ElevenLabs' CEO will squeeze margins for share in a voice market he expects to even out

A patient booked into Poland's public health system gets a phone call reminding them of the appointment. The caller is an ElevenLabs agent. Staniszewski told TechCrunch it is integrated with models the Polish side had already tuned on its own knowledge, and the data stays in the country ^[16]. The calls go out because 18% of booked patients never show up ^[16].

The pitch is grander. Staniszewski wants ElevenLabs to be the first to pass the Turing test for conversational AI, and he said that takes emotional intelligence: "You need to understand the emotions of the other side, to be able to slow down or speak up" ^[11]. That pitch assumes callers reward the most human-sounding voice. What callers actually do, by TechCrunch's account, is often fail to notice they are talking to an AI at all ^[1].

Once the voice clears that bar, Staniszewski himself puts call quality somewhere else. For informational calls, "you can use a lot of the open-source models because your knowledge base defines what a good experience is," he said ^[14]. ElevenLabs lets customers pick that reasoning layer from a menu of frontier and open-weight models ^[13]. The exception he named is financial services, where a caller needs authentication, transaction details or a refund: "There's no room for error. Here, frontier models will still lead" ^[15].

So a voice agent is at least two purchases, and the voice is the one its own seller expects to converge. At TechCrunch Disrupt last year, Staniszewski predicted audio models would be commoditized within a couple of years ^[8]. His timeline is now longer ^[3]. "There is still a lot of work to be done, and the quality delta you can achieve just on the model level is still significant," he said ^[9]. "If we think longer term, probably three, five years from now, those differences will be smaller" ^[10]. Asked about Decagon, he said: "The lines become more blurry" ^[17].

Investors are paying for the lead as it stands. The reported valuation is about 37 times the company's stated revenue pace ^[2]. At least 55% of revenue is classic enterprise ^[12], or roughly $330 million a year ^[1]. Klarna shares the customer list with Deutsche Telekom, Cisco, Adobe and a growing list of governments ^[3]. The interview does not include resolution rates or a price per call for any of these deployments.

For a buyer, two axes sort each call type. One is whether the call moves money or only answers questions. The other is whether your team or the vendor owns the knowledge base and the call flow.

Informational calls running on your own knowledge base make the voice the most replaceable part. I'd buy those on price with short terms, and treat the margin ElevenLabs says it is willing to give up as leverage ^[7]. The tradeoff is switching work at each renewal, through the three to five years in which Staniszewski says model differences still count ^[10].

Transactional calls on your own flow need the frontier reasoning model he says still leads there ^[15]. Lock-in on the voice is the smaller risk in that box.

When the vendor owns the flow, in either column, the buyer is renting the part that decides whether a call goes well ^[14]. Decagon shows a platform can build on ElevenLabs and then move its queries to its own models ^[4].