ElevenLabs at $22bn: buyers should judge voice agents on resolution

ElevenLabs said on 30 September that a $300m (£226m) employee tender offer values the company at $22bn (£16.6bn), double its valuation at its Series D in February. Wellington and T. Rowe Price led the deal, and EQT, Goldman Sachs and GIC were among six investors buying in for the first time. A tender is a sale of shares by staff and existing holders; as City AM notes, the company itself raises no new money.

What the company says is driving much of its growth is ElevenAgents, the conversational agent platform it has built alongside its voice models. ElevenLabs says enterprise customers now account for 55% of its revenue, that its agents handle more than 15 million conversations a week – three times February’s level – and that ElevenAgents’ annual recurring revenue (ARR) has more than tripled over the same period. It names Stripe, Deutsche Telekom, DoorDash’s SevenRooms and the insurer Admiral among the organizations using the agents, and says its agents handle jobs such as refunds, insurance renewals, phone-plan upgrades and healthcare bookings.

“AI should interact with people the way we interact with each other,” said Mati Staniszewski, co-founder and chief executive of ElevenLabs, in the company’s announcement.

The figures are the company’s own, and they measure volume. For the CIOs, CTOs and customer operations leaders deciding whether to let an AI agent answer the phone, volume is the wrong test. The question to put to any voice agent supplier, ElevenLabs included, is how many callers got what they rang for, and what happened to those who did not. The valuation shows that investors think voice agents have arrived in the contact center. It does not show how to buy one well.

Speed is only half a quality measure

The announcement offers one figure on quality. A recent analysis of ElevenLabs’ customers, it says, showed voice agents resolve issues 31% faster than chat agents on average. The post gives no sample, no method and no figure for how many issues were resolved at all. Speed matters on a phone line. The risk is that a quick call which leaves the problem open simply moves the cost to the next call, or to a complaint.

Admiral measures the outcome

A more useful document for buyers is one ElevenLabs published on 10 August: a write-up of a webinar in which Admiral’s generative AI team explained how it runs AI agents on customer calls. It is vendor marketing about a customer the vendor chose, and should be read that way. But the controls it describes are specific.

Admiral, which handles millions of customer interactions a year in markets including the UK, Italy, France and Spain, made 90% first contact resolution its primary target, according to the write-up. Containment – calls the agent handles without passing them to a person – was a secondary measure. Admiral tracks whether the customer got what they called for separately from whether the call was escalated, so a deliberate handoff to a human is never counted as a failure. Its bar for production is that the agent performs as well as its human staff, or better, before it goes live.

The boundaries are built into the design. In a settlement-quote agent for Admiral Money, a screening sub-agent checks every caller, and customers who are vulnerable or in arrears go to a priority queue with a human while the team learns more edge cases. Getting a quote takes about 2.5 minutes with the agent against about 5 minutes with a person, and the write-up credits the saving to an API-driven workflow rather than to AI alone. At L’Olivier in France, changes to the agent go to 1% of traffic first and step up to 100%, with simulation tests and a manual preview before anything reaches live traffic. If the data does not support a change, traffic goes back to zero at once.

Buyers should own the terms

ElevenLabs says its forward-deployed engineers and researchers work with organizations to refine its models and agents for their conditions. That help will shape many deployments. The risk is that the supplier then also defines what success looks like. Before a voice agent takes live calls, buyers should settle four things themselves:

• The main metric. Resolution, measured separately from escalation and reported to the buyer, rather than conversations handled.
• The go-live bar. The agent matches the human team on the same cases before it takes customers.
• Who the agent must pass on. Named groups of callers, such as vulnerable customers or those in arrears, go straight to a person.
• How changes ship. New versions reach a small share of calls first, with a rollback that takes one step.

None of this depends on ElevenLabs’ valuation, and all of it applies to its rivals. The $22bn price is, in effect, a bet that voice agents will take on a large share of customer calls. Buyers do not have to share that bet to use the technology. They do have to decide what a good call looks like before the supplier does. Organizations that count resolved calls will know quickly whether an agent works. Those that count conversations will find out from their complaints.