ElevenLabs Bets Disclosure Builds Trust While Margins Shrink

Before a company rolls out an AI voice agent to answer its phones, someone has to decide whether callers get told they're talking to software. That decision — simple on paper, thorny in practice — sits at the center of a TechCrunch interview with ElevenLabs co-founder and CEO Mati Staniszewski, conducted at the Nrth conference in Toronto. His answer has implications well beyond his own company: ElevenLabs' voice technology already handles first-line phone support for 35 million U.S. customers of Klarna, along with deployments at Deutsche Telekom, Cisco, Adobe, and multiple national governments, according to the interview.

The Core Tension: Growth Now, Margins Later

Staniszewski confirmed ElevenLabs is pacing at $600 million in annual recurring revenue and is reportedly valued at $22 billion, a figure that has arrived just four years after founding. When pressed on gross margins — a question that matters enormously to anyone evaluating whether a vendor's pricing is sustainable — he declined specifics but was candid about strategy: he said he doesn't mind margins compressing further if it expands market share, framing the priority as proving value to customers now over protecting profitability today. That's a founder-stage answer from a company operating at enterprise scale, and it's worth flagging plainly: this is a stated intent, not an audited financial disclosure. Buyers negotiating multi-year voice-AI contracts should treat any current pricing as provisional, not a floor.

What's Confirmed, What's Reported, What's Speculation

Separating the record matters here because the story blends three different kinds of claims.

Confirmed by the CEO in the interview: the $600 million ARR figure, the 55%-plus enterprise revenue split, the Klarna and government deployments, the company's practice of letting customers choose between frontier lab models, open-weight models, or their own fine-tuned models depending on the use case, and the claim that every customer undergoes KYC screening.

Reported but not confirmed: the $22 billion valuation is attributed to backers, not stated directly as fact by Staniszewski. Likewise, a 2028 IPO timeline has been reported elsewhere, but when asked directly, Staniszewski would only say the company is "preparing the foundation" and that timing is genuinely undecided — he explicitly called his own use of the word "years" vague.

Speculation, offered as personal prediction rather than data: Staniszewski's belief that in five years, society will have shifted to expecting AI agents by default, making disclosure less fraught than it is today. That's a forecast about cultural adaptation, not a documented trend, and founders should weigh it as such.

The Disclosure Question, Concretely

Here's a scenario that clarifies the stakes. A mid-sized insurer deploys a voice agent to handle claims status calls. Two design choices produce very different trust outcomes:

• Silent deployment: The agent sounds human, doesn't announce itself, and a customer later learns — via a support forum complaint or a journalist's inquiry — that they'd been talking to software during a sensitive claims conversation. The reputational cost compounds because the deception, however unintentional, becomes the story.
• Disclosed with an opt-out: The system states upfront that the customer can speak with a human but faces a wait, and offers the agent as an immediate alternative. Staniszewski's account of ElevenLabs' own deployments claims that in this framing, customers overwhelmingly choose the agent and are pleasantly surprised by the experience.

The second pattern is the one Staniszewski says he favors industry-wide, and it lines up with a broader principle in trust design: disclosure paired with genuine choice tends to convert skepticism into acceptance, while disclosure imposed as a formality after the fact does not. That said, ElevenLabs has an obvious commercial incentive to normalize agent-first interactions, so operators should validate this pattern with their own customer data rather than assume it transfers automatically.

Why Vendor Overlap Should Worry Procurement Teams

One detail in the interview deserves more attention from buyers than it's likely to get: ElevenLabs' own customer Decagon trained its voice product on ElevenLabs and now competes with it directly. Staniszewski's response — that the old boundaries between "model company, platform company, application company" have blurred, citing Anthropic's trajectory as a parallel — is a reasonable industry observation, but it also signals something practical for enterprise buyers. If you build a voice-AI stack on a vendor whose own tooling can be used to train a rival product, your competitive moat may be thinner than your contract implies. This is inference on UpShaqo's part, not a claim made in the source, but it follows directly from the facts disclosed.

The Model Choice Matters More Than the Marketing

Staniszewski's answer on frontier versus open-weight models offers a genuinely useful operating framework, not just a talking point. For informational, non-transactional interactions — status updates, FAQ-style queries — open-weight models can perform adequately because the deployer's own knowledge base constrains the acceptable range of answers. For anything involving authentication, financial transactions, or refunds, he argues frontier models remain necessary because there's "no room for error." The Polish government deployment he describes — automated reminder calls addressing an 18% no-show rate in public health appointments — illustrates a lower-stakes, informational use case where model flexibility and data residency requirements can both be satisfied simultaneously.

Practical Controls for Teams Evaluating Voice AI

Based strictly on what's documented here, a few controls follow naturally:

• Require explicit disclosure language in any customer-facing voice deployment, and test whether offering a human-vs-agent choice changes acceptance rates in your own environment.
• Match model tier to transaction risk: reserve frontier models for authenticated, financial, or irreversible actions; open-weight or fine-tuned models are plausible for lower-stakes informational calls.
• Ask any voice-AI vendor directly whether your training data or usage patterns could inform a competing product, and get contractual protection in writing rather than accepting general assurances about "blurred lines."
• Treat vendor margin commentary as strategic signaling, not a pricing guarantee — renegotiate rather than assume today's rates hold as the vendor matures toward a potential IPO.

What Remains Unresolved

The interview leaves open exactly how ElevenLabs' gross margins compare to peers, whether the 2028 IPO timeline will hold, and how regulators in the U.S. and Europe will eventually formalize disclosure requirements that Staniszewski currently frames as voluntary best practice. It's also unclear how enforceable his claim about preventing agents from creating other agents will prove as the underlying frontier models — which ElevenLabs doesn't train — continue to advance in autonomy. Those are the questions worth tracking as voice AI moves from novelty to default infrastructure in customer service.

Sources

TechCrunch AI — ElevenLabs' CEO on margins, IPO timing, and telling customers they're talking to a bot