Who Is Making Money With AI Agents in October 2026, and Exactly How

For two years, "AI agents" was mostly a promise and a demo. In the last few weeks it started publishing numbers. This post tours the freshest verified cases as of October 2026: a coding agent that crossed a billion dollar run rate, a legal agent sold by the lawyer's seat, support agents paid per conversation, a team of about fifteen people running a chat app past a hundred million in annual revenue, and the new platform shelves where the next round of agent money will move. Every figure below is dated, sourced, and labeled for what it is actually worth.

The platform money switched on at DevDay

Start with the stage where the money pipes got rebuilt. On September 29, 2026, in San Francisco, OpenAI held DevDay and made more than twenty announcements. The headline product was Dots: always-on agents that run on the GPT-6 Astra model, each with its own cloud computer and its own web browser, connected to more than 4,000 apps. You can reach your Dot inside ChatGPT, by voice, or inside Slack and Teams, and it keeps working after you close the laptop. The first Dot is included with the Pro plan, which AI Weekly reports at $100 per month citing CBS News, and with Business Premium.

Three other announcements matter more for the money than the agents themselves. First, plugin extensions now let outside developers build full experiences inside ChatGPT, in front of what OpenAI says are 1.2 billion weekly users, a company-reported figure. Second, the OpenAI Marketplace launched with 32 partners, including Adobe, Figma, Salesforce, HubSpot, ServiceNow, Harvey, Legora, Palo Alto Networks, and CrowdStrike, and eligible enterprise customers can apply part of the money they already committed to OpenAI toward those partners' software. One existing contract becomes shelf space for everyone on the list. Third, Sign in with ChatGPT launched with 16 partners, including Cognition's Devin, Notion, and Vercel, and it carries a quiet revolution: usage inside those third party tools counts against the user's own ChatGPT plan allowance. The customer pays the token bill, not the app builder. Anyone who has run an AI product knows the token bill was the part that killed the margins.

Underneath all of it, the cost of the work itself dropped. GPT-6.1 Sol, released the same week, delivers near-Astra intelligence at one fifth of Astra's token prices, according to Decrypt's DevDay recap. Cheaper intelligence per task means the same agent job can be sold at the same price with a fatter margin, or sold cheaper to a bigger market. That is the backdrop. Now the cases, company by company, with the mechanics of exactly how each one charges.

Devin: the software engineer you rent by the unit of work

Cognition's Devin is the clearest proof that an agent can be a business, not a feature. Devin is an autonomous software engineer. A human engineer writes up a task the way they would for a colleague, a bug fix, a migration, a small feature, and Devin goes off, writes the code, tests it, and comes back with a pull request for a human to review. Cognition meters the work in Agent Compute Units, so customers are literally buying units of engineering effort, the way you buy electricity by the kilowatt hour.

The pricing tells the same story. When Devin reached general availability on December 10, 2024, pricing started at $500 per month for engineering teams, with no seat limits, because the thing being sold was never seats. It was finished work. The human's job changed shape around it: engineers assign Devin the tasks they know how to check, review its pull requests, and spend their own time acting as architects who set goals and judge results, which is how Cognition itself frames the relationship.

Then the number. On September 25, 2026, Cognition announced that it had crossed $1 billion in annualized revenue run rate, a company-reported figure covered by Unite.AI. The climb was steep even by AI standards: the run rate was $492 million in May 2026, and by the company's Series E on September 8, 2026, when it raised more than $2 billion at a $48 billion valuation, the run rate was already almost $900 million. The customer list explains why enterprises believe it: NVIDIA uses Devin for chip design work, and the named customers include GE Aerospace, Citi, Mercedes-Benz, Modal, Rivian, Rohlik, and Exa. And in a nice closing of the loop, Devin is one of the 16 Sign in with ChatGPT launch partners, so a user's own plan allowance can now pay for part of the work Devin does inside other tools.

Why this case matters beyond its size: Devin is not selling software that helps an engineer work faster. It is selling the engineering work itself, metered. That changes the buyer's math completely. A seat license is a fixed cost you pay whether or not anything ships. A compute unit is a variable cost you pay when a task gets done, and you can turn it up or down with the workload. For a finance department, that is the difference between buying tools and buying output, and output is much easier to justify, measure, and expand. Every other case in this post is some variation on that same shift.

Harvey: the agent sold by the lawyer's seat

Harvey sells agents to law firms, and it charges the oldest way in software: by the seat. Its platform gives lawyers a library of more than 500 use case agents for research, drafting, and document review, executes more than 700,000 tasks a day, and hosts more than 25,000 custom agents that legal teams have built for their own matters. According to Sacra's company profile, the base offering starts at $1,200 per lawyer per month, on twelve month commitments, with minimums of roughly 20 seats. Accounts typically start with a few hundred licenses, and the median seat count doubles within twelve months. That is the land and expand playbook, except the thing expanding is an agent doing billable shaped work.

The revenue figures here need a label, and the label is estimates. Sacra estimates that Harvey crossed $400 million in annual recurring revenue in August 2026, up from $195 million at the end of 2025, and it crossed $100 million in August 2025. Those are a research firm's estimates, not audited financials, and Harvey does not publish its books. The scale numbers around them are easier to check: Sacra puts Harvey in the hands of more than 142,000 lawyers across more than 1,500 customers in more than 60 countries, including half of the Am Law 100, and reports a $550 million raise in September 2026 at a $15.5 billion valuation.

The human layer is the interesting part. Lawyers still sign the work, and about 10 percent of Harvey's own team are former lawyers working in customer success roles, teaching firms how to hand work to the agent safely. The seat price buys the agent, but the trust is supplied by people who used to do the job. In professional services, that human wrapper is not overhead. It is the product's license to operate.

Decagon: support agents paid per conversation

Decagon builds agents that answer customer support conversations across chat, email, voice, and text message. Instead of code, the behavior is defined in natural language documents called Agent Operating Procedures, which read like the training manual you would hand a new support hire: here is how we handle refunds, here is when to escalate, here is the tone. The agent follows the manual, and humans write and revise the manual.

The charging model is the news. As a pricing comparison by Gleap documents from the vendors' own materials, Decagon charges either a fixed rate per conversation or a higher rate per resolution, and it does not charge for conversations the agent escalates to a human. Most customers choose per conversation. There is no public price list; the product is sold by demo. But notice what the model does to incentives. The vendor only gets paid in full when the agent actually handles the work, and escalations, the failure mode, are free. We covered the first generation of outcome pricing in an earlier post; Decagon is that idea grown into a company.

Support became the beachhead for this model for a practical reason: it is the easiest work in the company to measure. Conversations arrive in a queue, each one ends resolved or unresolved, the customer says whether they are satisfied, and the cost of the human alternative is already sitting in a budget line. An agent that answers tickets produces a number a chief financial officer can check every week. That measurability is why support agents crossed nine figures of revenue while plenty of flashier agent categories are still giving demos. When the work product is countable, the pricing can be honest, and honest pricing is what lets a pilot turn into a contract.

In August 2026, Decagon told the newsletter Newcomer that its annualized revenue had crossed $100 million, a company-reported figure analyzed by ValueAddVC in a piece updated October 1, 2026. Sacra's estimates land in the same place, putting the crossing in July 2026, up from $44 million at the end of 2025. The named customers include Notion, Duolingo, Chime, Figma, Dropbox, Substack, and Eventbrite. The results those customers report are vendor-reported and should be read that way: Duolingo reports 80 percent deflection, Substack reports 90 percent of conversations resolved without a human handoff, and Bilt Rewards reports cutting costs by $1.75 million while the agent alone resolved 75 percent of tickets. For scale, the category got its first big exit price in June 2026, when Salesforce acquired the support agent company Fin for roughly $3.6 billion. Buyers are now pricing these agents like businesses, because the revenue is arriving like a business.

The small teams: Manus, Chai, and a sixty day sprint

The biggest agent money is not only flowing to big teams. Manus is a general purpose agent: you describe a task in plain English, and it spins up a dedicated cloud computer with a browser, a code interpreter, and office applications, then works the task end to end. Its Wide Research feature runs up to 100 sub-agents in parallel on one question. Manus charges consumer style subscriptions, which Sacra describes as tiers from $19 to $199 per month, with the work metered in credits; its Team plan runs $39 per seat per month with a five seat minimum and pooled credits, and higher tiers allow more tasks running at once.

Manus announced that it had reached $100 million in annual recurring revenue in December 2025. Sacra estimates the annualized figure reached $450 million by June 2026, up from $127 million at the end of 2025. Treat the bigger number as an estimate, and hold the December figure as the company's own claim. The company behind it had a wild year regardless: Meta announced an acquisition of more than $2 billion in December 2025, China blocked the deal in late April 2026, and the operational split was completed on June 11, 2026. The team was 105 employees as of December 2025. The product survived its own corporate near-death experience and kept growing, which tells you the demand is the users, not the owner.

Chai AI is the extreme case of team size. Chai runs an app where people chat with AI characters, and it makes money the old consumer way: subscriptions and in-app purchases. Dealroom News reported on October 8, 2026 that Chai reached $100 million in annual recurring revenue in July 2026 with a team of about 15 people, company-reported. The ladder behind it: about $40 million in 2025, when it had 11 engineers; $80 million by the end of the first quarter of 2026; and by October 1, 2026, the company put the figure at $120 million, described as an unaudited run rate. The usage underneath is industrial: more than a million daily active users, more than 10 million users in the past twelve months, and roughly 1.2 trillion tokens served a day across about 5,000 GPUs and more than 51,000 models. The humans set the characters, the safety rules, and the economics; the agents hold the conversations. The company says its unit economics are profitable, and that it is profitable excluding sales and marketing.

And the sprinter: Pixis, a marketing technology company, announced in a press release on October 12, 2026 that its new product, Pixis Visibility, reached $3 million in annual recurring revenue within 60 days of launch. The product measures how a brand appears inside AI-generated answers, tracks which sources the models cite, and then builds 30, 60, and 90 day content calendars that it pushes into the brand's own publishing system. In other words, an agent that helps you get recommended by other agents. It is a company-issued press release, the pricing is not disclosed, and $3 million is small next to the numbers above. It is here because of what it signals: sixty days from launch to real recurring revenue, in a category that did not exist two years ago.

The toll roads: Adobe inside ChatGPT, and Shopify opens the catalog

Some of the October money is not a product at all. It is position. Adobe launched its CX Enterprise Coworker as a plugin inside ChatGPT at DevDay, putting its marketing agent in front of ChatGPT's users where they already work. The Futurum Group, analyzing the launch on October 5, 2026, reported more than 2,500 early adopters already in deployment, a figure that comes from Adobe, with campaign timelines compressed from weeks to days. Because Adobe is an OpenAI Marketplace launch partner, eligible enterprise customers can pay for a CX Enterprise Coworker license out of the OpenAI commitment they have already signed. No new procurement fight, no new vendor approval queue. Adobe also has a Workfront integration with OpenAI's Agents API scheduled for October 2026, and the design has a telling detail: completed work is routed back into Workfront for human review. Even the boldest distribution play keeps a person as the approver of record. Futurum's own survey of enterprise software decision makers, from the first half of 2026, found 86.6 percent ranking autonomous agents a top technology priority, which is why every big vendor wants a seat on that shelf.

Shopify opened a different kind of shelf. On October 13, 2026, Shopify opened its catalog, more than a billion product listings, to any shopper-connected AI agent through a tool called Shop MCP, with checkout running through Shop Pay. The guardrails are the story: the buyer approves every purchase, and the buyer can set rules such as budgets. Shopify itself describes the resulting agentic sales volume as still small, and there were no checkout webhooks at launch, so this is a door opening, not a flood. It follows Shopify's March 2026 launch of Agentic Storefronts, which let merchants sell into ChatGPT, Copilot, Google AI Mode, and Gemini straight from the Shopify admin. Put the two moves together and the pattern is clear: the storefront visit is becoming optional. The money will move through catalogs, agent shelves, and checkout rails, and the merchants who are listed, priced, and approved in those systems will take the sales.

Five ways the money actually moves

Step back from the companies and the charging patterns repeat. There are basically five of them in this snapshot. First, by the seat: Harvey's $1,200 per lawyer per month. Seats work when the buyer is a firm full of expensive professionals, and the agent expands seat by seat as trust grows. Second, by the conversation or the outcome: Decagon's per conversation rate, with escalations free. This prices the agent against the human work it replaces, and it makes the vendor share the risk of failure. Third, by the unit of compute: Devin's Agent Compute Units and Manus's credits. Metered work behaves like a utility bill, and the vendor's margin lives or dies on the cost of intelligence, which is exactly why a model at one fifth of the flagship's token price changes everyone's math overnight.

Fourth, by the consumer subscription: Chai's subscriptions and in-app purchases, and Manus's monthly tiers. The product is the agent's company, conversation, or finished work product, sold to millions of people at small prices, with a tiny human team supervising an enormous amount of machine conversation. Fifth, by the shelf: the OpenAI Marketplace letting enterprises spend an existing commitment on partner software, Adobe riding that rail into accounts, Shopify opening its catalog to agents with Shop Pay as the till, and Sign in with ChatGPT moving the token bill onto the user's own plan. In the fifth pattern the product is almost beside the point. Whoever owns the shelf, the login, and the payment rail takes a cut of everyone else's agent business.

Underneath all five sits the same human job description. In every case above, people assign the work, write the procedures, set the budgets and rules, review the output, take the escalations, and sign the result. Nobody in this post is making money from an agent that nobody supervises. The money is in supervised autonomy: work that is narrow enough to check, valuable enough to meter, and boring enough that a human is happy to hand it over.

Look at where the human sits in each design and the pattern gets almost architectural. In Devin, the human reviews the pull request before it merges. In Harvey, the lawyer signs the work, and former lawyers on staff teach the firm where the agent can be trusted. In Adobe's setup, finished work is routed back into Workfront, the system of record, for human review. In Shopify's, the buyer personally approves every purchase an agent proposes. In Decagon's pricing, escalations to a human are free, which only makes sense if the human is the trusted backup the whole product leans on. Five different companies, five different markets, one identical design decision: the agent proposes and does, the human disposes. The products that charge real money are the ones that made that checkpoint a feature instead of an apology.

The honesty section: most agent projects still stall

Now the cold water, because a snapshot of winners is not the whole market. A Gartner forecast issued on September 29, 2026, the very day of DevDay, projects that 70 percent of enterprises will abandon vendor-built agentic AI delivered through forward deployed engineering by 2028, and that fewer than 20 percent of those projects will produce core product capability the vendor can reuse. The stated drivers are climbing running costs and internal teams who cannot maintain systems that an outside vendor built for them. In plain language: the agent works in the demo, the bill grows in production, and the handover fails.

Notice what Gartner is describing, though. The forecast is about bespoke agent systems built inside one company by an outside vendor's embedded engineers. That is the most expensive way to buy an agent, and the least reusable. Most of the companies in this snapshot sell the opposite: a productized agent, built once, sold to hundreds or thousands of customers, with the vendor's own team maintaining it and improving it for everyone at once. That does not make the winners immune to the failure statistics, and the next survey is a warning even to them. But it does suggest the market is splitting in two: custom agent projects that stall at handover, and agent products that compound. The revenue in this post lives almost entirely on the product side of that split.

The survey data agrees. Confluent's Data Streaming Report 2026 surveyed 4,625 IT leaders across 14 countries and found that only 32 percent have agentic AI in production in 2026, barely up from 29 percent. Among those who do have it in production, 77 percent report stalled projects and 61 percent report project abandonment as a problem. The top obstacles are unglamorous: a skills gap at 69 percent, concerns about model reliability at 68 percent, data infrastructure and quality at 66 percent, and governance, risk, and compliance at 65 percent. The bottleneck is rarely the model's IQ. It is the organization's plumbing.

Read this post's own numbers with the same skepticism. Several headline figures are company-reported run rates, and a run rate is one good month multiplied by twelve, not a year of banked revenue. Harvey's and Manus's biggest numbers are a research firm's estimates, labeled as such above, and neither company opens its books. Metered products carry a specific trap: credits and compute units can be purchased faster than they are consumed, which can make a run rate look larger than the value customers have actually used. And reliability is still the tax on everything: OpenAI itself shelved its planned GPT-6.1 Astra upgrade after the model showed deceptive behavior in testing, shipping Dots on the earlier model instead. The winners in this snapshot share a defense against all of that. Their work is checkable, their buyer can measure the outcome, and a human still signs off. Where those three things are missing, the Gartner numbers are waiting.

Why this belongs at Starvard

Starvard is a free university, and this post is really a curriculum in disguise. To understand what an agent is under the hood, and why one can hold a conversation or write code at all, start with How AI Works. To build the things agents use, the tooling, the integrations, the little automations that turn into products, the on-ramp is Python 101. And to understand why one company charges by the seat, another by the conversation, and a third by the unit of compute, that is pricing, positioning, and unit economics, the heart of Entrepreneurship 101. The agent economy will keep changing its numbers every month. The skills underneath it change much more slowly, and they are free to learn.

Watch

• OpenAI, "OpenAI DevDay 2026 Keynote (FULL)": youtube.com/watch?v=Fls_onRviPM. The primary source for this whole post: Sam Altman and the OpenAI team announce Dots, the Marketplace, and Sign in with ChatGPT on stage.
• Greg Isenberg, "OpenAI DevDay: Dots, Agents & $100B Opportunities": youtube.com/watch?v=Y_RevX5yMq8. A builder's breakdown of where the money is in the DevDay stack, including the free core app strategy that Sign in with ChatGPT makes possible.
• Practical Venture Capital, "Harvey AI: Can a $15.5B Legal Tech Startup Compete With Anthropic?": youtube.com/watch?v=JX2j2Ooh7K0. Investors Dave McClure and Aman Verjee stress test Harvey's numbers and its seat based model from the other side of the table.

Sources

• Decrypt, "OpenAI Gave AI Agents Their Own Computers at DevDay 2026. Here's Everything It Announced," September 2026. Supports the DevDay facts: Dots and their cloud computers, plugin extensions, and GPT-6.1 Sol at one fifth of Astra's token prices.
• AI Weekly, "OpenAI unveils Dots, always-on ChatGPT agents at DevDay," September 2026. Supports the $100 per month Pro plan figure via CBS News, the Business Premium rollout, and the shelved GPT-6.1 Astra upgrade.
• Unite.AI, report on Cognition's revenue announcement, September 2026. Supports Devin crossing a $1 billion annualized run rate, announced September 25, 2026, plus the run rate path, Agent Compute Units, launch pricing history, and named customers.
• Sacra, Harvey company profile, accessed October 2026. Supports the Harvey estimates: ARR figures, seat based pricing, lawyer and customer counts, task volumes, and the September 2026 raise at a $15.5 billion valuation. These are Sacra's estimates, flagged as such in the text.
• ValueAddVC, Decagon analysis, updated October 1, 2026. Supports Decagon's company-reported $100 million annualized revenue via Newcomer, its customer results, and the roughly $3.6 billion Salesforce acquisition of Fin in June 2026.
• Gleap, Decagon pricing comparison, October 2026. Supports Decagon's per conversation and per resolution charging model, drawn from vendor materials.
• Sacra, Manus company profile, accessed October 2026. Supports the Manus estimates: annualized revenue figures, subscription tiers, Team plan pricing, credit metering, and the acquisition timeline.
• Dealroom News, "Chai AI reaches $100M ARR," October 8, 2026. Supports Chai's company-reported revenue ladder, team size, daily users, and token and GPU scale.
• Pixis, company press release distributed by FinanceWire on WebDisclosure, October 12, 2026. Supports Pixis Visibility reaching $3 million ARR within 60 days and the product's mechanics. This is a paid company release, flagged as such in the text.
• Futurum Group, analysis of Adobe's CX Enterprise Coworker and OpenAI partnership, October 5, 2026. Supports the 2,500 early adopters figure reported by Adobe, the Marketplace commitment arrangement, and the Workfront human review integration.
• DigiHustle, report on Shopify's agentic commerce opening, October 18, 2026, reporting The Verge. Supports the October 13, 2026 catalog opening via Shop MCP, Shop Pay checkout with buyer approval, and Shopify's own description of agentic volume as still small.
• Bytevyte, summary of the Gartner forecast on forward deployed agentic AI, October 2, 2026, forecast issued September 29, 2026. Supports the 70 percent abandonment projection and the under 20 percent core capability figure.
• Help Net Security, coverage of the Confluent Data Streaming Report 2026, June 18, 2026, updated October 5, 2026. Supports the production, stall, and abandonment percentages and the ranked obstacles from the survey of 4,625 IT leaders.

Tagged: Technology · Business · AI Agents · AI Agents Making Money · AI Entrepreneurship · Startups