Are AI Voice Agents Ready for B2B Work Yet?

Are AI Voice Agents Actually Production Ready?

As of September 2026, for narrow jobs, yes. The platform pieces are generally available rather than experimental. OpenAI's Realtime API is out of beta with a production model, and ElevenLabs is running voice agents for named enterprise customers. The open question is no longer whether they work. It is what you should trust them with.

We get asked about this because voice sits right next to the website work we do. A company that already runs a chat widget and a lead routing workflow wants to know whether adding a phone answering agent is a sensible next step or a support nightmare.

Here is what the vendors have actually shipped, what the technology cannot do yet, and how we would pilot one without putting a customer relationship at risk.

What Changed in the Last Year?

Voice moved from a demo stack to an API with production features. OpenAI's developer documentation describes a generally available Realtime API built around a model named gpt-realtime, with a GA snapshot labelled gpt-realtime-2025-08-28. That naming alone tells you the platform is now versioned and supported rather than experimental.

The feature list is what matters. OpenAI documents support for WebRTC and SIP connections, so an agent can live in a browser or answer an actual phone line. It documents image input on the GA model, and support for Model Context Protocol servers, which lets a voice agent reach the same tools your other automations use.

One addition deserves attention. OpenAI documents async function calling, described as preventing the model from hallucinating while an operation is pending. Anyone who has built a voice flow knows why that exists. The old failure mode was an agent inventing an answer while waiting for a database lookup, out loud, to a customer.

How Much Money Is Behind This?

Enough that the category is not going away. ElevenLabs announced a 500 million dollar Series D on 4 February 2026 at an 11 billion dollar valuation, led by Sequoia Capital. In the same post it said it closed 2025 with over 330 million dollars in annual recurring revenue.

The customer list in that announcement is the more useful signal for a B2B buyer. ElevenLabs names Deutsche Telekom, Square, Revolut, Duolingo, NVIDIA, Meta, Salesforce and Harvey among organisations using its platform, and describes its agents being used for customer support, conversational commerce, citizen engagement, internal training and inbound sales.

Read that use case list carefully, because it is a map of where this is working. Support, commerce and inbound sales are all high volume, script adjacent and easy to escalate out of. None of them is complex negotiation.

What Can a Voice Agent Genuinely Do in B2B?

Handle the front of the funnel and the front of the queue. Qualifying an inbound call, booking a meeting, confirming details, answering questions that live in your documentation, and routing the caller to the right human. These are bounded tasks with clear success conditions.

The strongest case we see is inbound call qualification for companies whose buyers still phone. A voice agent that answers immediately at 9pm, captures the requirement accurately and books time with a human beats a voicemail box by a wide margin. The bar it has to clear is not a great salesperson. It is nobody answering.

Outbound is a different question, and we would be careful. An unsolicited synthetic voice call is a brand decision and often a regulatory one, and the rules vary by country. That is a conversation for your legal team before it is a conversation for your automation team.

Where Do Voice Agents Still Fail?

Wherever the conversation stops being a form. Interruptions, accents the model handles poorly, background noise, callers who change their mind mid sentence, and anything requiring the agent to say it does not know. Text chat forgives a pause. Voice does not, and every awkward gap is audible.

The harder failure is subtler. A voice agent that is confidently wrong is worse than a chatbot that is confidently wrong, because there is no transcript in front of the caller to check and no scroll back. The customer remembers what they think they heard, and so does your dispute later.

We would not put a voice agent in front of pricing negotiation, contract terms, complaints, or anything involving a distressed customer. Those are the conversations where the human relationship is the product.

What Are the Technical Limits You Should Design Around?

Three published ones shape the architecture. OpenAI's documentation states a maximum session duration of 60 minutes, a 32,768 token context window for the realtime model, and an idle timeout capability for detecting unresponsive users. Design your flows to finish well inside those bounds.

The context window is the constraint people underestimate. A voice session accumulates transcript quickly, and a long call plus a large system prompt plus tool results will hit that ceiling. Keep the prompt tight, summarise as you go, and hand off before the window becomes the reason the agent gets confused.

OpenAI also documents EU data residency support, which is often the first question a European client asks and the fastest way to end a procurement conversation if the answer is no. Check it before you build, not after.

Is a Voice Agent a Security Problem?

It is the same problem as any tool using agent, with a worse audit trail. Once a voice agent can call your systems through Model Context Protocol or a custom tool, it inherits every risk in the OWASP Top 10 for LLM Applications, where prompt injection is listed as LLM01, the first entry.

Voice adds its own wrinkle. A caller is untrusted input, and spoken instructions reach the model exactly like typed ones. Someone can simply say the injection out loud. The defences are the same ones we apply everywhere: least privilege on the agent's credentials, validated output formats, and human approval on anything irreversible.

OWASP also published a Top 10 for Agentic Applications on 9 December 2025, naming identity and privilege abuse and tool misuse among its highlighted threats. A voice agent with write access to your CRM is squarely in that territory. We go through the controls in our guide to MCP servers for web teams.

Should Voice Replace Your Website Chat?

No. It should sit beside it, and probably behind it. Chat is cheaper to run, easier to log, easier to correct mid conversation and far easier for a buyer at their desk who does not want to talk. Most B2B research happens silently, and forcing voice on that is a worse experience.

The exception is the phone line you already have. If calls come in and go unanswered, or land with someone who takes a message, a voice agent is competing against a genuinely poor baseline rather than against a good chat widget.

Our honest read is that voice is an additional channel for a specific caller, not an upgrade to your existing one. If your chat experience is weak, fix that first, because it is cheaper and it serves more of your buyers. We covered that groundwork in our piece on running an AI chatbot on your website.

How Would We Pilot One?

On one call type, with a hard escalation rule, for a fixed period. Pick the single highest volume, lowest stakes reason people call you. Give the agent one job on that call. Give it a phrase that hands the caller to a human, and make sure that handoff actually works before anything else ships.

Then measure three things. What proportion of calls completed without escalation, what proportion of the captured data was correct, and what callers said about the experience afterwards. That third one needs asking directly, because a completed call is not the same as a satisfied caller.

Record and review transcripts weekly for the first month, in full, by a person. This is tedious and it is the only way to find out what your agent is actually saying. The downstream data quality question matters too, which we cover in our piece on AI lead qualification in your CRM.

Where Is This Going Next Year?

Towards voice agents that carry real tools rather than scripts. The direction is already visible in what the platforms shipped: phone connectivity through SIP, tool access through Model Context Protocol, and async function calling so the agent can wait for a system without filling the silence with invention.

What we would not bet on is voice replacing your other channels. The interesting version of this is a voice agent as one interface onto the same automation layer that serves your chat, your forms and your internal tools. Build the layer well and the channel becomes a choice rather than a rebuild.

If you are weighing up a voice pilot, or you want the automation layer underneath it built properly first, we are happy to think it through with you. Find us at phoenix.studio.