Where is the money in the agentic AI market?

Where is the money in the agentic AI market right now?

The biggest money in agentic AI currently comes from coding, customer support, legal work and enterprise software, where companies are already paying to have AI complete specific tasks. The broader dream of autonomous agents running entire departments is much further from becoming a proven business.

Menlo Ventures estimated that enterprises spent $37 billion on generative AI in 2025, up from $11.5 billion in 2024. Of that total, $19 billion went to applications and $18 billion to infrastructure. These figures cover generative AI more broadly, including conventional assistants and model APIs, so we cannot treat the entire $37 billion as agentic AI revenue. Spending totals blur agents with ordinary assistants; we break down how much money went in, quarter by quarter, in our report on agentic AI.

The breakdown is revealing. Coding tools attracted $4 billion in enterprise spending, while horizontal agent platforms accounted for approximately $750 million. General-purpose copilots generated $7.2 billion, nearly ten times more than dedicated agent platforms.

Meanwhile, Cognition, the company behind Devin, reported more than $1 billion in annualized revenue in September 2026. Salesforce disclosed over $1.5 billion in Agentforce annual recurring revenue, although that figure includes a broader range of AI offerings. Sierra and Harvey have also reached hundreds of millions in estimated recurring revenue.

How big is the agentic AI market today?

The standalone agentic AI market is still much smaller than the broader AI industry, and there is no reliable single figure that captures every dollar companies spend on autonomous agents.

Part of the problem is how vendors classify their products. A coding assistant that suggests code, an agent that independently fixes a bug and a platform coordinating dozens of agents may all be counted under AI software spending.

Menlo Ventures' 2025 enterprise survey gives us a useful starting point. Its $750 million estimate for horizontal agent platforms was dwarfed by spending on coding tools, copilots and foundation-model APIs. However, the estimate does not include every agent embedded in departmental software.

Enterprise AI category	2025 spending	What the figure tells us
General-purpose copilots	$7.2B	Largest horizontal software category
Coding AI	$4.0B	Strongest departmental application category
Vertical AI applications	$3.5B	Includes legal, healthcare and other industries
IT operations AI	$700M	Infrastructure and incident automation
Marketing AI	$660M	Campaign and marketing workflows
Customer success AI	$630M	Service and retention workflows
Horizontal agent platforms	$750M	Explicitly marketed agent platforms
Foundation-model APIs	$12.5B	Model consumption across many AI products
AI infrastructure software	$1.5B	Retrieval, orchestration and related tools

These are overlapping categories rather than amounts that can be added together. Coding is part of departmental AI, while horizontal agent platforms are part of horizontal AI. Foundation-model spending also supports applications counted elsewhere.

Why are companies spending more on AI agents now?

Large companies are buying AI agents faster because the technology has become useful for coding, IT operations and other workflows that previously required constant human intervention.

McKinsey's August 2026 global AI survey found that 40% of respondents from organizations generating more than $1 billion in annual revenue were scaling AI agents in at least one business function. A year earlier, the figure was 27%.

That 13-percentage-point increase represents roughly 48% growth in the proportion of large organizations reporting scaled adoption. Going from 27% to 40% in a year shows how fast adoption moves; we look back over the past twelve months in our agentic AI funding report.

The contrast with smaller businesses is striking. Among organizations below $1 billion in revenue, the proportion remained around 22%. The adoption gap between the two groups widened from approximately five to 18 percentage points.

McKinsey also found that about one in five respondents reported scaling software coding agents across their organizations, rising to 31% among large enterprises. IT, knowledge management and software engineering were among the functions where agent deployment was most common.

However, Gartner reported in October 2026 that only 17% of organizations had deployed AI agents, although more than 60% expected to do so within two years. The studies use different populations and definitions, but both show how much anticipated demand has yet to turn into established deployment.

Is AI coding the most lucrative agentic AI business?

AI coding is currently the strongest candidate for the largest established agentic AI application market, with $4 billion in estimated enterprise spending and unusually fast revenue growth among specialist companies.

Menlo Ventures calculated that coding accounted for 55% of departmental AI application spending in 2025. Spending rose from approximately $550 million in 2024 to $4 billion in 2025, an increase of more than sevenfold.

That expansion coincided with coding models becoming capable of understanding larger codebases and completing multi-step development tasks.

Cognition provides an even more recent example. Its annualized revenue run rate increased from $492 million in May 2026 to almost $900 million in early September, before exceeding $1 billion later that month. That is approximately a doubling in four months, although the company also expanded through its acquisition of Windsurf.

Microsoft reported 50 million GitHub Copilot users in its fiscal 2026 annual report. That number covers the wider coding-assistance product and should not be confused with 50 million paying customers for fully autonomous agents.

Software development has another practical advantage: code can be compiled, tested and reviewed before deployment. That makes it easier to measure whether an agent completed its assignment correctly than in many legal, financial or customer-facing workflows.

Can AI customer support agents become a billion-dollar business and actually save money?

AI customer support is already producing companies with hundreds of millions in recurring revenue, and the business can scale because customers pay for resolved interactions that would otherwise require human work.

Sierra is one of the clearest examples. According to Sacra, its annual recurring revenue increased from approximately $26 million at the end of 2024 to $130 million at the end of 2025 and an estimated $200 million in May 2026.

That amounts to roughly 7.7 times growth in less than a year and a half. Sierra reached its first $100 million ARR in approximately seven quarters after launching commercially in February 2024.

Salesforce is selling similar capabilities through Agentforce, while Intercom offers Fin, an AI customer service agent that can answer questions and execute configured workflows.

Intercom's standard price starts at $0.99 per billable outcome. Salesforce publishes a $2-per-conversation option for Agentforce. Sierra generally uses negotiated enterprise contracts linked to usage or outcomes.

Consider a company handling one million eligible customer interactions annually. At the published rates, one million billable outcomes would generate $990,000 for Intercom, while one million billed conversations would generate $2 million for Salesforce, before discounts or other fees.

The economics can also work for customers. If a human-handled request costs $5 and an AI agent successfully resolves the same request for $0.99, automating 500,000 cases would generate approximately $2 million in theoretical gross annual savings.

That calculation excludes integration, supervision and additional software costs. It also assumes each automated resolution actually replaces human work.

Intercom's billing documentation includes several types of outcomes, including certain completed procedures, handoffs and inferred resolutions. Companies therefore need to monitor repeat contacts, escalations and customer satisfaction rather than relying entirely on a vendor's resolution count.

Intercom also highlights Synthesia as a customer that reduced support resolution times from five days and five hours to four hours and 37 minutes after deploying Fin. That is a vendor-published customer result rather than an independently measured industry average.

Is legal AI already making serious money?

Legal AI is already one of the most lucrative specialized AI software markets, with Harvey estimated to generate around $400 million in annual recurring revenue.

According to Sacra, Harvey grew from approximately $50 million ARR at the end of 2024 to $195 million at the end of 2025, before reaching an estimated $400 million in August 2026.

Its recurring revenue therefore increased roughly eightfold in less than two years.

Harvey's products help legal professionals research cases, analyze documents, draft contracts and handle due diligence. Sacra reported more than 142,000 lawyers using the platform across over 1,500 customers in more than 60 countries.

Another interesting figure comes from customer expansion. Sacra estimated that Harvey's median number of licensed seats doubles within 12 months of deployment. Existing customers are clearly an important part of Harvey's growth, alongside new firms joining the platform.

Harvey primarily charges through enterprise subscriptions tied to users, unlike the outcome-based pricing increasingly common in customer support. Subscriptions for Harvey, outcome pricing for support: we compare which business models are winning in our agentic AI report.

Where is the money in healthcare AI agents?

Healthcare AI agents are finding commercial opportunities in clinical documentation and administrative work, where hospitals already spend heavily on repetitive tasks.

Microsoft's fiscal 2026 annual report said Dragon Copilot was on pace to document more than 100 million patient encounters during the calendar year. The company has also reported strong growth in automated encounters across its healthcare products.

That usage shows meaningful operational scale, although Microsoft does not separately publish Dragon Copilot's revenue or the average amount collected per encounter.

Abridge, Ambience Healthcare and Suki are among the specialists selling clinical documentation software. Their products turn conversations between clinicians and patients into structured medical notes and help organize information for existing healthcare systems.

The commercial opportunity extends into billing preparation, coding, administrative follow-up and other tasks connected to patient care.

Healthcare workflows require strict controls over patient information, accurate documentation and reliable integration with electronic health records. The clearest revenue opportunities therefore involve handling administrative work that clinicians can review and approve.

Public disclosures still make it difficult to rank the leading healthcare agent startups by recurring revenue or profitability.

Are AI agents making more money for software companies or model providers?

Foundation-model providers currently capture far more measurable enterprise spending than dedicated horizontal agent platforms, and the growth of AI agents gives them another source of demand.

Menlo Ventures estimated that enterprises spent $12.5 billion on foundation-model APIs in 2025, compared with approximately $750 million on horizontal agent platforms.

The comparison is imperfect because model APIs power many products beyond autonomous agents. Still, the gap illustrates how much money flows into the underlying technology.

Gartner's May 2026 forecast expected worldwide spending on AI models to rise from approximately $15.5 billion in 2025 to $32.6 billion in 2026. The firm identified greater model consumption through multi-step workflows as a contributor to the increase.

An agent can use models repeatedly while completing one assignment. It may retrieve information, decide on an action, call external software, inspect the result and correct mistakes.

Menlo's enterprise survey also found that Anthropic held approximately 40% of enterprise LLM API spending in 2025, compared with 27% for OpenAI and 21% for Google. Those three companies collectively accounted for roughly 88% of the market in that estimate.

Their eventual profits are harder to determine because model providers face significant infrastructure costs and pricing competition.

Can Salesforce and Microsoft dominate the agentic AI market?

Salesforce and Microsoft have a major advantage in enterprise agentic AI because millions of employees already work inside their products. Both companies can sell automation without asking customers to adopt an entirely new software ecosystem.

Salesforce reported more than $1.5 billion in Agentforce annual recurring revenue during its second quarter of fiscal 2027, growing over 240% year over year.

Agentforce ARR now includes Salesforce's broader AI offerings, Slackbot and Headless 360. The headline figure therefore cannot be attributed entirely to autonomous agents.

However, Salesforce also reported 3.2 billion Agentic Work Units during the quarter, up 97% from the previous quarter. That represents nearly twice as much recorded activity in three months under Salesforce's measurement system.

The company disclosed 7 billion Agentic Work Units delivered cumulatively, while bookings for Agentforce One Edition and Agentforce for Apps more than doubled quarter over quarter.

These are three different indicators: recurring contracted revenue, recorded AI work and customer purchasing activity. Together, they provide stronger evidence of commercial adoption than revenue announcements alone.

Still, Agentic Work Units are Salesforce-defined measurements rather than individually billed customer tasks. The broadened ARR definition also makes historical comparisons less straightforward.

Microsoft has a different route into the same market. Its 2026 annual report disclosed more than 30 million paid Microsoft 365 Copilot seats. Its subsequent earnings call reported nearly 40 million agents registered through Agent 365 across tens of thousands of companies.

The registered-agent figure does not tell us how many agents are actively used or generate incremental revenue.

Salesforce agents can interact with CRM data, while Microsoft's agents can use identity, productivity and security systems already deployed by their customers. Independent vendors can still win, as Sierra and Harvey demonstrate, but startups selling broad-purpose enterprise agents must compete against tools that organizations can activate within software they already own. Independents face tools companies already own, and we look at who is ahead and who is falling behind in our agentic AI market report.

Are AI agents replacing monthly software subscriptions with pay-per-task pricing?

AI agents are making pay-per-task pricing more common, especially in customer service, but traditional subscriptions remain important across coding, legal software and enterprise automation.

The difference depends largely on whether customers and vendors can agree on what counts as completed work.

Intercom charges $0.99 for standard Fin outcomes, while its newer sales qualification product can charge $9.99 for a qualified lead routed to a salesperson. Salesforce offers conversation-based and action-based Agentforce pricing.

Harvey, meanwhile, primarily sells subscriptions to legal professionals. UiPath generates revenue from its broader software subscriptions and associated automation offerings.

Company or product	Pricing approach	Publicly disclosed example
Intercom Fin support	Per outcome	$0.99 per standard outcome
Intercom Fin for Sales	Per qualification	$9.99 per qualified lead
Salesforce Agentforce	Per conversation	$2 per conversation
Salesforce Flex Credits	Per action	$0.10 per standard action
Salesforce Agentforce add-on	Per user	$125 per user/month
Salesforce Agentforce Coworker	Per user	$50 per user/month
Sierra	Negotiated enterprise usage and outcomes	No universal public price
Harvey	Enterprise subscriptions	Negotiated seat-based contracts
UiPath	Platform subscriptions and services	Enterprise contracts

One subtle difference is how vendors define an outcome. Intercom's Fin pricing documentation includes confirmed or inferred resolutions, configured handoffs and other completed procedures. A qualified sales lead commands a much higher price than a typical support resolution because the expected business value is different.

Can AI agent infrastructure become a billion-dollar business, including security?

AI agent infrastructure can support billion-dollar companies, while security and governance are becoming important parts of the market. Much of the established revenue, however, belongs to broader automation, cloud and software platforms.

Menlo Ventures estimated $1.5 billion in enterprise spending on AI infrastructure software in 2025, covering retrieval, orchestration, storage and related technologies.

Microsoft provides a more recent indication of demand. During its fiscal 2026 fourth-quarter earnings call, it reported 100,000 Microsoft Foundry customers and said Foundry revenue had more than doubled year over year.

The company also reported that the number of Foundry customers using at least one trillion tokens on an annualized basis had quadrupled.

UiPath demonstrates a different way to monetize agent infrastructure. Its platform coordinates traditional automation, AI agents, business systems and human supervision. During its second quarter of fiscal 2027, UiPath reported $1.938 billion in annual recurring revenue, up 12% year over year.

That figure covers the full UiPath platform, much of which existed before the current wave of AI agents.

Security introduces another opportunity as agents gain permission to execute code, access private records and modify business systems.

Microsoft's Agent 365 is designed to extend existing identity, security and management systems to AI agents. The company reported nearly 40 million registered agents shortly after launching the platform, although registration does not establish paid adoption.

Gartner's October 2026 research estimated that only 17% of organizations had deployed agents, while more than 60% expected to do so within two years. Gartner also forecast that over 40% of agentic AI projects could be canceled by the end of 2027 because of costs, unclear value or inadequate controls. That is a projection rather than an observed cancellation rate.

Independent infrastructure and security startups face competition from Microsoft, Salesforce and established cybersecurity providers such as Palo Alto Networks and CrowdStrike. Specialists have a clearer opportunity where enterprises need to coordinate agents across multiple vendors or maintain independent controls.

Are AI sales agents as profitable as customer support agents?

AI sales agents have an attractive potential market, but their commercial success is harder to prove than customer support automation because generating outreach is much easier than generating actual customers.

Companies such as 11x, Artisan and Unify offer tools that automate prospect research, outreach and parts of sales development.

A sales agent might send thousands of personalized messages, but that activity has little economic value unless it produces meetings, qualified opportunities and paying customers.

Intercom's Fin for Sales, for example, charges $9.99 for certain successful lead qualifications, compared with $0.99 for standard support outcomes. The tenfold difference reflects how much more valuable a qualified commercial opportunity can be than a routine support interaction.

Yet even a qualified meeting does not guarantee revenue. Companies need to track conversion rates, deal sizes and customer acquisition costs before deciding whether an AI sales agent outperforms a human-led or conventional automated process.

McKinsey's 2026 survey found that marketing and sales were among the functions where organizations most frequently reported AI-related revenue improvements. Those findings cover AI more broadly and cannot be assigned specifically to autonomous outbound agents.

Which agentic AI companies are actually making the most money?

The largest observable agent-related revenue figures belong to established software platforms and fast-growing specialists, but no clean ranking exists because companies report different products and financial metrics.

Cognition has become one of the clearest examples of rapid monetization in coding. Sierra has built a substantial customer-service business, while Harvey has reached a similar scale in legal software.

Salesforce and UiPath report larger recurring revenue figures, although both include broader offerings beyond autonomous agents. Bigger numbers at incumbents do not settle who leads; we cover who the top companies are today in our agentic AI funding report.

Company	Recent revenue indicator	What we can conclude
UiPath	$1.938B ARR	Established automation platform; not pure agent revenue
Salesforce Agentforce	Over $1.5B ARR	Broad AI product family including agent capabilities
Cognition	Over $1B annualized revenue run rate	Major coding AI business, including Windsurf
Harvey	~$400M ARR	Sacra estimate for legal AI
Sierra	~$200M ARR	Sacra estimate for customer-service AI
Microsoft 365 Copilot	Over 30M paid seats	Large paid AI suite, without separate agent revenue
Microsoft Foundry	100,000 customers	Significant platform adoption, agent revenue undisclosed
Intercom Fin	From $0.99 per standard outcome	Transparent pricing, total agent revenue undisclosed

Salesforce and UiPath disclose public-company metrics. Cognition's figure is a company-reported annualized run rate, while the Harvey and Sierra figures are third-party estimates from Sacra.

A run rate annualizes recent activity and does not mean a company has already collected that amount over a full year.

Are agentic AI companies actually profitable?

Most leading independent agentic AI startups have not disclosed enough financial information to establish their profitability, despite reporting exceptionally fast revenue growth.

Cognition, Sierra and Harvey provide recurring revenue figures or estimates, but public information does not allow us to calculate comparable gross margins, operating profits or cash generation.

UiPath offers a useful reference point from the established automation industry. Its second quarter of fiscal 2027 produced $410 million in revenue, an 80% GAAP gross margin and $32 million in GAAP operating income.

Those are actual public-company financial results, although UiPath's business includes conventional automation alongside agentic products.

The economics of autonomous agents can differ from conventional SaaS. A customer may pay a fixed subscription while the vendor incurs variable model and infrastructure costs every time the agent works.

McKinsey's 2026 survey found that approximately one in five organizations was limiting AI usage because of operating expenses, including token consumption. Around one in ten reported cost constraints associated specifically with AI agents.

A coding agent that completes a task successfully after a few model calls can have very different economics from an agent that repeatedly fails, retries and consumes expensive inference before requiring human assistance.

Could AI agents take money away from traditional software companies?

AI coding agents are already changing software purchasing decisions, and some established software vendors could lose revenue as businesses build tools internally instead of buying them.

McKinsey's August 2026 survey found that 32% of respondents said their organizations had decided against buying at least one software product or feature because agentic coding tools enabled them to develop the functionality in-house.

That is an unusually direct indication of AI replacing potential software purchases.

The effects will probably differ substantially by product category. Small internal utilities, simple dashboards and narrow workflow applications are becoming easier to recreate with coding agents.

Complex systems such as Salesforce, SAP and electronic health record platforms are much harder to replace because they contain years of business data, permissions, integrations and established processes.

Some agentic AI spending will therefore come from genuinely new budgets, while some will be redirected from software products, outsourced development and existing IT services.

Are investors overpaying for agentic AI startups?

Investors are paying extremely high prices for several leading agentic AI startups, and some valuations already require exceptional revenue growth to continue for years.

Cognition raised more than $2 billion at a $48 billion valuation in September 2026. The company subsequently announced annualized revenue exceeding $1 billion. We look at what these companies are worth, not only Cognition, in our agentic AI market report.

Using a $1 billion run rate as a reference gives a valuation of roughly 48 times annualized revenue, although the actual multiple is slightly lower because Cognition reported exceeding that threshold.

Sierra raised $950 million in May 2026 at a $15.8 billion valuation. Against Sacra's estimated $200 million ARR, that represents approximately 79 times recurring revenue.

Harvey's estimated $400 million ARR and reported $15.5 billion valuation imply roughly 39 times ARR.

These comparisons combine figures from different reporting periods, and private-market run rates are not necessarily audited or defined consistently. They are useful indications of investor expectations rather than precise like-for-like valuation multiples.

The valuations become easier to understand when we examine the growth. Cognition approximately doubled its reported annualized revenue in four months. Harvey's estimated ARR roughly doubled from the end of 2025 to August 2026. Sierra's estimated ARR increased by more than 50% between the end of 2025 and May 2026.

Investors are paying for the possibility that these companies will become major software vendors before their markets mature. At these multiples, though, even strong growth may not be enough if margins disappoint or competition puts pressure on prices.

Which parts of agentic AI have the most money to capture next?

Coding, customer support and specialized professional work currently offer the best-supported revenue opportunities, while enterprise orchestration and agent security could become larger businesses as adoption spreads.

Coding already attracted $4 billion in estimated enterprise spending in 2025, and Cognition's reported growth confirms that specialized vendors can capture substantial revenue.

Customer support provides another strong opportunity. Sierra has reached an estimated $200 million ARR, while Intercom and Salesforce have established measurable usage-based pricing.

Legal AI has a smaller potential customer base than general enterprise productivity software but unusually valuable work. Harvey's estimated $400 million ARR shows that specialist products can become sizable businesses without targeting every industry.

Healthcare has impressive operational adoption, including Microsoft's projected 100 million Dragon Copilot-documented encounters, but publicly available revenue figures remain less complete.

Infrastructure and security could benefit from the spread of multi-agent deployments. Gartner's forecast that more than 60% of organizations expect to deploy agents within two years suggests a substantial potential increase in demand, although expectations do not guarantee actual purchases.

Broad, fully autonomous business agents remain harder to rank. Their potential market is large, but evidence of reliable, profitable operation across complex organizations is still limited.

So where is the money in the agentic AI market?

The money in agentic AI is already flowing into coding, customer support, specialized professional software and the enterprise platforms that let agents work inside existing business systems. These are currently the strongest commercial opportunities, while the broader autonomous-agent vision still has much more to prove.

Coding has the clearest large spending base, with Menlo Ventures estimating $4 billion in enterprise spending in 2025 and Cognition reporting more than $1 billion in annualized revenue.

Customer support offers particularly attractive transaction economics. Sierra's estimated $200 million ARR, Intercom's outcome pricing and Salesforce's growing Agentforce business show that companies will pay directly for useful automated work.

Harvey demonstrates the potential of specialized professional markets, while Microsoft and Salesforce show how established software vendors can distribute agent capabilities at enormous scale.

Foundation-model providers are also collecting substantial revenue because agents consume inference as they perform more tasks.

The main uncertainty now concerns how much of this activity will become profitable, recurring business. Rapidly growing revenue does not resolve questions about model costs, reliability, customer retention or the sustainability of today's private-market valuations.

Our conclusion is that the agentic AI market is already commercially significant, but the money is concentrated in products that automate well-defined, valuable work. Coding and customer support have the strongest current evidence, legal AI is emerging as a major specialist business, and large enterprise platforms have the distribution advantage.

The next winners will be decided increasingly by how much useful work customers can get done for what they spend, rather than by how autonomous an agent looks in a demonstration.