Conversational AI: what are the top startups?

Which companies actually count as conversational AI startups now?

For this ranking, conversational AI startups are companies whose core product either runs real conversations for businesses or provides the voice and real-time infrastructure those conversations depend on.

That definition gives us three different groups that should be compared carefully. Sierra, Decagon, Parloa, PolyAI and Ada sell AI agents that talk directly with customers and resolve work. Retell AI and Vapi provide the infrastructure companies use to build and operate voice agents. ElevenLabs sits between those two worlds because it owns important voice technology while increasingly selling complete conversational agents. LiveKit sits deeper in the stack, handling real-time audio and communications infrastructure.

We exclude OpenAI, Anthropic and Google because conversational products are only part of much broader foundation-model businesses. We also keep Character.AI and Replika outside the main ranking. Consumer companionship has very different economics from automating customer support, sales calls or appointment scheduling.

That distinction has become more important lately because the companies can look almost identical from the user's side while selling very different products underneath.

Part of the market	What customers actually buy	Startups we include
Enterprise AI agents	Automated conversations and completed customer tasks	Sierra, Decagon, Parloa, PolyAI, Ada
Voice-agent platforms	Infrastructure for building and running phone agents	Retell AI, Vapi, Bland AI
Voice and speech stack	Speech models plus increasingly complete agents	ElevenLabs
Real-time infrastructure	Low-latency audio, video and agent runtime	LiveKit

Did conversational AI finally move beyond chatbots?

Yes, enterprise conversational AI has moved well beyond scripted chatbots: the leading systems now resolve tasks, call external systems and stay involved across multi-step workflows.

The adoption numbers have changed quickly. Salesforce's latest State of Service research surveyed 3,075 service professionals and found that 66% of customer-service organizations represented in the study were using agentic AI, up from 39% a year earlier. Seventy percent of organizations using AI agents said they were seeing measurable value within 60 days.

We should still be careful with the word "agentic." A Rasa survey of 30 enterprise conversational-AI leaders found that 67% were expanding or scaling their programs, while confidence in AI handling complex conversations averaged only 4.37 out of 7. Sixty-three percent preferred hybrid architectures combining LLMs with deterministic logic; only 13% preferred fully agentic systems.

Customers are also less impressed by a fluent conversation than vendors sometimes assume. Research conducted by NewtonX for Ada across 2,000 consumers and 500 enterprise decision-makers found that just 24% of consumers had had their issue resolved by AI alone. At the same time, 59% preferred immediate AI support over waiting for a person when the AI could actually solve the problem.

That gap explains a lot of what the strongest startups are building now. Sierra has agents that process returns, dispute charges and work through mortgage-related tasks. Decagon recently added Browser Actions so its agents can operate web-based systems that lack clean APIs. Ada is pushing the same reasoning layer across voice, messaging and email. The market has moved from "can the bot understand me?" toward "can the agent finish what I asked?"

Is voice becoming the main battleground in conversational AI?

Yes, voice is currently the hardest and fastest-moving battleground in conversational AI, and public disclosures from just Retell AI, Vapi and Bland imply roughly 95 million to 215 million AI calls per month.

Retell says it powers more than 50 million real-time AI phone calls each month. Vapi says it processes between one million and five million calls per day, equivalent to roughly 30 million to 150 million per month. Bland recently told Fortune that it handles more than 3.5 million calls per week, another roughly 15 million per month.

Those figures use different definitions and should not be added as if they were audited market statistics. Even using them conservatively, though, they show an order of magnitude that barely existed a few years ago. And the estimate leaves out PolyAI, Sierra, ElevenLabs, Ada and the many agents running through LiveKit.

Capital is moving in the same direction. LiveKit raised $100 million at a $1 billion valuation after becoming the real-time infrastructure behind products including ChatGPT Voice Mode. PolyAI raised $86 million in its latest round. Vapi raised $50 million. Bland raised another $50 million. ElevenLabs reached an $11 billion valuation while telling investors that enterprise voice-agent deployments were helping drive its growth.

Voice also creates a much harsher product test. A text bot can pause for several seconds without looking broken. A phone agent cannot. Interruptions, latency, background noise, turn-taking, telephony failures, accents and escalation all happen in real time. That is why specialized companies can still survive even though everyone has access to increasingly capable foundation models.

If you want more recent data on this point, please see our latest conversational AI market report.

Is Sierra actually the top conversational AI startup today?

Sierra is currently the strongest pure-play enterprise conversational AI startup we found.

Sierra reported crossing $100 million in ARR seven quarters after launch and then said it entered its third year above $150 million after posting a $50 million quarter. Sacra later estimated that the company had reached roughly $200 million in ARR. Sierra also says more than 40% of the Fortune 50 now use its platform and that its agents have handled billions of interactions.

Its valuation reflects those numbers. Sierra raised $950 million at a post-money valuation of roughly $15.8 billion, after being valued at $10 billion only months earlier and $4.5 billion before that. Even against Sacra's $200 million revenue estimate, investors are valuing Sierra at close to 80 times annualized revenue. Expectations could hardly be higher.

What strengthens Sierra's case today is what has happened after the fundraising. Over the past few weeks, Sierra has released tools for agents that maintain context over days or weeks, navigate external IVR phone systems, run outbound follow-ups, connect banking information through Plaid and manage staged releases with software-like governance. It also introduced Voice Personas and a Context Engine designed to carry information from one customer interaction into the next.

Sierra is expanding from customer-support automation toward something closer to an operating layer for customer relationships. Decagon is moving in a similar direction, but Sierra currently has the bigger business, deeper Fortune 50 penetration and the strongest financing position.

Is ElevenLabs bigger than Sierra in conversational AI?

ElevenLabs is larger than Sierra by reported recurring revenue, although a straight comparison exaggerates ElevenLabs' position because its business extends far beyond conversational agents.

ElevenLabs ended 2025 at roughly $350 million in ARR and said it had passed $500 million during the first four months of 2026. That is about $150 million of additional recurring revenue, or roughly 43% growth, in four months. No pure-play enterprise conversational-AI startup in our research publicly reports comparable revenue.

The mix matters. ElevenLabs also sells speech synthesis, dubbing, transcription, creative audio tools and media products. We cannot treat all $500 million as conversational-agent revenue. The company itself says enterprise voice-agent deployments are helping drive growth, but it does not disclose how much ARR comes specifically from ElevenAgents.

The direction of the product is clearer than it was a year ago. ElevenLabs recently expanded ElevenAgents into omnichannel use cases, published a healthcare appointment-scheduling workflow and highlighted enterprise deployments with companies such as Deutsche Telekom. ElevenAgents Spotlight now reviews conversations in real time so teams can measure resolution, conversion and customer satisfaction and then improve the agent.

So we rank ElevenLabs second overall. It has the biggest disclosed business in the group and owns unusually strong voice technology, while Sierra still has the cleaner claim to leadership in enterprise conversational agents themselves.

If you want more recent data on this point, please see our latest conversational AI market report.

Can Decagon realistically catch Sierra?

Decagon is the only pure-play startup that currently looks capable of closing Sierra's lead quickly.

The commercial trajectory is unusually fast. Decagon raised $250 million at a $4.5 billion valuation after being valued at $1.5 billion less than a year earlier. The company said it added more than 100 global enterprise customers during 2025, including Avis Budget Group, Block and Deutsche Telekom. Sacra estimates annualized revenue reached roughly $100 million in July, up from about $44 million at the end of 2025.

If that estimate is directionally right, Decagon more than doubled its revenue run rate in roughly seven months. Sierra remains larger, but Decagon has reached nine-figure annualized revenue only a few years after being founded.

The recent product activity is useful too. Decagon has added Browser Actions for agents that need to work through web software, Assist for human representatives dealing with escalations, automated optimization tools and its own work on low-latency text-to-speech inference. That combination suggests Decagon is building wider across the customer-service workflow instead of remaining a thin orchestration layer around external models.

We would still put Sierra comfortably ahead today. Decagon's case rests on speed: if a company already around $100 million in annualized revenue can keep anything close to its recent growth rate, the distance between the two could shrink much faster than their valuations imply.

Is Parloa really in the same league as Sierra and Decagon?

Parloa belongs in the top tier now, especially when we look at global enterprise distribution rather than U.S. valuation alone.

Parloa passed $50 million in ARR and reported 150% net revenue retention, meaning its existing customer base was spending substantially more from one year to the next. It then raised $350 million at a $3 billion valuation, only months after reaching a $1 billion valuation. The company says its technology has powered more than one billion interactions across more than 100 countries and 140 languages.

The more interesting part is how Parloa is distributing the product. Within a short period it deepened its relationship with SAP, integrated with Five9 and partnered with Alvaria for compliant outbound customer engagement. Those companies already sit inside large enterprise contact centers, so Parloa can reach customers without convincing them to throw away the rest of their stack.

That strategy could be particularly effective in Europe and regulated industries, where replacing an entire contact-center architecture is difficult. It also gives Parloa a different path from Sierra's more vertically integrated approach.

Parloa still has a smaller business than Sierra and Decagon based on the revenue evidence available to us. Its 150% net retention and growing distribution network make fourth place much less debatable than it would have been a year ago.

Is Retell AI the breakout voice startup people are missing?

Retell AI is the most capital-efficient breakout company in this ranking, and the numbers are extreme enough to deserve more attention than its funding profile suggests.

Retell disclosed a $4.6 million seed round and later said it had reached $40 million in annualized revenue with a team of 25. The company subsequently reported reaching $50 million in ARR during 2025 and powering more than 50 million real-time AI phone calls per month. Sacra estimates the annualized revenue figure has since reached roughly $60 million.

Using the company's $50 million ARR disclosure, Retell reached more than ten dollars of recurring revenue run rate for every dollar of disclosed venture funding. That ratio does not mean Retell generated ten dollars of cash for every venture dollar; ARR and capital raised measure completely different things. It does show how little outside capital the company needed to reach meaningful scale.

Retell is also moving deeper into the reliability problem. Its vCX-Hard benchmark tests models against difficult moments drawn from real contact-center calls, where even the best model reportedly clears only about 88%. Conductor, another newer product, is designed to build voice agents with more awareness of operational constraints.

We would rather own that combination of revenue, call volume and production data than a much larger funding announcement with very little usage evidence. Retell therefore earns fifth place in our ranking despite having far less capital and publicity than several companies below it.

If you want more recent data on this point, please see our latest conversational AI market report.

Is Vapi becoming the default infrastructure for AI phone agents?

Vapi currently has the best evidence of becoming a horizontal developer platform for AI phone agents.

The strongest proof comes from Amazon Ring. According to TechCrunch, Ring evaluated more than 40 AI voice vendors before choosing Vapi, and it now routes 100% of inbound calls through the platform. A Ring executive said customer-satisfaction scores improved after deployment.

Vapi says it has processed more than one billion calls, currently handles between one million and five million calls per day and has been used by more than one million developers. TechCrunch reported, citing an investor source, that Vapi is running at a "healthy" eight-figure ARR level. The company raised $50 million at a valuation of around $500 million, bringing total funding to $72 million.

The newer products show where Vapi wants to build an advantage. Model Intelligence recommends models using data from production calls, Monitoring helps teams inspect live systems and Simulations tests agents against different caller behaviors before deployment. These products become more useful as Vapi sees more traffic.

The Ring contract gives us unusually clean evidence because there was an actual competitive selection process involving dozens of vendors. Vapi won a production workload where failure would be obvious to millions of customers. For infrastructure, that is much stronger evidence than another polished voice demo.

Is PolyAI being overtaken, or is it quietly getting stronger?

PolyAI is getting stronger commercially even while newer conversational AI startups attract more attention.

PolyAI disclosed more than 100 enterprise customers and over 2,000 live deployments when it raised $86 million late last year. By May, the company was saying it served more than 200 enterprise customers across 25 countries. The exact starting number was "100+," so we cannot claim the customer base literally doubled, but the move past 200 still shows substantial expansion in a short period.

The deployments are also unusually concrete. PolyAI works with FedEx, Marriott, Caesars, PG&E and UniCredit. Fogo de Chão now uses a PolyAI agent across all 88 U.S. restaurants; the company says the system has a 95% guest-satisfaction score and an 88% booking-completion rate. A release published last week added a direct Epic integration being used across more than 1,100 PDS Health dental and medical practices.

That healthcare deployment also shows how far the product has moved. PolyAI now says its dialog agents can operate across more than 75 languages for appointment scheduling, refill requests and other patient interactions.

PolyAI's $750 million valuation looks modest beside Sierra, Decagon and Parloa. Commercially, though, more than 200 enterprise customers and thousands of deployments keep it firmly in the top group. We place PolyAI sixth because Retell's growth efficiency is harder to ignore, not because PolyAI has lost relevance.

Is Bland AI growing as fast as the hype suggests?

Bland AI is already a real enterprise voice company, although its public call-volume data supports a scale story more clearly than an acceleration story.

Fortune reported that Bland has more than 250 enterprise customers, handles over 3.5 million calls per week and raised another $50 million, taking total funding above $100 million. Customers include Samsara, Kin Insurance and CNO Financial Group. Bland also says it owns more of the technical stack than many rivals, including proprietary voice, speech-recognition and language models.

The volume is clearly large. A bit of caution here: Fortune also reported that Bland processed more than 175 million AI calls last year. A current pace of 3.5 million calls per week annualizes to roughly 182 million calls. Because both figures are minimums rather than exact totals, we cannot derive a precise growth rate, but those disclosures alone do not prove that call volume is exploding right now.

Bland has passed the stage where we need to ask whether anyone uses the product. Hundreds of enterprises and millions of weekly calls answer that question. What we still lack is the same transparent revenue trajectory we have for ElevenLabs, Retell, Parloa or Sierra.

Bland belongs in the top ten. We would need clearer revenue growth or a sharper jump in usage before moving it into the top five.

Has Ada actually made the jump from chatbots to AI agents?

Ada has successfully crossed from the old chatbot era into agentic customer service, and its latest growth figures are stronger than its market visibility suggests.

Ada reported more than 100% year-over-year revenue growth, 108% growth in agentic-AI ARR and 146% net revenue retention. Voice ARR grew twelvefold. The company says it now has more than 550 enterprise AI agents deployed globally.

Those figures make Ada one of the better examples of an earlier conversational-AI company adapting rather than getting displaced. Ada spent years in chatbot automation, then moved its platform toward agents that reason across workflows and channels. Its newer Reasoning Engine applies the same instructions, policies and context across voice, messaging, email and other customer interactions.

The historical usage numbers are huge but need context. Ada says it has powered more than 6.4 billion interactions and nearly one billion conversations since 2016. Most of that cannot be treated as evidence of current agentic-AI adoption. We give much more weight to the 108% growth in agentic ARR, 146% retention and twelvefold increase in voice ARR because those figures describe what is happening now.

Ada consequently stays in our top ten. It lacks the recent valuation momentum of Sierra or Parloa, but commercially the company looks much healthier than the "legacy chatbot vendor" label would suggest.

Can Salesforce and Fin squeeze these conversational AI startups out?

Salesforce and Fin are the biggest distribution threat to conversational AI startups, yet independent vendors are still winning major enterprise workloads.

Salesforce already has an installed base that no startup can reproduce. Agentforce reached roughly $1.2 billion in ARR in Salesforce's April quarter, although that number covers a broader agent platform rather than conversational customer service alone. Salesforce can sell agents alongside CRM, Service Cloud, Data Cloud and the rest of its enterprise stack.

Fin is a more direct comparison. Intercom says more than 7,000 teams use Fin, the product is approaching $100 million in recurring revenue and it resolves almost two million customer issues per week. Intercom has also trained its own customer-service model, Apex, rather than depending entirely on frontier models from other labs.

One number deserves extra care. Fin reports an average resolution rate of 76%, but Intercom changed the metric so conversations where Fin never had the opportunity to answer are excluded from the resolution-rate denominator. Intercom says the underlying automation rate did not change. We therefore would not compare that 76% directly with resolution figures from Sierra, PolyAI or Ada unless the measurement definitions were identical.

Acquisitions create another threat. NiCE paid approximately $955 million for Cognigy and can now distribute Cognigy's technology into an existing network of more than 25,000 customers. That is a much faster route to market than building a new AI platform internally.

Still, startups keep winning major accounts. Ring chose Vapi after evaluating more than 40 vendors. Sierra says it has penetrated more than 40% of the Fortune 50. Decagon added more than 100 enterprise customers in one year. Enterprise buyers are clearly willing to go outside their incumbent stack when the specialist product works better.

If you want more recent data on this point, please see our latest conversational AI market report.

What can conversational AI startups build that competitors cannot easily copy?

The hardest conversational AI businesses to copy are increasingly the ones that own either the business outcome or the production-data loop around millions of conversations.

At the application layer, Sierra and Decagon are embedding themselves into the actual work. When an agent verifies a customer, checks account data, makes a change in a billing system, completes a return and follows up later, replacing the vendor means rebuilding much more than a chat interface. Outcome-based pricing pushes the vendor even closer to the operation itself.

Voice infrastructure creates a different kind of depth. Vapi now recommends models using production data and tests agents through simulations. Retell benchmarks models against difficult moments extracted from real customer calls. ElevenLabs' Spotlight reviews live voice and chat conversations to identify where agents fail and what should change.

The pattern is pretty clear across those three companies: conversation data is turning into evaluation infrastructure. A company that handles millions of calls learns where models break, which configurations work for different tasks, how callers interrupt, which accents cause errors, when tools fail and how latency changes outcomes. Competitors can access the same GPT or Claude model; they cannot instantly recreate the same production history.

LiveKit represents another defensible layer. Once developers build their real-time audio, media routing and agent architecture around infrastructure that already powers systems such as ChatGPT Voice Mode, switching has an operational cost even if the underlying AI models remain interchangeable.

Thin wrappers are becoming harder to defend. The stronger positions today sit either close to the customer outcome or deep enough in the infrastructure that removing the vendor becomes painful.

Is conversational AI funding getting overheated?

Yes, parts of conversational AI look overheated, especially where valuations have moved far ahead of already-fast revenue growth.

Take the latest large rounds across eight companies we track closely: Sierra raised $950 million, ElevenLabs $500 million, Parloa $350 million, Decagon $250 million, LiveKit $100 million, PolyAI $86 million, Vapi $50 million and Bland $50 million. Together those rounds add up to about $2.34 billion.

The more unusual pattern is the speed of the repricing. Sierra moved from a $4.5 billion valuation to $10 billion and then roughly $15.8 billion. Decagon jumped from $1.5 billion to $4.5 billion in less than a year. Parloa moved from $1 billion to $3 billion within months. ElevenLabs went from $3.3 billion to $11 billion in about a year.

Revenue has also risen very quickly, so calling the entire market a bubble would be lazy. The multiples show how differently investors are pricing the companies. Using reported revenue where available and clearly identified estimates elsewhere, ElevenLabs is around 22 times ARR, Decagon roughly 45 times, Sierra close to 80 times, while Parloa's valuation is below 60 times its disclosed ARR floor.

Those are software multiples with very little room for ordinary execution. The market can justify them only if conversational AI keeps taking a much larger share of customer-service and communications spending.

Company	Valuation	Current revenue evidence	Rough valuation / ARR
ElevenLabs	$11B	>$500M reported ARR	~22x
Decagon	$4.5B	~$100M estimated annualized revenue	~45x
Parloa	$3B	>$50M reported ARR	<60x using $50M floor
Sierra	~$15.8B	~$200M estimated ARR	~79x

If you want more recent data on this point, please see our latest conversational AI market report.

Which conversational AI startups are actually on top right now?

Sierra is our number-one conversational AI startup today, followed by ElevenLabs, Decagon and Parloa.

Sierra takes first place because we see the strongest combination of enterprise penetration, revenue growth, product depth and strategic ambition in a company focused directly on conversational agents. Its recent push into long-running agents, voice, external phone systems and customer context strengthens that position.

ElevenLabs ranks second because its commercial scale is impossible to ignore. More than $500 million in ARR makes it the largest company here by disclosed recurring revenue, but its broader audio business prevents us from treating the full figure as conversational-agent revenue.

Decagon is third and probably has the best chance of moving higher quickly. Its estimated revenue more than doubled in roughly seven months, its customer list is expanding and its product is spreading across automated agents, human-agent assistance, browser actions and voice technology.

Parloa deserves fourth. The combination of more than $50 million in ARR, 150% net revenue retention and distribution through SAP, Five9 and Alvaria gives it a serious global enterprise position.

Retell takes fifth because we think the market underweights its capital efficiency. PolyAI follows with one of the deepest production footprints in enterprise voice. Vapi ranks seventh on the strength of its developer adoption, enormous call volume and the Ring deployment. Ada comes next after a convincing transition into agentic AI. LiveKit makes the list because real-time conversational infrastructure is becoming strategically important far beyond customer service. Bland completes the top ten: large enough to matter today, with enough unanswered questions around revenue growth to keep it below the leaders.

The ranking will probably keep moving quickly. The gap between Sierra and Decagon could narrow, Retell could become much harder to overlook if its revenue keeps compounding, and Vapi could become more valuable as thousands of companies build on top of it. For now, Sierra has the cleanest claim to overall leadership.

Rank	Startup	Why we rank it here
1	Sierra	Strongest overall enterprise conversational-AI position
2	ElevenLabs	Largest disclosed revenue base and major voice advantage
3	Decagon	Fastest credible pure-play challenger to Sierra
4	Parloa	Strong global enterprise growth and distribution
5	Retell AI	Exceptional revenue and usage relative to capital raised
6	PolyAI	More than 200 enterprise customers and deep voice deployments
7	Vapi	Leading horizontal infrastructure contender for AI phone agents
8	Ada	Strong agentic-AI growth from an established enterprise base
9	LiveKit	Critical real-time infrastructure underneath major AI products
10	Bland AI	Proven enterprise call scale, with less financial transparency

Who is the author of this content?

NEW MARKET PITCH TEAM

We track new markets so founders and investors can move faster

We build living "market pitch" documents for emerging markets: AI, synthetic biology, new proteins, and more. Instead of outdated PDFs or hallucinated LLM answers, our clients get a clean, visual, always-updated view of what's really happening: key players, deals, regulations, and signals that matter. Learn more about us.