ElevenLabs, Voice Cloning and What Actually Changes for Agents

In February 2026, ElevenLabs closed a $500 million round at an $11 billion valuation, led by Sequoia — tripling its worth in a year, according to Sacra’s breakdown. The company passed $500 million in annual recurring revenue and is used by 41% of Fortune 500 companies.

Numbers like that usually become a headline and stop there. What interests me is something else: what changes on the desk of people who actually run agents.

I spent two decades building search engines before becoming an AI founder. I’ve seen this curve: a feature becomes a product, a product becomes a platform, a platform becomes infrastructure nobody argues about anymore. Synthetic voice went through that in record time. In 2023 it was a conference demo. In 2026 it’s the front line of bank support.

The proof isn’t in the demo, it’s in the support queue

Two 2026 deployments say more than any benchmark.

In January, Revolut put ElevenLabs voice agents into customer support across the UK and Europe, covering more than 4 million customers in 30+ languages, and reported an 8x reduction in time-to-resolution. In February, Klarna launched a voice agent as first-line phone support for 35 million US customers, reporting resolutions up to 10x faster, as compiled by TextToLab.

Notice the detail almost everyone skips: neither of those numbers is about voice. They’re about resolution. Voice is the door. What solves the problem is what stands behind it.

Voice is an interface, not a product

Here’s the most expensive mistake I see companies make in 2026: treating voice cloning as if it were the project.

Cloning a voice today is a commodity. You record a few minutes, upload, and a convincing voice comes out the other side. It’s impressive the first time and irrelevant by the tenth. Because a cloned voice on top of a stupid agent is a stupid agent that now sounds exactly like you.

What sets the Revolut deployment apart is everything underneath: the agent knows who the customer is, has access to history, knows the company’s policy, and knows when to stop and hand off to a human. That isn’t an audio problem. It’s a memory, context and governance problem.

That’s exactly why, at Arvor, voice comes second. First we organise the knowledge — BRAIN MAKER turns documents, decisions and history into a queryable memory with the source attached. Then we wire agents to that memory on WhatsApp and Telegram, where people already talk. Only then does voice earn its keep: the lawyer dictates on the way out of the hearing, the doctor dictates between patients, the engineer asks with a hard hat on and hands full.

In all three cases the value isn’t a beautiful voice. It’s the right answer arriving at a moment when typing was impossible.

The part nobody wants to discuss

Cheap, convincing voice cloning carries a social cost, and it has already arrived.

Voice notes are social proof in a lot of markets — Brazil especially. People recognise a relative by voice and act on it. When three minutes of public audio is enough to reproduce someone convincingly, the impersonation scam stops depending on a good actor and starts depending on a file.

That’s not an argument against the technology — it’s an argument against using it thoughtlessly. If you’re going to put a synthetic voice speaking on behalf of your company, decide first: how does the customer confirm it’s really you, what will the agent never ask for by voice, and where does human review sit. Those three questions matter more than any model choice.

Where to start

If you want to test the voice layer, ElevenLabs is the market reference today — voice cloning, two TTS families, speech recognition in 90+ languages and its own conversational agents platform. You can try it at try.elevenlabs.io.

Disclosure: that’s an affiliate link. If you subscribe through it, Arvor earns a commission at no extra cost to you. We recommend it because we use it — the commission doesn’t change the technical assessment above.

Now the honest advice: don’t start with voice. Start with what the agent needs to know in order to answer well. Voice on top of disorganised knowledge just makes the wrong answer arrive faster, with a better accent.

If that’s your problem — scattered knowledge, a specialist turning into a bottleneck, an agent that sounds great and misses the point — it’s literally what we do in agentic consulting: diagnosis, organisation and agents in production. Not slides.