asked for a person instead of an AI, and 59% find AI agents frustrating on customer service calls, against 11% for human agents. Those figures come from a company that sells human answering, and Upfirst sells AI answering, so the honest reading is that callers accept AI for simple tasks and want a person close by for everything else.

The most useful failure story is McDonald's. Its first automated drive-thru ordering system ran at about a hundred US locations before it was shut down in 2024, and the company is now testing a new voice assistant. In a September analysis for CIO, Foundry analyst Alberto Bellé argued that the first system “didn't fail because the AI made too many mistakes. It failed because it didn't know it was making mistakes.” He cites the QSR Drive-Thru Report: AI-taken orders were 81% accurate, rising to 95% when the system got stuck and handed the order to an employee. Most errors came from customised or unusual requests. The lesson applies to any business: an agent that recognises uncertainty and transfers the call beats one that keeps going and gets it wrong.

Tell callers it is an AI, and keep a person within reach

Disclosure matters more than a human-sounding voice. In the Upfirst survey, 80% of consumers want an AI to say at the start of the call that it is an AI, and among sceptical callers being told it was an AI was the top trust factor while sounding like a natural human voice ranked last. Earlier Upfirst analysis of 450,702 calls found that receptionists that disclosed they were an AI saw roughly 20% fewer hang-ups, though the company notes that this shows an association rather than a cause.

The handoff is the other half. In Upfirst's call data, 16.4% of calls included an attempt to hand the caller to a human, so a working agent routes people rather than trapping them. Design the transfer so the person who picks up sees what the caller already said, and set clear triggers: a complaint, a payment, a repeated misunderstanding or a caller who simply asks for a person. If you also run a website chatbot, our earlier guide on AI chatbots for websites covers the same handoff from the text side.

Outbound calls and the law

Answering calls and placing them are very different legally. In the United States, the FCC confirmed in a February 2024 declaratory ruling that AI-generated voices count as an “artificial or prerecorded voice” under the Telephone Consumer Protection Act, so AI calls need the prior express consent of the person called, and marketing calls need prior express written consent. An FCC order published in the Federal Register on 26 October, effective 25 November 2026, updates how people withdraw that consent: callers may designate one clear opt-out method, such as a key press or voice command on the call, must disclose it on the call, and must honour requests within ten business days. An opt-out from a telemarketing call still stops all telemarketing calls from that caller.

Outside the US, transparency rules are tightening too, including the EU AI Act's requirement to tell people when they are talking to an AI, which we will cover in a separate guide. If you plan outbound AI calls in any country, get consent records and opt-out handling right before you worry about voice quality.

An eight-point checklist before you switch one on

• Start with the calls you miss: after-hours, busy periods and overflow, not the calls your team already handles well.
• Limit the first version to questions, messages and bookings, and send payments and complaints to a person.
• Have the agent say it is an AI in its first sentence.
• Give callers a short, obvious route to a human, and pass the conversation so far to whoever picks up.
• Ground answers in your real hours, prices and policies, and update them when anything changes.
• Test in your callers' actual languages, accents and noise levels before going live, and listen to real recordings every week.
• For outbound calls, keep written consent records, a disclosed opt-out method and a process to honour opt-outs within ten business days.
• Measure transfers, repeat calls and complaints, not only the share of calls the agent handled on its own.

Where CoreCodify fits

AI voice agents are now good enough to answer the calls a small business keeps missing, and cheap enough to try. They still fail when they are asked to do everything, hide that they are an AI or keep talking when they have not understood.

At CoreCodify we design and build custom websites, web apps and AI features, including voice and chat agents connected to your booking, CRM and order systems with clear handoffs to your team. If you want a voice agent that works for your callers and your business, hire me for custom work at usamify.com.