Which AI agent businesses make over $10K/month now?

Is $10K/month still a meaningful threshold for an AI agent business?

For AI agent businesses today, $10K a month is already a pretty low bar: several independent companies are doing more than $10 million a month, and the largest is approaching $75 million.

Cognition gives us the most extreme comparison. Earlier this year, the company said its annualized revenue run rate had reached $492 million. Its latest financing disclosure puts that figure at roughly $900 million. Annualized, $10,000 per month is $120,000 a year, so Cognition is running at about 7,500 times our qualification threshold.

The same gap appears elsewhere. Sierra has disclosed more than $150 million ARR. Legora crossed $100 million ARR. Harvey said recently that it added more than $100 million of ARR in a single quarter. Retell AI is estimated at around $80 million of annualized revenue. Unify finished last year at eight figures of ARR.

So $10K still tells us whether people are willing to pay for an agent, especially when we look at smaller companies. For the market leaders, it has stopped being a useful measure of scale. The more interesting lines are closer to $1 million, $10 million and $100 million ARR.

Which AI agent businesses clearly make more than $10K/month today?

Cognition, Sierra, Harvey, Legora, Retell AI, Unify, Vapi, Bland AI and several other agent businesses clear $10K per month comfortably, with the strongest cases clearing it by hundreds or thousands of times.

We need one boundary before comparing them. We count a business when agents are central to what customers buy: software that can take a task, work through multiple steps and use tools or data with some autonomy. That includes Devin writing software, Sierra resolving customer problems, Harvey running legal workflows and Retell powering phone agents. We would not count an ordinary SaaS company simply because it added an AI chatbot.

The cleanest numbers deserve more weight. Sierra's ARR disclosures come directly from the company. Legora's move beyond $100 million ARR was reported alongside its financing. Unify publicly said it reached eight figures. Cognition directly reported its annualized revenue run rate. Retell's number, by contrast, comes from Sacra's model, so we should treat $80 million as an estimate rather than equivalent evidence.

Even after applying that filter, the list is much longer than a few famous outliers.

AI agent business	Latest useful revenue evidence	Rough monthly equivalent	Confidence
Cognition	~$900M annualized revenue run rate	~$75M	High on figure, lower on ARR comparability
Harvey	Added $100M+ ARR in one recent quarter	Well above $8M	High
Sierra	$150M+ disclosed ARR	$12.5M+	High
Legora	$100M+ ARR	$8.3M+	High
Retell AI	~$80M annualized revenue estimate	~$6.7M	Medium
Unify	Eight-figure ARR	$833K+	High
Vapi	Eight-figure ARR reported around financing	$833K+	Medium-high
Bland AI	Reached $2M ARR very early	~$167K at that milestone	High for milestone
Gumloop	~$1.1M estimated annual revenue	~$92K	Low-medium

Can we trust the huge ARR numbers coming from AI agent companies?

We can trust that several AI agent businesses are genuinely large, but we should not treat every “ARR,” “run rate” and “annualized revenue” number as the same thing.

Revenue labels have become sloppy across AI. Investors interviewed by Business Insider have complained that startups increasingly mix recurring subscriptions, consumption revenue, annualized monthly usage and contracted business under similar labels. One exceptional month multiplied by 12 can tell us something useful about scale, but it is not the same as twelve-month subscriptions already under contract.

Cognition is a useful example. The company's latest figure is roughly $900 million of annualized run-rate revenue. Earlier, it reported $492 million on the same basis. That growth is extraordinary, but run rate usually means taking a recent revenue period and extrapolating it. Sierra's reported ARR therefore gives us somewhat different information from Cognition's run rate.

Retell requires another level of caution because Sacra estimates its annualized revenue rather than Retell publicly reporting the number itself. Gumloop's $1.1 million figure is also an outside estimate.

None of that threatens the basic conclusion. A business reported at $10 million, $80 million or $900 million annualized revenue is not going to fall below our $120,000 annual qualification threshold because of a technical definition. It does matter when we try to decide who is actually largest.

For that reason, we give direct company disclosures the most weight, followed by financial reporting tied to fundraising or management comments, and then outside revenue estimates.

Is Cognition now the biggest independent AI agent business?

Cognition appears to be the largest independent agent-focused company by current reported revenue, with its annualized run rate jumping from $492 million to roughly $900 million in only a few months.

The speed is more interesting than the headline alone. Cognition originally proved that Devin itself could monetize: the company said Devin went from $1 million ARR to $73 million ARR in roughly nine months. The later Windsurf transaction then created a much broader coding business.

By May, Cognition reported a $492 million annualized revenue run rate and said enterprise usage of Devin had been growing 50% month over month for six consecutive months. Its latest financing announcement pushed the run rate to about $900 million. Over that short interval, the reported number increased by roughly $408 million, or 83%.

There is one reason not to call $900 million “Devin revenue.” Cognition now owns Windsurf, so part of the combined business comes from an AI coding environment rather than autonomous-agent usage alone. The pre-Windsurf Devin figure of $73 million ARR is cleaner evidence for the commercial strength of the agent itself.

Even with that qualification, coding has produced something unusually large. Devin reached tens of millions in recurring revenue before the combined company approached the billion-dollar run-rate range.

Are customer-service AI agents the clearest commercial winners?

Customer-service agents are currently one of the strongest AI agent businesses we can find because Sierra, Decagon and Fin have all turned automated support work into large revenue streams.

Sierra offers the cleanest evidence. It went from four initial design partners to more than $150 million ARR in roughly two years and says more than 40% of the Fortune 50 now use its technology. Its agents handle jobs ranging from mortgage refinancing to insurance claims and retail returns.

Decagon reached eight-figure ARR earlier in its growth and has since expanded across large consumer businesses. Outside estimates now put its annualized revenue much higher, although we would give those estimates less weight than the company's disclosed eight-figure milestone.

Fin gives us a useful test because it grew inside Intercom rather than as an independent startup. Stripe's customer case study says Fin created an eight-figure revenue line in less than a year while resolving more than one million support conversations per week. Intercom has since gone much further, renaming the company around Fin and making the agent its central product.

Three separate businesses have therefore found substantial demand around essentially the same job: handling customer conversations and resolving problems without waiting for a human agent.

The economics are easy for buyers to understand. Companies already know what their contact center costs, how many conversations arrive, how long agents spend on them and how much outsourcing costs. An AI agent can be tested against an existing budget rather than requiring management to invent a new one.

Are legal AI agents already becoming $100M businesses?

Legal AI agents have clearly crossed the $100 million barrier, with Harvey and Legora independently reaching nine-figure recurring-revenue territory.

Legora passed $100 million ARR roughly 18 months after launching its product commercially. TechCrunch reported the milestone alongside the company's financing this year, giving us a relatively clean and recent revenue anchor.

Harvey has moved even faster at larger scale. The company said recently that it added more than $100 million of ARR during a single quarter. Harvey also says 80% of the Am Law 100 now use its platform, and its newest financing valued the company at $15.5 billion.

Usage inside law firms is becoming deeper as well. Harvey says customers run more than 25,000 custom agents. At one Australian law firm that tested Harvey across 13 practice groups, 70% of participating lawyers were using it daily and 88% came back week after week before the firm moved to an enterprise-wide rollout.

That helps explain why legal AI has monetized so quickly. Lawyers are expensive, document-heavy work repeats constantly, and firms can review an agent's output before anything reaches a client. Saving even a small share of lawyer hours can support a large software bill.

The existence of two independent $100 million-plus companies makes legal AI harder to dismiss as one unusually successful startup. We are looking at a category that has already produced repeatable enterprise spending.

Are voice AI agents real businesses or still mostly impressive demos?

Voice AI agents are already real businesses: Retell AI is estimated around $80 million of annualized revenue, Vapi has reached eight figures, and Bland AI crossed $2 million ARR very early in its life.

Retell gives us the clearest growth curve. Sacra estimates that the company finished 2024 around $5.4 million, reached roughly $45 million annualized revenue by the end of the following year, moved to $60 million in April and then approximately $80 million recently. The latest estimate is almost 15 times the 2024 level.

Vapi gives us a second large example. Around its latest funding, the company was described as an eight-figure ARR business and said its platform had already processed more than one billion calls.

Bland shows how quickly the economics can work at a smaller scale. The company says it reached $2 million ARR just four months after getting started. It now reports more than 3.5 million conversations a week.

Those companies sell slightly different layers of the voice stack, yet customers are paying for the same broad capability: letting software conduct phone conversations that previously required people.

Voice also benefits from falling infrastructure costs. Speech recognition, language-model inference and synthetic voices have become cheaper and faster at the same time. A phone agent that would once have been too slow, expensive or awkward to deploy can now handle high-volume production traffic.

Voice-agent business	Revenue evidence	Current scale clue
Retell AI	~$80M annualized revenue estimate	Up from ~$5.4M in 2024
Vapi	Eight-figure ARR	1B+ calls processed
Bland AI	Hit $2M ARR within months of launch	3.5M+ conversations per week

Can a general-purpose AI agent make serious money?

General-purpose AI agents can become large businesses, but the evidence is thinner than it is for legal, support, coding and voice agents.

Genspark is the obvious counterexample to anyone who thinks an agent has to own one narrow workflow. The company has reported a remarkably fast revenue climb, moving through successive ARR milestones after monetization rather than relying on a single isolated spike. Recent reports have placed it well into nine-figure annualized revenue territory.

That is commercially important because Genspark's proposition is much broader than “handle customer support” or “review this contract.” It attempts to execute research, travel, communication and other multi-step tasks from one general interface.

The problem is replication. We can point to several large customer-service agent businesses and two legal AI companies above $100 million. There are multiple meaningful voice-agent companies too. We have fewer independent general agents with similarly transparent revenue.

So Genspark proves that a broad agent can monetize at huge scale. It has not yet proved that general-purpose agents are as repeatably good a business category as agents connected to a specific expensive job.

Are AI sales agents actually working, or mostly automating spam?

AI sales agents are making real money, but sales is a shakier automation target than customer support, legal work or coding.

Unify is the strongest revenue example we found. The company says it reached eight figures of ARR last year while growing around 15% month over month. Even the bottom of “eight figures” means more than $833,000 a month, far above our $10K threshold.

The product has also changed in an interesting way. Unify originally leaned heavily into automated outbound. Its more recent positioning brings humans back into the loop, with agents researching accounts, assembling data and handling repetitive work while sales representatives take over where judgment and persuasion matter.

Unify itself has been unusually explicit about why. The company says fully automated AI SDRs still struggle to break into complex enterprise accounts because selling requires empathy, context and creativity. Its own new-business representatives work alongside agents rather than disappearing.

That tells us something useful about the category. Sales agents can automate research, prospecting, personalization and workflow execution very well. Closing a complicated deal remains much harder.

The businesses can still become large because sales teams spend heavily on those repetitive steps. We simply have stronger evidence today for agents replacing complete units of work in support or voice than for agents replacing an entire enterprise salesperson.

What do the AI agent businesses making the most money have in common?

The highest-revenue AI agent businesses today tend to work on expensive, frequent and easy-to-measure jobs such as writing code, resolving support cases, reviewing legal material or handling calls.

Look at how different Cognition, Sierra, Harvey and Retell appear on the surface. One serves engineers, another customer-service teams, another lawyers and another voice developers. Their commercial setup is surprisingly similar.

All four enter markets where companies already spend a lot of money on people. All four handle work that happens repeatedly rather than once a quarter. And in each case, customers can inspect the result: code either works, a customer problem gets resolved, a legal document gets reviewed, or a call reaches an outcome.

The pattern gets weaker when we move into vague productivity. “Make employees more productive” can be valuable, but measuring the gain is difficult. “Resolve 30,000 customer conversations” is much easier to price and defend to a CFO.

Gusto's use of Gumloop gives us a good smaller-scale example. Gusto says agents built with Gumloop generated more than $1.5 million of additional ARR in three months, cut preparation time on one workflow from 30–45 minutes to under five minutes, and produced a 31% higher win rate on deals using those briefings. Those outcomes are much more concrete than simply saying employees “saved time.”

The agent businesses growing fastest sit close to a measurable unit of work.

Is outcome-based pricing helping AI agent companies grow faster?

Outcome pricing is giving some AI agent businesses access to a much larger revenue pool because they can charge for work performed rather than only for employees using software.

Traditional SaaS usually grows with seats. A customer has 500 employees, perhaps 300 of them need the application, and the vendor charges per user each month.

Agents break that relationship. A customer-service company might employ 200 support people yet process millions of conversations. If software can resolve those conversations itself, charging per resolution opens a revenue base that can keep growing even when the customer does not hire more staff.

Sierra explicitly markets outcome-based pricing: customers pay for results rather than tokens. Fin charges around successful customer resolutions. Decagon also supports pricing connected to conversations and resolutions.

We can see why vendors like this structure. When an agent does twice as much useful work, its bill can increase without the customer buying twice as many seats.

Customers can understand what they are paying for too. The conversation moves from “Should we buy another AI subscription?” toward “Is paying this amount to resolve a support case cheaper than our existing method?”

That is a much easier commercial argument.

Do AI agent companies need the best underlying model to win?

The businesses making serious agent revenue today show that owning the best foundation model is far from enough; integration, workflow design, data access and reliability are increasingly where customers make their buying decision.

Decagon is one of the clearest examples. The company has said roughly 90% of its workloads run on open-source models. Its customer-service product therefore cannot depend entirely on exclusive access to one frontier model.

Harvey also works across model providers while building legal-specific agents, integrations and data infrastructure. Its recent partnerships with Docusign and Everlaw push Harvey deeper into contracts and litigation evidence, making the product more useful because of what it can access and do inside a legal workflow.

The same idea shows up at Gumloop. Its latest pricing model passes model tokens and compute through at cost and adds an orchestration fee. Customers can also bring their own API keys from OpenAI, Anthropic, Google, xAI, DeepSeek and others. Gumloop is effectively saying that the model can change while the workflow layer remains valuable.

This is one reason the “AI wrapper” label has become less useful for the companies reaching scale. Customers care about whether an agent can safely pull the right company data, take an action, survive edge cases and complete the workflow.

A thin interface around one API remains easy to copy. A deeply installed agent that touches company data, permissions and workflows is much harder to swap out.

Could cheaper AI models hurt agent businesses?

Cheaper AI models should help the strongest agent businesses because their cost of completing work can fall faster than the value customers place on that work.

Voice AI shows how this can happen. Retell became commercially viable partly because speech recognition, language models and synthetic voices became fast and cheap enough to combine into a usable phone experience. Lower model costs widened the set of calls worth automating.

Gumloop provides an even more direct current example. The company recently changed its pricing so model tokens and compute are passed through at cost, while Gumloop charges an 8% orchestration fee. That makes the underlying AI infrastructure almost a commodity input while Gumloop earns money for coordinating the workflow.

There is still a real risk for simple agent products. If two companies perform essentially the same task and one relies on expensive inference, falling model prices can quickly invite cheaper competitors.

Companies such as Harvey, Sierra and Cognition have more room to defend themselves because customers are buying completed work inside complicated enterprise systems. A cheaper model can improve their gross margin or let their agents do more work for the same price.

So model deflation currently looks more favorable for mature agent platforms than destructive. The pressure will fall hardest on products whose only real advantage was access to an expensive model that everyone can now obtain cheaply.

Which AI agent businesses really make over $10K/month now?

Yes, many AI agent businesses make over $10K per month now, and the strongest evidence suggests that dozens more probably clear the threshold even though private companies rarely disclose revenue at that size.

Cognition is the largest example we can defend from recent reporting, with roughly $900 million of annualized run-rate revenue across Devin and Windsurf. Sierra has publicly passed $150 million ARR. Legora crossed $100 million. Harvey recently said it added more than $100 million ARR in one quarter alone. Retell is estimated around $80 million annualized revenue. Unify and Vapi are already eight-figure ARR businesses. Bland crossed $2 million ARR early enough that $10K MRR was never much of a hurdle.

The freshest evidence also makes the category pattern clearer. Cognition's run rate has nearly doubled in a few months. Harvey has just raised again at a $15.5 billion valuation while reporting that 80% of the Am Law 100 use its product. Retell's estimated annualized revenue has continued climbing from roughly $45 million at the end of last year to around $80 million. These are businesses still expanding rather than old revenue anecdotes being recycled.

The most reliable winners cluster around coding, customer service, legal work and voice. Each category gives the agent a frequent job, a big existing labor budget and an output that customers can check.

Sales agents and broad general-purpose agents also make money, although the evidence is more uneven. Genspark shows that a broad agent can reach very large revenue. Unify shows that AI-heavy sales software can reach eight figures, while its recent shift toward human-plus-agent selling also exposes the limits of full autonomy.

At the smaller end, Gumloop is estimated around $1.1 million in annual revenue, which would still mean roughly $92,000 per month. We have much less confidence in that estimate than in Sierra's direct ARR disclosure, but it illustrates how far below the famous unicorns the commercial market extends.

The answer to the original question is stronger than “yes.” $10K a month is already an early traction milestone in AI agents. The companies shaping the market today are fighting over whether they can reach $10 million, $100 million or eventually $1 billion of annual revenue.

AI agent business	What customers pay it to do	Best current revenue evidence	Our read
Cognition	Write and execute software work	~$900M annualized run rate across Cognition/Windsurf	Largest reported independent agent-focused business
Harvey	Legal research, drafting and workflows	Added $100M+ ARR in one quarter	Clearly nine-figure business
Sierra	Resolve customer-service work	$150M+ disclosed ARR	One of the cleanest agent revenue cases
Legora	Legal research and workflows	$100M+ ARR	Clearly nine figures
Retell AI	Power voice agents	~$80M annualized revenue estimate	Very large, but estimated
Unify	Automate sales research and outbound work	Eight-figure ARR	Clearly above threshold
Vapi	Voice-agent infrastructure	Eight-figure ARR reported	Clearly above threshold
Bland AI	Automate phone conversations	$2M ARR reached early	Clearly above threshold
Gumloop	Build workplace agents and automations	~$1.1M annual revenue estimate	Above threshold, lower-confidence figure