Voice AI Is Eating the Call Center, One Angry Caller at a Time

Think about the last time you called a company you were annoyed with. The hold music. The menu that would not let you say "agent." The 14 minutes before a human picked up and asked you to repeat everything you had already typed. That specific, universal misery is now the target of one of the best-funded races in enterprise software, and a handful of companies believe they can end it by removing the human entirely.

The list reads like a startup roll call: Observe.AI in Redwood City, PolyAI in London, Botpress out of Quebec City, Verloop.io in Bengaluru, Forethought in San Francisco, plus the older heavyweights LivePerson, Oracle and Omilia. On paper they all do "conversational AI for customer service." In practice they are playing three very different games, and the money is starting to pick favorites.

Three camps, one phone call

---

The quickest way to make sense of this market is to stop treating it as one market. There are really three jobs on offer, and most companies started with one before creeping into the others.

Observe.AI is the clearest example of a company that changed lanes. It started by listening - transcribing and scoring contact-center calls so managers could see which agents followed the script. In 2025 it pushed into the other two camps at once, shipping VoiceAI agents that handle calls outright and GenAI Insights that route "voice of customer" findings to the rest of the business. When the tool that grades your agents can also replace them, the category lines get blurry fast.

The money follows the voice

---

If you want to know where investors think the value sits, watch the voice specialists. In December 2025, PolyAI raised $86 million at a $750 million valuation, a 50% jump from its previous round, crossing $200 million raised in total. Its reported ARR doubled to around $40 million, with U.S. revenue nearly tripling. The founders - Nikola Mrksic, Pei-Hao Su and Tsung-Hsien Wen - met at Cambridge's Machine Intelligence Lab studying dialogue systems, which is a polite way of saying they have been trying to make computers hold a conversation since before it was fashionable.

The number that makes executives lean forward is not the valuation. It is the claim that at PolyAI's largest deployments, voice agents do the work of roughly 1,000 full-time staff, across about 45 languages. Its customer list - Marriott, Caesars, PG&E, Foot Locker - is the kind of roster that turns a novelty into a line item. Verloop.io is chasing the same prize from India, marketing voice agents that handle calls in 80-plus languages. Omilia and Oracle bring the enterprise credibility and the compliance paperwork that regulated industries insist on.

It is worth pausing on why voice, specifically, is where the capital concentrated. Text was the easy win - chat windows and help-center bots have existed for years, and customers largely tolerate them. Voice is harder and worth more. A phone line carries the callers who are frustrated enough to skip the FAQ, the ones on hold, the ones costing a company real money in staffed minutes. Automate a text chat and you save a few cents. Automate a live call that would otherwise tie up a person for eight minutes and you save something an operations director will notice on a spreadsheet. That is the arbitrage every voice specialist is quietly pricing against.

The incumbent problem

---

Not everyone is riding the wave. LivePerson has been in this business since 1995, back when "conversational AI" meant a live-chat window on a website. It should have owned this moment. Instead its revenue fell about 22% in fiscal 2024 to $312.5 million as enterprise customers churned, and the company spent the following year restructuring around its core and its debt. Then, in 2026, SoundHound agreed to buy it for around $43 million - roughly what a hot startup raises in a single Series A.

That deal is the tell. A once-dominant public company, home to on the order of a billion messages a month, changing hands for the price of a seed-stage bet. The lesson is not that LivePerson built bad technology. It is that in AI, being early buys you nothing if you get slow. SoundHound wanted the enterprise relationships and the digital-messaging volume to bolt onto its own voice agents. The distribution was worth more than the brand.

Botpress is the counter-story from the opposite end. Founded the same year as PolyAI and Observe.AI, it went open source under an MIT license and let developers run its agent platform for free, with no per-seat cost. It raised a comparatively modest $25 million Series B in 2025 at roughly a $120 million valuation, betting that the winning strategy is reach, not managed service. Forethought, which won TechCrunch Disrupt's Battlefield back in 2018, has raised north of $100 million and stayed narrow, focused on autonomously resolving support tickets rather than answering the phone.

Does it actually work?

---

Here is the honest part. The technology has moved faster than the trust. The old chatbot followed a rigid decision tree and shattered the moment you said something it did not expect. Today's systems sit on large language models, so they follow messy natural speech, keep context across a conversation, and can take real actions - issue the refund, move the reservation, close the ticket. For routine, high-volume calls, they are genuinely good now.

The trouble starts at the edges. Complex problems, emotional ones, the caller who is already furious - those still need a person, and the best vendors design for that handoff instead of pretending it will not happen. The uncomfortable truth for every company in this story is that the software is often ready before the customer is willing. What they are really selling is not a voice. It is permission to let a machine handle the moment a human used to.

There is also a quieter product buried in all of this: the data. Every automated conversation becomes a labeled record of what customers want, where they get stuck, and which promises the company keeps. Observe.AI built its early business on exactly that insight - that the transcript is often worth more than the call. A contact center that used to be a black box of manager spot-checks turns into a searchable record of demand. For a lot of buyers, the analytics are the reason to sign, and the automation is the part they roll out slowly once they trust what the transcripts are telling them.

Who actually wins

---

There is no single winner here, and anyone who tells you otherwise is selling something. The positions are too different. PolyAI is winning the funding-momentum and voice-quality race. Botpress is winning developer reach. Observe.AI is winning the "we already sit inside your contact center" argument. LivePerson lost the pace it needed and became a component in someone else's stack.

If you run a support operation, the practical question is not which logo is hottest. It is build versus buy. Platforms like Botpress hand technical teams full control and no per-seat tax. Managed voice players like PolyAI or Omilia get an enterprise live faster with less engineering. Analytics-first tools like Observe.AI layer onto what you already have. The right answer depends on how much control you want and how quickly you need to move.

What is not in doubt is the direction. The call center used to be a cost you offshored. Now it is a data source you automate, and every conversation that used to end in a manager's spot-check now ends in a training set. The 2026 consolidation - SoundHound swallowing LivePerson - is almost certainly the first of several. Somewhere out there, a person is still on hold, listening to a saxophone loop, about to be handed to a machine that has heard this complaint a million times and will not sigh. Whether that feels like progress depends entirely on how the handoff goes.