Making Money With AI Agents: Who Is Actually Earning, Doing What, and What It Really Takes

Somebody is getting rich on AI agents, and it is mostly not who the ads say. This is the documented record: the solo founder who sold for $80 million, the agent companies posting nine figure revenue, the freelancers riding an 18,347 percent demand spike, and the far larger pile of projects that quietly died.

Type "make money with AI agents" into a search engine and you will drown in promises. Courses promise passive income. Gurus promise a fully automated business by Friday. Almost none of them show receipts, and receipts are the only interesting thing here, because the real record turns out to be more useful than the fantasy. Real people and real companies are earning serious money with AI agents right now. They are just earning it in specific, copyable ways, and the ways are not the ones being sold in the ads.

So let us do the boring, valuable thing. Let us go through the documented cases, with names, numbers, and sources you can check. Then let us look at the other ledger, the failure statistics, because any honest map of this territory has to show both. Finally, let us turn the cases into paths: what a person reading this could actually sell, to whom, with what tools, and for roughly what it costs to start.

First, the machine: four ways money actually moves

Strip away the branding and there are only four ways anyone earns money from an AI agent. Every case in this article is one of these, and confusing them is how people waste a year.

You sell the agent itself. A customer pays a subscription for software that does a job: answers support tickets, builds apps, qualifies leads. This is the SaaS model with an agent inside, and it is where the biggest numbers in this article live.

You sell the outcome. Instead of renting software, the customer pays per finished unit of work: per resolved conversation, per booked appointment, per completed report. This pricing model barely existed before agents, because software could not finish work. Now it can, sometimes, and the price list is public.

You sell the building. Most companies cannot build agents themselves, so they pay people who can: freelancers, consultants, and agencies. The money here is fees and retainers, and the demand signal is measurable in public marketplace data.

You keep the savings. A company runs agents on its own work and keeps the difference: fewer outsourced contracts, fewer support hours bought, more output from the same team. This is real money, but it is the easiest to exaggerate, because nobody audits a savings claim. Treat every savings figure in this article as company reported, because it is.

Notice what is not on the list: buying an agent, pointing it at the internet, and collecting money while you sleep. That is not a business model. That is a demo. The four models above all share the same hidden ingredient, which the cases will keep pointing at: a human or a company standing behind the agent, responsible for the result.

The solo builder: one person, $80 million, six months

The single most instructive money story of the agent era is not a giant company. It is one man in Tel Aviv.

Maor Shlomo had already built and left a data startup called Explorium. In February 2025, after a year of army reserve duty, he launched Base44, a tool that lets anyone build a working app by describing it to a chatbot. He was, in Inc. magazine's words, a solo founder who "handled every aspect of Base44, from sales to engineering to customer support," waking every two to three hours at night to check the platform was still running. One month after launch, Base44 had generated nearly $1.5 million in revenue from subscriptions priced between $20 and $200 a month. In June 2025, roughly four months after launch, Wix acquired Base44 for $80 million in cash.

The details matter more than the headline. By the sale, Base44 had roughly 250,000 to 300,000 users and about $3.5 million in annual recurring revenue, and it was profitable, making close to $189,000 in its best month, according to Calcalist's interview with Shlomo. He owned 100 percent of the company. He had about eight employees by the end, and roughly $25 million of the purchase price was set aside as a retention pool for them. So "solo" is partly a slogan. But the core fact survives every asterisk: one owner, no venture funding, a team you could fit in a minivan, and an eight figure exit built on a product that charged money from the start.

Why did this work when hundreds of other AI app builders launched in the same months and went nowhere? The coverage keeps landing on the same three things. He sold a concrete outcome, an app without coding, not a technology. He charged from day one, so demand was proven by payments rather than applause. And he iterated against real usage at a speed a larger company could not match. Shlomo is not an anomaly in kind, either. Carta, which manages the ownership records of tens of thousands of startups, reports that the share of new startups founded by a single person rose from 23.7 percent in 2019 to 36.3 percent in the first half of 2025. Agents and AI coding tools are a large part of why: the minimum viable team keeps shrinking.

Watch: Solo founder, $80M exit, 6 months: The Base44 bootstrapped startup success story, from Lenny's Podcast. Shlomo walks through the whole run himself, including the parts the headlines skip, like not writing front end code for three months because the AI wrote it.

The agent companies: nine figures of recurring revenue

One level up from the solo builder sit the companies whose entire product is agents, and the striking thing about 2025 is how fast their revenue became real.

Sierra, founded by former Salesforce co-CEO Bret Taylor and former Google executive Clay Bavor, sells AI agents that handle customer service for large companies: billing disputes, returns, product questions, across chat and voice. Sierra launched in February 2024. On November 21, 2025, it announced $100 million in annual recurring revenue, reached in seven quarters, up from roughly $20 million the year before. Its named customers include SoFi, Ramp, and Brex, and it has pushed hard into regulated industries like finance and health care, where an agent that mishandles a conversation creates legal problems, not just unhappy customers. Sierra's revenue is company reported, as all private company ARR is. But the customer list is checkable, and enterprises do not sign contracts of that size for demos.

The incumbent response shows the same money from the other side. In its results for the quarter ending October 31, 2025, Salesforce reported that Agentforce, its agent platform, had passed $500 million in annual recurring revenue, up 330 percent year over year, with more than 9,500 paid Agentforce deals closed and over 18,500 total deals since launch. Agentforce and its data sibling together reached $1.4 billion in ARR. Whatever discount you apply to platform accounting, a public company reporting half a billion dollars of agent revenue in an earnings release is not describing a pilot program.

Then there is the toolmaker. Lovable, based in Stockholm, sells the thing Base44 sold, an AI app builder, at much larger scale. It passed $100 million in ARR in July 2025, eight months after launch, doubled to $200 million by November, and reached $400 million in February 2026 with 146 full time employees. TechCrunch did the division: $2.77 million of ARR per employee, which beats the figure Gartner predicted a new class of unicorn companies would reach by 2030. Lovable's users build apps; Lovable's business is selling the agent that builds them.

What do Sierra, Salesforce, and Lovable have in common, underneath the different products? None of them sells "AI." Sierra sells resolved customer problems. Salesforce sells finished work inside systems its customers already pay for. Lovable sells finished apps. The agent is the engine. The product is a result somebody already had a budget for. Y Combinator's bet says the same thing at seed scale: 67 of the 144 startups in its Spring 2025 batch, 46 percent, were building AI agents, and the ones getting funded were overwhelmingly narrow and vertical, an agent for one industry's one expensive workflow, not a general purpose robot.

Watch: Building more effective AI agents, from Anthropic. Two of the engineers behind Claude's agent work explain what actually holds up in production, which is the knowledge underneath every company in this section.

Selling finished work by the unit

The most quietly revolutionary money in this space is not a company. It is a price list.

In August 2024, Zendesk became the first major support vendor to charge by the outcome: customers pay only when its AI agent resolves an issue on its own, at published rates of $1.50 per resolution on committed volume and $2.00 pay as you go, with escalations to humans billed at zero. Intercom went further. Its agent, Fin, costs $0.99 per outcome, charged at most once per conversation, with a 50 outcome monthly minimum when it runs on another company's help desk, and failed attempts are free. Salesforce's Agentforce lists $2.00 per conversation. Think about what those numbers mean. A unit of finished customer service work, which used to require a salaried human, now has a wholesale price of about one to two dollars, printed on a web page.

For buyers, that pricing is why agent budgets get approved: the cost is tied to a result, so the return is arithmetic instead of faith. For sellers, it is the template worth copying at every scale. The freelancer who charges per resolved ticket backlog, per qualified lead delivered, per reconciled month of books, is using the same invention as Intercom, at kitchen table scale. The era of charging for software seats is ending wherever software does the work itself. What replaces it is charging for the work.

The buyer's side of this market has its own documented receipts, and the best known is Klarna. The payments company says its AI support agent now does the work of 853 full time agents, up from 700 at the start of 2025, has saved the company $60 million, handles two thirds of all customer inquiries, improved response times by 82 percent, and cut repeat issues by 25 percent. Those are Klarna's own figures, reported on its earnings calls, and they should be read as such. But keep reading, because Klarna is also the honesty section's best witness, and we will meet it again there.

The builders for hire: freelancers and the agency wave

Now for the layer of this economy closest to an individual reader: the people companies hire to build agents for them.

The demand signal is the cleanest public number in this whole article. Fiverr's Spring 2025 Business Trends Index, drawn from tens of millions of searches on the platform, found that searches for AI agent development had risen 18,347 percent in six months. Searches for experts in the workflow tools that agents are assembled from rose with it: Go High Level up 1,489 percent, Make.com up 1,083 percent. Fiverr's own explanation is the important part: most companies, in the words of its head of verticals, do not know exactly what AI agents are or how to use them, and that knowledge gap is what they are paying freelancers to close.

The supply side has faces. Mark Kashef, an AI consultant on Fiverr and founder of Prompt Advisors, told the platform that nearly 30 percent of his Fiverr orders were for AI agent development, with demand shifting from simple chatbots to multi agent systems spanning web, WhatsApp, and voice, and that automation, voice, and integration requests made up 49 percent of his project pipeline. That is one seller's self report, quoted by the platform that profits from the marketplace, so read it as a signal, not a census. The honest limits of the public data are worth stating plainly: no platform publishes what the median agent builder actually earns, and the loudest income claims in this niche come from people selling courses about the niche. A creator's video about his own $45,000 month is marketing material for his program, not an audited result. Both things can be true at once: the demand is real and enormous, and the typical income is unknown and certainly smaller than the screenshots.

The agency version of the same trade packages the building into an offer for one kind of business: a support agent for Shopify stores, an intake agent for law firms, a booking and reminder agent for clinics, sold for a setup fee plus a monthly retainer for maintenance and improvement. The maintenance is not a detail. Agents drift, business rules change, models update underneath them, and the retainer is what turns a one off project into income that repeats. The agencies that survive are the ones that sell that continuing responsibility, not the ones that install a chatbot and vanish.

The other ledger: what the failure record says

Everything above is true. Here is the rest of the truth, in the same font size.

MIT's Project NANDA studied the enterprise side in its 2025 report, The GenAI Divide, reviewing more than 300 public deployments and interviewing dozens of organizations. Despite an estimated $30 to 40 billion of enterprise spending on generative AI, about 95 percent of organizations reported no measurable effect on profit and loss. The funnel underneath is the useful part: about 60 percent of organizations evaluated custom or vendor built AI systems, about 20 percent got one to a pilot, and only about 5 percent reached production. The report calls itself preliminary, its numbers come from interviews rather than audited reporting, and critics have questioned its methods, fairly. But the pattern shows up in every serious survey, including ones with no stake in the answer. S and P Global Market Intelligence found that 42 percent of companies had abandoned most of their AI initiatives in 2025, up from 17 percent in 2024, and that the average organization scrapped 46 percent of its proofs of concept before production.

Gartner, in June 2025, predicted that more than 40 percent of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. The same release named the supply side's disease, agent washing: vendors rebranding ordinary chatbots and automation as agents. Of the thousands of vendors claiming to sell agents, Gartner estimates only about 130 are real. If you are buying, that is the base rate. If you are selling, being one of the real ones is the whole marketing strategy.

And then there is churn, the quiet killer that ARR announcements are designed to hide. Remember Klarna's $60 million in savings and its agent doing the work of 853 people. In May 2025, between those boasts, Klarna reversed course and started hiring humans again, in an on demand model, after customers complained the agent's answers were generic and the company had let go of the people holding the institutional knowledge. Forrester analyst Kate Leggett called Klarna almost the poster child for bad AI deployment, noting the overpivot to cost cutting landed suspiciously close to its IPO. Klarna kept the agent and restored the human path, which is the actual lesson: the money was real, the quality bill came due anyway, and the winning configuration was agent plus humans, not agent instead of humans. Why agents fail technically, in production, in ways that cost people money, is a whole subject of its own, and we documented it in What Breaks: The Documented Failure Modes of AI Agents.

Hold both ledgers at once and a pattern sharp enough to use emerges. The failures are broad, vague, and aimed at everything. The successes are narrow, specific, and aimed at one expensive workflow with a measurable result. MIT's data has the sharpest version of this: buying from specialized vendors and building in partnership succeeded about 67 percent of the time, roughly triple the success rate of internal builds, and the winners picked one pain point and executed it well. The market is not paying for potential. It is paying for finished, verified, boring work.

Four paths you can actually copy

Everything below is assembled from the cases above. None of it requires venture funding. All of it requires the unfashionable ingredients the winners share: a narrow problem, a customer who already spends money on it, and a human accountable for the result.

Path 1: the freelance builder. Sell agent builds to small and mid sized companies on the marketplaces where the demand spike is measured, or directly to businesses you already know. The offer: one workflow, automated, for a fixed fee, plus a monthly care retainer. The stack is the one buyers are searching for by name: Make.com or n8n for the wiring, a model API from OpenAI or Anthropic for the reasoning, and the client's own tools, their CRM, inbox, and calendar, as the agent's hands. Cost to start is genuinely small: free tiers cover your first builds, and your real investment is the weeks of learning, which is exactly the knowledge gap Fiverr says companies are paying to close. Land the first client with a working demo built on their kind of data, not a slide deck.

Path 2: the niche agency. Take Path 1 and remove the word "companies." Pick one industry you understand, one workflow that costs it money every week, and one product shaped offer: setup fee, monthly retainer, outcome language in the contract where you can manage it. A support agent for stores on one platform. An intake and follow up agent for one kind of practice. The Fiverr data shows where the generic demand is; the Sierra story shows why vertical wins, because the agent that knows how one industry's work actually flows is the one that survives contact with it. Your moat is not the model. It is the industry's edge cases, learned client by client.

Path 3: the micro product. The Base44 and Lovable path, at whatever scale the market gives you. Build one agent powered tool that does one painful job, charge a subscription from day one, and price with a usage component so your heaviest users cannot eat your margin in API costs, the pricing structure that kept Base44 profitable at $20 to $200 a month. Build it with the AI coding tools themselves, Cursor, Claude Code, or an app builder, which is how one person now ships what used to take a team. Cost to start: a domain, a few subscriptions, API usage you only pay when users show up. Carta's numbers say the solo founder is becoming the normal founder. The honest warning from the same story: distribution is the job. Shlomo grew through building in public and product led sharing, not through the product being good in a corner where nobody saw it.

Path 4: the inside job. If you already have a job, the fastest documented route to agent income is a raise, a promotion, or a new role, not a resignation letter. MIT's winners chose one pain point, partnered smartly, and measured outcomes; the person inside a company who does that for one workflow, killing an outsourcing contract or a manual process with a number attached, becomes the company's agent person while everyone else is still forming committees. The back office is where MIT found the clearest returns, in operations and finance work nobody puts in a keynote. Unglamorous, measurable, and promotable. This path also compounds into Paths 1 and 2 later, with a portfolio of results instead of promises.

Watch: How to Build & Sell AI Agents: Ultimate Beginner's Guide, from Liam Ottley. Whatever discount you apply to income claims in this genre, and you should apply one, Ottley's free guide is the most complete walkthrough of the agency model in these paths, from what an agent is to getting the first client.

And the paths not to take, briefly, because the failure ledger earned them a mention. Do not sell a thin wrapper around a chatbot and call it an agent; Gartner's 130 real vendors out of thousands is the market learning to tell the difference. Do not promise autonomy you cannot supervise; the documented failures in our companion post are what unsupervised agents do with production access. And do not build the broad thing first. Every success in this article started narrower than its founder wanted. Every failure in this article started with everything.

Why this belongs at Starvard

Agents are the first technology where a single person can plausibly do the work of a small company, and the gap between the people earning from that and the people buying courses about it is mostly knowledge. Starvard's course on AI agents teaches how agents are actually built, evaluated, and kept safe. Our AI course covers the underlying technology in plain language. And because every path in this article is finally a small business, our course on money covers the unglamorous half: pricing, margins, and knowing whether the number on the screen is profit.

Watch

• Lenny's Podcast, "Solo founder, $80M exit, 6 months: The Base44 bootstrapped startup success story": youtube.com/watch?v=L9KvV_UOs3A. The founder of Base44 tells the whole story in his own words: the revenue, the stack, the all nighters, and the sale.
• Anthropic, "Building more effective AI agents": youtube.com/watch?v=uhJJgc-0iTQ. The engineering reality under the business stories: what agent designs survive production, from the team that builds Claude.

Sources

• Inc., "How Base44 Found a Buyer 4 Months After Launching," December 2025. The solo founder account: nearly $1.5 million in revenue in the first month, subscriptions from $20 to $200 a month, and the $80 million Wix acquisition four months after launch: inc.com
• Nomixy, "Bootstrapped AI SaaS Lessons from Base44," summarizing TechCrunch's deal reporting, Calcalist's interview with Shlomo, and Carta's Founder Ownership Report. Base44 at sale: roughly 250,000 to 300,000 users, about $3.5 million ARR, close to $189,000 profit in its best month, eight employees, about $25 million in retention; solo founders rose from 23.7 percent of new startups in 2019 to 36.3 percent in the first half of 2025: nomixy.com
• WebProNews, "Sierra AI Startup Hits $100M ARR in 21 Months with Enterprise Agents," November 2025, reporting Sierra's announcement and TechCrunch's coverage. $100 million ARR announced November 21, 2025, seven quarters after launch, up from roughly $20 million the year before, with customers including SoFi, Ramp, and Brex: webpronews.com
• Insider Finance, "Salesforce Earnings Lift Guidance on AI Demand," December 3, 2025, reporting Salesforce's fiscal Q3 2026 results. Agentforce ARR above $500 million, up 330 percent, more than 9,500 paid deals, over 18,500 total deals since launch, and $1.4 billion combined ARR for Agentforce and Data 360: insiderfinance.io
• TechCrunch, "As Lovable hits $200M ARR, its CEO credits staying in Europe for its success," November 19, 2025. Lovable doubled ARR to $200 million four months after passing $100 million in July 2025: techcrunch.com
• TechCrunch, "Lovable says it added $100M in revenue last month alone, with just 146 employees," March 11, 2026. $400 million ARR in February 2026, 146 full time employees, $2.77 million of ARR per employee, and the milestone sequence of $100 million, $200 million, and $300 million before it: techcrunch.com
• Macha, "Intercom Fin Pricing 2026: $0.99 per Outcome, Explained," checked against Intercom's published pricing. Fin charges $0.99 per billable outcome, at most once per conversation, with a 50 outcome monthly minimum on other help desks, and $9.99 per sales qualification: getmacha.com
• Future Career Info, "AI Agent Economy: 67 Startups Bet on Boring Niches," 2026. Outcome based price lists across the industry (Zendesk at $1.50 to $2.00 per autonomous resolution from August 2024, Intercom at $0.99 per outcome, Salesforce at $2.00 per conversation) and PitchBook's count of 67 agent startups among the 144 in Y Combinator's Spring 2025 batch: futurecareerinfo.fyi
• Computerworld, "Demand for freelancers explodes, thanks to agentic AI," May 15, 2025, reporting Fiverr's Spring 2025 Business Trends Index. Searches for AI agent development up 18,347 percent in six months, Go High Level up 1,489 percent, Make.com up 1,083 percent, and Mark Kashef's account of his own order mix on the platform: computerworld.com
• CX Dive, "Klarna says its AI agent is doing the work of 853 employees," November 20, 2025. Klarna's claims of $60 million saved, two thirds of inquiries handled, response times improved 82 percent, repeat issues down 25 percent, alongside the May 2025 return to human hiring and Forrester's critique: customerexperiencedive.com
• Daylight Studio, "MIT 95% AI Failure Study: What It Really Says," analysis of MIT Project NANDA's The GenAI Divide: State of AI in Business 2025. About 95 percent of organizations saw no measurable profit and loss impact despite $30 to 40 billion in spending, the 60, 20, and 5 percent funnel from evaluation to production, partnerships succeeding about 67 percent of the time, and the report's stated limits, plus S and P Global's finding that 42 percent of companies abandoned most AI initiatives in 2025: thedaylightstudio.com
• Gartner, press release, June 25, 2025. The prediction that over 40 percent of agentic AI projects will be canceled by the end of 2027, the agent washing problem, and the estimate of about 130 real agentic AI vendors among thousands claiming the label: gartner.com