Who are the top AI agent startups by revenue today?

Why is it so hard to rank AI agent startups by revenue?

Ranking AI agent startups by revenue today is genuinely messy because the biggest companies increasingly mix autonomous agents with search, assistants, coding environments and traditional software.

Cognition is the clearest example. Devin can independently work through software-engineering tasks, yet Cognition now also owns Windsurf, an AI coding environment with its own subscription revenue. Glean sells enterprise agents alongside search and an AI assistant. Harvey increasingly automates legal workflows, although lawyers also pay for research, drafting and document analysis. Replit and Lovable can build substantial pieces of software from a prompt, even though most people still describe them as AI coding platforms.

The revenue numbers create another problem. Private AI companies variously report ARR, annualized revenue, monthly revenue multiplied by twelve and loosely defined run rates. These figures are especially easy to confuse when a startup is growing 10% or 20% in a month.

We therefore need two views of the market. For the strict ranking, we count companies where autonomous task execution sits at the heart of the product customers buy. We then look separately at larger AI platforms where agents have become a major part of the product.

That gives us a cleaner answer than throwing every company with an “agent” button into one table.

What actually counts as an AI agent startup?

We count Cognition, Sierra, Manus, Decagon and Parloa as clear AI agent startups because customers mainly pay them to delegate work that the software then carries out.

Devin receives engineering tasks and works through them with limited supervision. Sierra, Decagon and Parloa handle customer conversations and can take actions across business systems. Manus receives broader digital assignments and operates through virtual computers to complete multi-step work.

Harvey, Legora and Glean sit one layer further out. Their products now contain serious agentic capabilities, but customers also pay for large amounts of non-agent work. Harvey handles legal research and drafting. Legora combines research, document review and workflow automation. Glean still has a major enterprise-search and assistant business.

Lovable and Replit are harder again. Their products increasingly behave like software-building agents, yet calling every natural-language coding platform an “AI agent startup” would stretch the category until it stopped being useful.

For the main ranking, we keep the definition fairly strict. We bring the broader companies back when their revenue helps explain where the agent market is heading.

Which AI agent startup makes the most revenue today?

Cognition is the clear revenue leader among independent agent-first startups, with an estimated annualized revenue pace of about $900 million.

Sacra's latest estimate puts Cognition at roughly $900 million in annualized revenue, up from the $492 million run rate the company itself disclosed in May 2026. Even if we ignore the newer estimate and use only Cognition's official $492 million figure, the company would still sit far above Sierra at roughly $200 million ARR.

The climb started well before Cognition became this large. Devin went from roughly $1 million ARR in September 2024 to $73 million by June 2025. Cognition then acquired Windsurf, which was contributing about $82 million ARR. The two businesses had less than 5% overlap among enterprise customers, according to Cognition, so the acquisition gave the company a large second distribution channel rather than simply combining two versions of the same customer base.

The combined company then accelerated. Cognition said enterprise ARR rose more than 30% within seven weeks of the Windsurf deal, and enterprise usage of Devin later grew about 50% month over month for six consecutive months before the company disclosed its $492 million run rate.

That history is why we are comfortable calling Cognition the leader even though the newest $900 million figure is an external estimate. There is enough distance between Cognition and the rest of the field that small measurement errors do not change first place.

Independent agent-first startup	Latest useful revenue indicator	Main product	Confidence
Cognition	~$900M annualized revenue estimate	Software-engineering agents + Windsurf	Medium-high
Sierra	~$200M ARR	Customer-service agents	High
Manus	$125M+ revenue run rate	General-purpose AI agent	Medium
Decagon	~$100M annualized revenue estimate	Customer-service agents	Medium
Parloa	$50M+ ARR	Customer-service and voice agents	High

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

How much of Cognition's $900 million revenue really comes from AI agents?

Cognition's roughly $900 million annualized revenue makes it the largest agent-first startup today, although a meaningful share comes from the Windsurf coding platform rather than Devin alone.

Before the acquisition, Devin had reached about $73 million ARR and Windsurf was around $82 million. The deal therefore roughly doubled Cognition's immediately visible revenue base.

What happened afterward is more important. Adding those two disclosed figures gives about $155 million. Cognition subsequently reached a company-reported $492 million annualized run rate, more than three times that simple starting point. Sacra now estimates around $900 million.

Windsurf clearly helped create the lead. It gave Cognition a large installed base of developers who already used AI interactively inside an editor, while Devin offered a second mode in which engineers could delegate longer tasks. According to Cognition, fewer than 5% of enterprise customers overlapped when the businesses combined, which left substantial room to sell each product into the other's accounts.

So today's Cognition is best described as an AI software-engineering company built around both agents and an agentic IDE. Calling the entire $900 million “Devin revenue” would exaggerate Devin's standalone size. But excluding Cognition from the top spot would miss how the company actually sells software now.

Is Sierra the biggest pure enterprise AI agent company?

Sierra is the clearest large-scale pure enterprise AI agent company, at roughly $200 million ARR.

Sierra crossed $100 million ARR seven quarters after commercial launch, then recorded its first $50 million revenue quarter and entered its third year above $150 million ARR. Sacra's latest estimate puts the company around $200 million.

The customer base makes that figure unusually useful. Sierra works with companies including Cigna, DIRECTV, Rivian, SiriusXM, SoFi, Sonos and Wayfair. The company has said roughly half of its customers generate more than $1 billion in annual revenue and about one quarter generate more than $10 billion.

Sierra also ties much of its pricing to actual agent activity. Customers can pay per conversation or successful resolution, so revenue expands when companies hand agents more customer interactions. That relationship is much closer to work performed than a conventional software seat.

At around $200 million ARR, Sierra has moved well beyond pilot-scale spending. Large companies are paying serious money to let AI agents handle production customer interactions, and Sierra remains the strongest standalone proof of that model.

Is Decagon actually catching Sierra in customer-service AI agents?

Decagon is growing fast enough to become a serious challenger, but Sierra still appears to generate about twice as much revenue today.

Sacra estimates Decagon reached roughly $100 million in annualized revenue in July 2026, up from about $44 million at the end of 2025. Sierra is around $200 million ARR on the same broad basis.

Decagon's underlying customer growth supports the direction of that estimate. The company added more than 100 global enterprise customers during 2025, including Avis Budget Group, Deutsche Telekom and Mercado Libre. Its Q3 2025 GAAP revenue and ARR were both more than triple their levels a year earlier.

The two companies now attack almost the same economic problem. Sierra and Decagon charge around conversations, successful resolutions and expanding customer-service volume. Both can therefore grow inside an account as more support traffic shifts from humans to agents.

Sierra still has the larger commercial footprint and a better documented sequence of revenue milestones. Decagon has cut the gap enough that customer-service AI agents now look like a real two-company race rather than a market Sierra has already locked up.

Are customer-service agents the biggest AI agent business right now?

Customer-service agents are the deepest pure-agent market, while coding produces much larger individual revenue leaders.

Sierra at roughly $200 million, Decagon near an estimated $100 million and Parloa above $50 million together represent more than $350 million in annualized revenue. That total comes from only three independent companies and excludes agent products sold by Salesforce, ServiceNow, Intercom and other incumbents.

Customer service works unusually well for agents because the jobs repeat constantly. Companies already have ticket histories, policies, CRM data and knowledge bases that agents can use. They can also measure whether an issue was resolved, transferred to a human or reopened.

Coding has produced fewer clear independent agent companies, but the revenue ceiling has been much higher. Cognition is the obvious example, and the broader coding market makes the gap even clearer: Lovable has passed $500 million in annualized revenue, while Cursor reached roughly $4 billion shortly before its acquisition by SpaceX.

For now, customer service gives us the broadest evidence that specialized enterprise agents can repeatedly generate substantial revenue. Coding shows how enormous one successful agentic workflow can become.

Agent market	Leading independent companies	Approximate current scale	What we see today
Software engineering	Cognition	~$900M annualized	Largest agent-first company
Customer service	Sierra, Decagon, Parloa	>$350M combined	Deepest group of pure-agent startups
General-purpose agents	Manus	$125M+ run rate	Large consumer/prosumer experiment
Sales agents	11x, Artisan and peers	Mostly far below $50M individually	Well behind early hype

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

Is Harvey one of the biggest AI agent startups by revenue?

Harvey is one of the biggest agentic AI companies today at roughly $350 million ARR, but we keep it outside the strict pure-agent ranking.

Harvey's core legal platform does much more than autonomous execution. Lawyers use it for research, document analysis, drafting and knowledge work. At the same time, the product is becoming increasingly agentic: Harvey has said customers were already running more than 25,000 custom agents on the platform.

The revenue trajectory is striking. Harvey crossed roughly $100 million ARR in 2025, reached around $200 million afterward and is now at about $350 million according to recent reporting and Sacra's estimates. The Times reported this week that Harvey has reached $350 million ARR and is used by more than 200,000 lawyers globally.

That makes Harvey larger than Sierra and roughly comparable with the next tier of major AI application companies. It also shows how quickly the distinction between an AI assistant and an AI agent can disappear once customers begin handing the software complete workflows.

If someone asks which independent AI companies are making the most money from agentic work, Harvey clearly belongs near the top. A strict ranking of companies built primarily around autonomous agents should still separate it from Cognition, Sierra and Decagon.

Is Legora catching Harvey in legal AI revenue?

Legora is growing much faster than most software startups, but Harvey still generates more than twice its estimated revenue.

Legora publicly crossed $100 million ARR after less than 18 months on the market. Sacra estimates that the company reached about $150 million ARR by June, compared with roughly $50 million at the end of 2025 and just $3 million at the end of 2024.

That works out to roughly 50-fold growth in about eighteen months from the end-2024 base. The company now serves more than 1,000 customers across around 50 markets, including White & Case, Linklaters and Barclays.

Harvey remains around $350 million ARR and has deeper penetration among the world's largest law firms and corporate legal departments. Legora's growth rate, however, has made this much less comfortable than the raw revenue gap suggests.

The more important pattern is happening inside both products. Legal AI began with research, summarization and drafting. Customers are now building workflows that let software handle longer sequences of legal work. Legal platforms are becoming agent businesses through usage rather than through branding.

Should Glean's $300 million ARR count as AI agent revenue?

Glean is a $300 million ARR enterprise AI company with a serious agent business, but we cannot say how much of that $300 million comes specifically from agents.

Glean announced in May 2026 that ARR had reached $300 million, up from $100 million only fifteen months earlier. It had already doubled from $100 million to $200 million in nine months, so the pace remained unusually fast even as the company became larger.

The product has also changed substantially. Glean started with enterprise search, added an AI assistant and now offers agent building, orchestration and governance. More than 85% of customers use Glean across at least five departments, while its Fortune 500 customer count has nearly doubled year over year.

That broad deployment is exactly why Glean is interesting in an agent discussion. A specialist company such as Sierra begins with one job and builds an agent around it. Glean begins with access to company data, permissions and systems, then lets teams build many agents on top.

Glean still does not disclose agent-specific revenue, so putting all $300 million into the strict leaderboard would create false precision. In the broader market for enterprise agentic software, it is already one of the largest independent companies.

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

How big is the Manus AI agent today?

Manus is still one of the largest general-purpose AI agent businesses we can verify, with more than $100 million ARR and a total annual revenue run rate above $125 million.

The Information reported that Manus crossed $100 million in subscription ARR only eight months after introducing paid plans. Usage-based revenue lifted the company's total annualized pace above $125 million, up from roughly $90 million a few months earlier.

That is a very different business from Sierra or Decagon. Manus asks consumers and professionals to delegate open-ended work such as research, website creation, data tasks and document production. The system has reported processing 147 trillion tokens and creating more than 80 million virtual computers for agent tasks.

Manus then went through a disruptive period around Meta's attempted acquisition, which regulators ultimately blocked before the businesses separated. Because of that upheaval, we would rather keep the last solid $125 million run-rate figure than project its earlier growth forward.

The $125 million figure is already enough to make the point. A general-purpose AI agent can generate nine-figure annualized revenue without being tied to one narrow enterprise workflow. Whether Manus can keep that revenue growing as specialized agents improve is still much less clear.

Why are coding agents making so much more money than AI sales agents?

Coding has become the richest agentic software category because developers can give AI valuable work all day and quickly check whether the result works.

Cognition is now estimated around $900 million in annualized revenue. Lovable, which sits just outside our strict agent definition, passed $500 million in annualized revenue in June and said users were creating about one million new projects per week. Replit had already reached $150 million annualized revenue in 2025 and has since talked publicly about heading toward a billion-dollar run rate, although we do not have a current figure precise enough to rank it.

Cursor shows the extreme version. Forbes reported that the company reached about $4 billion in annualized revenue shortly before SpaceX acquired it. Cursor has since stopped being an independent startup, but the number tells us how much money is flowing into agentic coding.

Sales agents remain tiny by comparison. 11x was reported around the low tens of millions in annualized revenue during its rapid-growth phase and later faced questions around customer retention and some customer claims. Artisan was around $9 million ARR earlier this year. Even allowing for private-company reporting gaps, neither category comes close.

Coding has a built-in feedback loop that most knowledge work lacks. An agent can modify code, run a test, inspect the error and try again. Developers can review the result before it reaches production. Every successful task saves the time of an employee whose labor is already expensive.

The gap is pretty stark. The leading coding businesses are already measured in hundreds of millions or billions of annualized revenue; the best-known autonomous sales startups are generally measured in tens of millions.

Are ARR headlines making AI agent startups look bigger than they really are?

Yes. Today's AI agent revenue rankings usually make companies look larger than their actual trailing revenue because most headline figures are annualized.

A startup generating $10 million this month can announce a $120 million annual revenue run rate even if it generated much less than $120 million during the previous twelve months. In a slow-growing SaaS business the distinction may be manageable. In an AI startup growing 10% or 20% every month, it can be huge.

Cognition's roughly $900 million estimate is an annualized pace, for example. It does not tell us that Cognition collected $900 million over the previous year. Manus explicitly separated $100 million of subscription ARR from more than $125 million of total annualized revenue once usage-based spending was included.

Glean illustrates another wrinkle. The company calls its $300 million figure ARR, although some customers use consumption or hybrid pricing. TechCrunch pointed out in May that a consumption-based component does not behave exactly like classic recurring SaaS revenue.

We still use these figures because private companies rarely publish comparable GAAP revenue. They are measures of current commercial scale, not money already earned.

Revenue metric	What we learn from it	What can mislead us
Trailing revenue	Money already generated	Lags badly during hypergrowth
ARR	Annualized recurring contracts or subscriptions	Assumes recurring revenue stays recurring
Annualized revenue	Latest revenue pace multiplied forward	Can race ahead of actual trailing sales
Revenue run rate	Broad estimate of current annual pace	Definitions vary by company
Revenue target	Management's ambition	Says little about revenue today

Which AI agent startups have the best-looking revenue?

Sierra has one of the cleanest revenue stories among pure AI agent startups because large enterprises repeatedly pay for agents that handle measurable customer interactions.

Sierra's contracts often scale with conversations or successful resolutions. That gives the company a direct path to expansion when a customer sends more support volume through its agents. Its customer list also skews toward major enterprises, which generally means larger deployments and more integration into day-to-day operations.

Cognition has far more scale, but the economics are harder to read from outside. Coding agents can consume large amounts of inference, and Cognition combines Devin usage with Windsurf subscriptions. We know the top line is exploding; we know much less about the gross margin underneath it.

Harvey and Glean have another advantage: customers embed these products across high-value enterprise workflows. Glean says more than 85% of customers already deploy the platform across at least five departments. Harvey is used by more than 200,000 lawyers according to recent reporting. Those deployment patterns should make casual abandonment harder.

We should not pretend we have public-company-quality data. None of these startups gives us the retention cohorts, gross margins and customer-concentration disclosures we would use to compare listed SaaS businesses properly.

Right now, Sierra gives us the easiest revenue to understand. Cognition gives us the biggest revenue number.

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

Are startups like 11x and Artisan anywhere close to the AI agent revenue leaders?

The best-known “AI employee” startups are still nowhere close to Cognition, Sierra or the biggest legal and coding platforms by revenue.

Artisan was around $9 million ARR earlier this year. 11x reached roughly the low tens of millions in annualized revenue during its early surge, followed by reporting about churn and disputes over some customer references. These are meaningful startup revenues, but the gap to Sierra's roughly $200 million and Cognition's estimated $900 million is enormous.

Funding headlines can disguise that difference. A company can raise $50 million or $100 million while generating only a few million dollars of recurring revenue, especially in a category investors expect to become large. Revenue tells us how much customers are already willing to pay.

This also changes how we should think about the “AI employee” story. The broad promise of replacing entire roles generated a lot of attention. The companies making the most money today usually sell something much more specific: resolve customer problems, complete coding tasks, review legal work or operate inside a defined business system.

The revenue hierarchy has become much clearer than the funding hierarchy.

What do today's top AI agent startups have in common?

The highest-revenue AI agent startups cluster around jobs where the work is expensive, the agent has enough context to act, and the result can be checked quickly.

Cognition has codebases, tests, error messages and developer review. Sierra and Decagon can use CRM data, company policies and support histories, then measure whether a customer issue was resolved. Harvey and Legora work with documents and legal knowledge while lawyers remain available to review important outputs. Glean gives agents access to company knowledge, permissions and connected enterprise systems.

These environments give AI repeated chances to produce measurable value. An engineering team can estimate how much developer time an agent saved. A contact center can count completed conversations. A law firm can compare the time needed to research or review documents.

Open-ended sales prospecting is much harder. A prospect can ignore a perfectly written message for dozens of reasons unrelated to agent quality, and a bad outbound campaign can damage a company's reputation or email domain.

The early revenue market is telling us something fairly concrete about autonomy. Agents are already commercially powerful when success can be observed and corrected. Revenue drops sharply once the software enters jobs where outcomes depend heavily on messy external behavior.

Are AI agent revenues concentrated in just a handful of startups?

Yes. AI agent revenue is already heavily concentrated at the top, despite the huge number of companies that now describe themselves as agent startups.

Our five-company strict group reaches roughly $1.4 billion in combined annualized revenue using the latest available figures: Cognition around $900 million, Sierra around $200 million, Manus above $125 million, Decagon around $100 million and Parloa above $50 million.

Cognition alone represents roughly two-thirds of that total. Even allowing for the fact that its figure includes Windsurf and comes from an external estimate, the concentration is striking.

The broader agentic-software group adds several more large companies. Harvey sits around $350 million ARR, Glean at $300 million, Legora around $150 million and Lovable above $500 million. Those figures show that significant agentic revenue exists outside companies explicitly branded around agents.

Meanwhile, dozens of well-funded agent startups remain below $50 million ARR. Some are probably growing extremely quickly, and private data will always leave blind spots. There is still no evidence of a flat market where ten or twenty companies are all within striking distance of the leader.

Today, the financial winners form a surprisingly short list.

So who are the top AI agent startups by revenue today?

Cognition is the clear number-one AI agent startup by revenue today, Sierra is the strongest pure enterprise-agent runner-up, and Manus, Decagon and Parloa make up the next measurable tier.

Cognition's latest annualized revenue estimate is about $900 million, although its last company-disclosed run rate was $492 million. Sierra is around $200 million ARR. Manus has disclosed more than $100 million subscription ARR and over $125 million in total annualized revenue. Decagon is estimated around $100 million. Parloa has passed $50 million ARR.

Once we widen the definition to businesses where agents are a major part of a broader AI product, the leaderboard gets much richer. Harvey is around $350 million ARR, Glean has officially reached $300 million and Legora is estimated near $150 million. Lovable has passed $500 million in annualized revenue, although we still classify it primarily as an AI software-building platform.

Cursor would dwarf most of these numbers after reaching roughly $4 billion in annualized revenue, but SpaceX has now completed its acquisition, so Cursor no longer belongs in a ranking of independent startups.

The most interesting conclusion is where the money has accumulated. Generic “AI employee” startups have attracted plenty of attention, yet the largest revenues sit in software engineering, customer service, legal work and enterprise knowledge. Those jobs give agents clear inputs, valuable tasks and outputs that somebody can verify.

Using a strict definition, Cognition leads by a huge margin. Sierra provides the strongest evidence that a pure enterprise-agent company can build a large recurring-revenue business. And the rise of Harvey, Glean, Legora and Lovable suggests the eventual winners may increasingly look like full software platforms with agents built deeply into them rather than standalone agent products.

Company	Latest useful annualized revenue indicator	How we classify it	Current position
Cognition	~$900M estimated	Agent-first coding company	Clear #1
Lovable	$500M+	Agentic coding platform	Outside strict ranking
Harvey	~$350M ARR	Legal AI platform with agents	Major broader leader
Glean	$300M ARR	Enterprise AI platform with agents	Major broader leader
Sierra	~$200M ARR	Pure enterprise agents	#2 strict ranking
Legora	~$150M estimated ARR	Legal AI platform with agents	Fast-growing broader player
Manus	$125M+ run rate	General-purpose AI agent	#3 strict ranking
Decagon	~$100M estimated	Pure enterprise agents	#4 strict ranking
Parloa	$50M+ ARR	Pure enterprise agents	#5 strict ranking

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

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.