Voice agents on websites: use cases, costs and pitfalls

TL;DR

• A website voice agent lets visitors talk to your site through the browser's microphone and hear spoken answers drawn from the same content as your text chat.
• Voice helps where typing is awkward or explanations run long. It gets in the way where visitors need to scan details, such as price tables or spelled-out email addresses.
• As of October 2026, published prices for the voice layer work out to roughly 1 to 31 cents per conversation minute, before the language model, long conversations and retries are added. Plan for latency, microphone permissions, AI disclosure and a text fallback.

What a website voice agent is

A visitor clicks a microphone button in your website's chat widget. The browser asks for permission to use the microphone. The visitor asks a question out loud; the agent works out the answer from your content and replies in a synthetic voice, usually with a live transcript on screen.

It is not a phone system. There is no phone number and no carrier, and it runs entirely in the browser. It is the same assistant as your text chat with a different way in and out.

There are two common ways to build one:

Either way, the agent should answer from the same knowledge base as your chat, with the same rules about sources and handoffs.

Use cases that make sense

Voice is not a better chat. It is a different one, and it fits some moments better than others.

Good fits

• Mobile visitors with a long question. Describing a situation ("we're a 12-person dental practice with two locations and we use Google Workspace") is faster said than typed on a phone.
• Guided product discovery. When the right answer depends on a few follow-up questions, a spoken back-and-forth can feel like talking to a knowledgeable salesperson.
• Hands-busy visitors. People in workshops, warehouses or on the road who can talk but can't type comfortably.
• Visitors who prefer speaking. For some people, including some with motor or reading difficulties, speaking is easier than typing. That only helps if voice is offered as an option, never as the only way in.
• Demos and walkthroughs. "Show me how to set up a report" while the page highlights each step.

Poor fits

• Comparing prices and specs. Audio is linear; tables are not. Let the agent say a one-line summary and show the table on screen.
• Collecting exact details. Email addresses, order numbers and spelled names go wrong easily when spoken. Switch to a text field.
• Public spaces and open-plan offices. Many visitors can't or won't talk to a website at their desk.
• Private topics. Few people want to say their health or finance questions out loud.

How it works in the browser

Microphone access goes through the browser's getUserMedia() API. According to MDN, it is available only in secure contexts (HTTPS), it needs the user's permission, which the browser asks for at least the first time, and the request can "neither resolve nor reject" if the user ignores the prompt (MDN). If your widget runs inside an iframe, the parent page has to allow the microphone explicitly through Permissions Policy:

<iframe src="https://widget.example.com/voice" allow="microphone"></iframe>

// Ask for the microphone only after the visitor clicks "Talk".
const stream = await navigator.mediaDevices.getUserMedia({ audio: true });

Two rules follow. Ask for the microphone only when the visitor chooses voice, never on page load. And handle every outcome of the prompt: allowed, denied, and ignored, with a timeout that falls back to text.

What it costs, as of October 2026

Voice vendors price in different units, which makes comparison awkward. Here are the published list prices we found, checked on October 20, 2026:

Sources: OpenAI pricing, ElevenLabs, Vapi, Retell AI. Prices change often; check before you budget.

To turn OpenAI's token prices into minutes, its Realtime documentation says audio in user messages is "1 token per 100 ms" and audio in assistant messages is "1 token per 50ms" (OpenAI). That is 600 tokens per minute of visitor speech and 1,200 per minute of agent speech.

A worked example

Take a five-minute conversation in which the visitor speaks for about two minutes and the agent for about two. Counting speaking time only:

Real bills come out higher than this simple math for a structural reason. OpenAI's cost guide explains that "the entire conversation is sent to the model for each Response," so "turns later in the session will be more expensive." The conversation so far is resent as input on every turn, up to the model's context window. Caching reduces the cost of that repeated input, but OpenAI describes caching as "best-effort and not guaranteed." Measure your actual cost per conversation in a pilot before you commit to a budget.

Keeping costs under control

• Cap session length and end sessions after a period of silence.
• Keep spoken answers short. At OpenAI's realtime rates above, a minute of agent speech costs about four times as much as a minute of visitor speech.
• Use a smaller model for routine questions.
• Limit sessions per visitor to stop abuse.
• Track cost per conversation and cost per outcome, not only cost per minute.

Pitfalls to plan for

Slow replies break the conversation

People are finely tuned to the rhythm of conversation. A study of ten languages found "a general avoidance of overlapping talk and a minimization of silence between conversational turns," with average gaps that differ by language only within 250 milliseconds of the cross-language mean (Stivers et al., 2009). A voice agent that takes a couple of seconds to start answering feels like a bad phone line. Measure the time from the moment the visitor stops speaking to the first sound of the reply, and test on ordinary mobile connections, not only office Wi-Fi.

Interruptions and echo

Visitors interrupt, and the agent must stop talking when they do. On laptops with speakers, the agent's own voice can leak back into the microphone and be heard as a new question. Test with real hardware, including cheap headsets and laptop speakers, before launch.

Recognition errors on the details that matter

Names, product codes, numbers and email addresses are where speech recognition errors hurt most. Read critical details back for confirmation, show the transcript on screen, and switch to a typed field for anything the visitor must get exactly right.

Microphone permission friction

Some visitors will deny the permission prompt, and some will ignore it. Treat both as normal. Keep text input visible at all times, and never make voice the only path to an answer.

Disclosure: say it is an AI

A natural-sounding voice makes it easy for visitors to assume they are talking to a person. In the EU, Article 50(1) of the AI Act requires providers to design AI systems that interact directly with people so that those people are "informed that they are interacting with an AI system," unless that is obvious from the context (AI Act Service Desk). The regulation's general application date is 2 August 2026 (Article 113). Whatever your jurisdiction, have the agent say it is an AI assistant in its first sentence and label the widget. This is not legal advice; check the details with your counsel.

Recordings and privacy

Decide before launch whether you store audio at all or only transcripts, for how long, and who can access them. Describe it in your privacy notice, and ask your counsel about consent requirements in the places your visitors come from.

Answers written for reading, spoken aloud

A paragraph that reads fine on screen is tiring to listen to. Instruct the agent to answer in two or three spoken sentences, then offer to show links or details on screen.

No way out

Some conversations need a person. Give the voice agent the same human handoff paths as your chat: book a call, continue in text, or leave a message for the team.

A launch checklist

• Voice is opt-in behind a clear "Talk" button; text input is always available.
• The site runs on HTTPS, and the iframe allows the microphone if the widget is embedded.
• The first sentence says the visitor is talking to an AI assistant.
• A live transcript is shown on screen.
• Spoken answers are short, with links and tables shown visually.
• Emails, phone numbers and order numbers are confirmed or collected as text.
• Handoff to a person works from voice.
• Session length caps, idle timeouts and per-visitor limits are set.
• Cost per conversation is tracked from day one.
• It has been tested on phones, on laptops with speakers, in noisy rooms and with a range of accents.

Measuring a voice agent

Measure it like any website agent, by outcomes rather than activity. Our guide to measuring an AI agent on your website covers the basics. Voice adds a few metrics of its own: the share of visitors who grant microphone permission, time to first audio, how often visitors interrupt, how often they switch from voice to text, and cost per minute next to cost per resolved conversation.

Accuracy rules don't change because the answer is spoken. A wrong answer said in a confident voice is still a wrong answer, so the controls in our guide to grounded answers apply in full. And if you are still deciding what the assistant should be allowed to do, start with chatbot vs AI agent.

ASDK's Voice agent, which will let visitors talk to your website using the same knowledge as chat, is coming soon.