What are the main business models in agentic AI?

What do agentic AI companies actually charge for today?

Agentic AI companies currently make money through a wide mix of subscriptions, usage fees, completed outcomes, deployment services, infrastructure consumption and, increasingly, transaction fees.

The easiest way to understand the market is to look at what customers are actually being billed for. Salesforce is a good example because Agentforce now supports several models at once. Customers can pay $500 for 100,000 Flex Credits, with a standard action costing about $0.10; $2 per conversation; $5 per user per month for one employee-agent license; $125 per user for flat-fee access to some agent products; or $2 for a successful Help Agent resolution.

Intercom has gone further toward value-based pricing. Fin currently costs $0.99 for a successful resolution, procedure handoff or lead disqualification, while a qualified sales lead costs $9.99. Cursor sits closer to traditional SaaS but increasingly mixes seats with consumption. Its individual plans run from $20 to $200 a month, while team seats start at $40 and include defined pools of agent usage.

Those differences tell us more than the labels. When an agent helps one employee, a subscription still makes sense. When the agent runs thousands of unpredictable actions, vendors prefer credits or metered usage. When one completed task has an obvious business value, companies can charge for the outcome. And when the agent directly participates in a purchase, the vendor can eventually take a cut of the transaction.

There is already a fairly clear map of agentic AI business models, even though the market has not settled on one universal pricing unit.

Business model	What customers pay for	Where it fits best	Examples today
Subscription + usage	User access plus extra AI work	Coding, professional agents	Cursor
Consumption	Actions, credits, tokens or compute	General agent platforms	Salesforce, Microsoft
Outcome	Successful completed task	Support, lead qualification	Intercom Fin, Salesforce Help Agent, Sierra
Software + services	Platform plus deployment work	Complex enterprise agents	Sierra, vertical AI companies
Infrastructure usage	Tracing, runtime, storage, evaluation	Agent developers	LangSmith and similar platforms
Transaction fee	Completed purchase or booking	Agentic commerce	ChatGPT Instant Checkout
Marketplace commission	Third-party agent components	Agent ecosystems	Salesforce AgentExchange

Is agentic AI already a real business, or mostly pilots?

Agentic AI is already generating billions of dollars in contracted and recurring business, although the average company is still much earlier in deployment than the market leaders.

The commercial numbers are difficult to dismiss. Salesforce's latest reported quarter put Agentforce ARR at $1.2 billion, up 205% year over year. ServiceNow reported in July that its broader AI business had crossed $1 billion in annual contract value, while agentic deployments had increased ninefold in nine months.

Startups are reaching unusual scale too. Bloomberg reported that Cursor passed a $2 billion annualized revenue rate in February after doubling in only three months. A later Dealroom report put the figure at roughly $4 billion by early June, although Cursor is private and that later number has not been publicly confirmed by the company. Sierra was estimated at around $200 million in ARR by May after being near $130 million at the end of 2025.

The enterprise adoption picture is much less extreme. Gartner's 2026 CIO research found that 17% of surveyed organizations had already deployed AI agents, with another 42% planning deployment within a year. Other enterprise surveys still show a much larger group experimenting than running agents broadly in production.

A few categories have already found very strong willingness to pay, especially coding, customer service and enterprise workflow automation. Most corporate functions are still figuring out what should be delegated to an agent, what needs human approval and how to control cost and risk.

So yes, agentic AI is a real software market today. Autonomous agents running most businesses is still a very different claim.

Are subscriptions still working for AI agents?

Subscriptions are still working well for AI agents, especially when one employee remains the obvious user, but flat unlimited seats become harder to justify as agents do more work on their own.

Cursor shows the direction clearly. Its current individual plans cost $20 for Pro, $60 for Pro Plus and $200 for Ultra. Team pricing starts at $40 per user per month, with a $120 Premium seat for heavy agent users.

But the subscription no longer buys an economically identical experience for everyone. Cursor separates its own models from third-party model usage and gives different plans different consumption pools. Its documentation says limited agent users often stay near $20 of monthly third-party usage, daily agent users commonly consume $60 to $100, and power users running several agents or automations can exceed $200.

Cursor adjusted the model again in June. Team seats now receive one pool for Cursor's first-party models and another for third-party APIs. The company also introduced the $120 Premium seat after finding that a small group of heavy users generated a disproportionate share of spending.

A developer occasionally asking an agent to explain a function and another developer running cloud agents for hours simply should not cost the vendor the same amount.

Subscriptions are therefore more likely to survive agentic AI than disappear. They give customers predictable procurement and vendors a recurring revenue floor. Heavy autonomous work, though, increasingly sits behind usage limits, premium tiers or overage charges.

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

Is usage-based pricing becoming the default for AI agent platforms?

Usage-based pricing is becoming the default for general-purpose agent platforms because the amount of work an AI agent performs varies too much for one flat price.

Salesforce charges in Flex Credits. A standard Agentforce action consumes 20 credits, which works out to about $0.10 at list pricing. Microsoft uses Copilot Credits: a classic answer can consume one credit, a generative answer two, an agent action five and tenant-graph grounding ten. Several types of activity can also happen inside the same interaction.

Agent infrastructure companies use a similar structure. LangSmith combines paid access with metered traces, storage, deployment compute and agent workloads. Model providers meter tokens and additional tools. Even Cursor, which still looks like a subscription product on the surface, now tracks third-party model use at underlying API prices.

The reason is variable work. One agent request might answer a simple question. Another might search files, retrieve data, call three tools, inspect the result, retry a failed action and then write back to another system. Billing both as one identical request stops making sense pretty quickly.

The weakness is predictability. Customers do not love giving an autonomous system permission to generate an unknown bill, which is why usage pricing increasingly comes wrapped inside committed spending, included allowances, alerts and hard limits.

The emerging model looks a lot like cloud computing: variable consumption, but with budgets and controls around it.

Can outcome-based pricing really work for AI agents?

Outcome-based pricing already works for AI agents in narrow workflows where everyone can agree on what success means, and customer support has become the clearest proof.

Intercom's current Fin pricing makes the model concrete. A successful support resolution costs $0.99. The company also charges $0.99 when Fin successfully completes certain handoffs or disqualifies an unsuitable lead, while a qualified lead costs $9.99.

Salesforce now uses the same idea for its Help Agent. A customer pays $2 when the agent resolves an interaction under Salesforce's defined criteria; an unresolved interaction is not billed as a resolution.

Sierra has built much of its commercial positioning around a similar principle. Instead of selling generic access to a chatbot, it negotiates around the business outcomes its customer-facing agents produce. The company's rapid growth suggests enterprise buyers are comfortable with this model when they can clearly see what they are buying.

Customer service works especially well because the denominator already exists. Companies know how many cases they receive, what percentage humans resolve, how much support labor costs and whether customers escalate after an AI response. That makes it possible to compare the price of an AI resolution with the existing cost of handling the same case.

The model gets shakier once attribution becomes fuzzy. If an AI agent writes part of a software release, contributes research to a lawsuit or touches a sales opportunity six months before closing, deciding what counts as a billable outcome becomes much harder.

Outcome pricing should become a major agentic AI model, but mostly in workflows with frequent, measurable and attributable results.

Agent task	Is success easy to measure?	Fit for outcome pricing
Customer support resolution	Very easy	Excellent
Lead qualification	Easy	Strong
Appointment booking	Easy	Strong
Debt collection	Easy	Strong
Code generation	Moderate	Mixed
Legal drafting	Moderate	Mixed
Research	Difficult	Weak
Long enterprise sales cycle	Difficult	Weak

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

Can AI agents charge against labor budgets instead of software budgets?

AI agents can increasingly tap labor budgets when they remove a measurable chunk of human work, which gives some agent companies much more pricing room than ordinary SaaS.

The difference can be huge. A normal productivity tool may compete with another $20 or $50 software seat. An agent that completes work previously costing a company $10, $50 or $500 creates a different pricing ceiling.

Customer support makes the calculation obvious. If a human-handled case costs several dollars once wages, management and overhead are included, a reliable automated resolution priced near $1 can still leave the buyer with substantial savings.

High-value professional work pushes that logic much further. Legal AI companies are selling into an industry where professional time can cost hundreds of dollars per hour. Harvey has reportedly reached hundreds of millions of dollars in recurring revenue, while Legora has also moved past $100 million ARR. Neither company needs to eliminate lawyers to justify a large software bill. Saving part of an associate's day across hundreds or thousands of expensive professionals is enough.

The phrase "digital labor" sometimes gets ahead of reality, though. Most agents today are taking over pieces of jobs rather than whole jobs. A support agent may handle common tickets and escalate the rest. A coding agent may implement or review code while a developer still owns the final result. A legal agent may prepare drafts while lawyers remain responsible for the matter.

For now, the strongest labor-based sales pitch is specific: the agent removes this piece of work at a lower cost. That is much easier to prove than claiming an AI system can replace an entire employee.

Are vertical AI agents quietly becoming services businesses?

Many vertical AI agent companies currently look like software businesses with a meaningful services layer because getting an agent into a real enterprise workflow still takes considerable human work.

Sierra openly pairs its platform with teams that help customers build, deploy and improve agents. CrewAI's enterprise offering has included structured onboarding and access to forward-deployed engineering. Legal, financial and healthcare AI vendors often need to connect proprietary data, adapt workflows, set permissions and evaluate outputs before customers will trust agents with important work.

That implementation burden is not surprising. A generic chatbot can be turned on in minutes. An agent that refunds customers, handles regulated documents or updates internal systems needs to understand much more about how one company operates.

The real question comes after the first deployment. Services become dangerous when every new customer requires the same large amount of custom engineering forever. Revenue can grow quickly while margins stay closer to consulting than software.

The better companies are trying to turn repeated deployment work into reusable product. Every integration becomes a connector. Every recurring workflow becomes a template. Manual evaluations can become automated tests. Customer-specific permission problems can improve the platform's governance layer.

The services component is not automatically a weakness. Right now it can be the fastest way to get agents into production. The test is whether implementation effort per customer falls as the company matures.

Why are coding agents making money so much faster than most AI agents?

Coding agents are making money unusually fast because software engineering gives AI almost everything it needs: expensive labor, digital inputs, machine-readable feedback and customers who already expect automation.

Cursor is the clearest commercial example. Bloomberg put its annualized revenue above $2 billion in February, double the level reported only three months earlier. By early June, Dealroom reported a figure around $4 billion, citing a person familiar with the company. Even if we stick to the lower publicly reported milestone, very few software companies have ever reached this scale so quickly.

The growth is increasingly enterprise-driven. Around 60% of Cursor's revenue was already coming from corporate customers when Bloomberg reported the $2 billion run rate. Later reporting suggested the enterprise share had risen further.

Coding is unusually agent-friendly because an AI system can inspect the repository, edit files, run tests, read compiler errors and try again. The environment constantly tells the agent whether its work is moving in the right direction. Many other knowledge jobs simply do not offer feedback that clean.

The commercial model is evolving alongside the product. Cursor moved Bugbot from a $40-per-seat subscription toward usage billing this year. The average code-review run now costs roughly $1 to $1.50 depending on the size and complexity of the pull request. Customers can choose deeper review effort and pay for more computation.

That transition is revealing. As coding agents move from helping developers type toward independently reviewing, testing and changing software, charging purely for access becomes less natural. More revenue can attach directly to the amount of work the agent performs.

Coding agents are giving us an early look at what mature agent economics may look like elsewhere.

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

Do model providers or AI agent apps keep more of the money?

AI agent applications can capture far more revenue than the underlying model call costs, but only if they control inference spending well enough to keep the model provider from eating too much of the margin.

Cursor gives us unusually clear evidence. TechCrunch reported that the company had previously operated with negative gross margins. The introduction of its own Composer model, along with greater use of cheaper alternatives, helped push parts of the business into positive gross margin.

Cursor's recent pricing changes reinforce the point. The company now separates first-party-model capacity from third-party API consumption and gives customers much more included usage on models it can operate more cheaply. That lets Cursor offer heavy agent usage without paying frontier-model prices on every action.

Harvey is dealing with the same problem at a different scale. A Wall Street Journal report said Harvey's monthly token consumption had jumped from roughly 1 trillion to 14.5 trillion within six months as the legal platform became more persistent and agentic. Harvey is responding partly by adding lower-cost open-weight models alongside Anthropic, Google and OpenAI models.

That 14.5-fold increase is a useful warning. Better agent products can trigger vastly more inference than ordinary chat products.

Application companies therefore need something customers value independently of the model: workflow integration, proprietary context, memory, permissions, domain-specific evaluation and distribution. Then they can choose whichever model gives them the best price-performance ratio for each task.

Thin wrappers have weak economics. Agent companies that own the workflow and intelligently route between models have much more room to build durable margins.

Can agent infrastructure become a big standalone business?

Agent infrastructure can become a substantial standalone business because companies running autonomous software need ways to trace, test, govern and control what those agents are doing.

LangSmith shows how this category already monetizes. Customers pay for the development platform and then consume additional tracing, storage, deployment and compute resources as their applications scale.

The demand becomes easier to see once agents start taking actions. Traditional application monitoring tells a company whether a service crashed or returned an error. An agent can run perfectly from a technical standpoint and still choose the wrong tool, expose sensitive data, spend far too much money or confidently complete the wrong task.

Governance is becoming a category of its own for the same reason. Gartner's 2026 research found only 17% of surveyed organizations had deployed agents so far, yet 42% planned to do so within a year. ServiceNow is already positioning its AI Control Tower around managing agents across the enterprise rather than simply selling individual bots.

The infrastructure opportunity spans runtime, evaluation, observability, identity, permissions, security and cost control. Companies will have more reason to pay for these layers once one organization has hundreds or thousands of agents rather than ten experimental ones.

Some parts will commoditize quickly. Model providers and cloud platforms can add basic tracing, and open protocols reduce the value of proprietary plumbing. A company whose entire product is a convenient way to connect two APIs may struggle.

The bigger businesses should form around problems that remain painful even after standards improve: knowing which agent acted, what data it saw, why it made a decision, what the action cost and whether the result was actually correct.

Do Salesforce, Microsoft and ServiceNow have an unfair advantage in agentic AI?

Salesforce, Microsoft and ServiceNow have a huge advantage in enterprise agentic AI because they already own much of the data, permissions and workflow context agents need. Startups still win when they reinvent the workflow itself.

Salesforce's latest figures make the distribution advantage obvious. Agentforce reached $1.2 billion in ARR, up 205% year over year, while more than half of Agentforce and Data 360 bookings came from existing Salesforce customers. Agentforce and Slack had also delivered 3.8 billion Agentic Work Units to date, up 111% quarter over quarter.

ServiceNow has the same structural benefit. Its AI products crossed $1 billion of annual contract value in the latest reported quarter, and agentic deployments grew ninefold in nine months. The platform already sits inside IT operations, employee workflows, security and other processes where an agent can immediately do useful work.

Microsoft can take the same route through Microsoft 365, Teams, SharePoint, Azure and Power Platform. In Copilot Studio, some agent activity by Microsoft 365 Copilot users is included rather than billed as a separate usage event, which gives Microsoft another way to push agents through an existing installed base.

But distribution is not enough when the product changes the job itself. Cursor became a multibillion-dollar business despite GitHub and Microsoft owning developer distribution. Sierra built a large customer-service business while competing with Salesforce, ServiceNow and established contact-center vendors. Harvey broke into major law firms even though those customers already had Microsoft and legal-information platforms.

The dividing line is getting clearer. Incumbents are strongest when the customer wants an agent inside software it already uses. Startups have more room when customers are willing to replace the old workflow with something built around AI from the start.

Data deepens the incumbent advantage, but integrations alone do not create an unbeatable moat. The stickier asset is accumulated workflow context: permissions, organizational memory, past decisions, evaluations and patterns showing how a specific company actually wants work done.

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

Can AI agent marketplaces become the next app stores?

AI agent marketplaces could become large businesses, but marketplace commissions remain much less proven today than selling agents, infrastructure or outcomes directly.

Salesforce's AgentExchange gives us an early version of the model. Companies can discover third-party actions, templates and agent components that connect Agentforce to products from outside vendors. Salesforce already understands this business through AppExchange, where developers have built commercial products around the Salesforce ecosystem for years.

Agents could make the market more granular. Instead of buying a whole application, a company might buy one recruiting workflow, one contract-review skill, one identity check or one specialized data connector.

That creates a natural commission model. The platform helps developers reach customers, handles discovery and distribution, and takes a percentage of paid sales. Salesforce already uses revenue-sharing arrangements for parts of its broader partner ecosystem.

But an agent marketplace becomes a serious business only when customers repeatedly buy through it and developers earn enough money to keep building for it. We do not yet have evidence of agent-component transaction volume remotely comparable with mature app stores.

Open standards also limit how much control one marketplace can capture. MCP, Agent2Agent and other protocols make it easier for capabilities to travel between platforms. That benefits developers and customers while making it harder for one gatekeeper to impose a large permanent tax.

For now, marketplaces belong in the list of real agentic AI business models, but they sit well behind subscriptions, usage and outcome pricing in commercial maturity.

Will agentic commerce make money from transaction fees?

Transaction fees are starting to become a real agentic AI business model because shopping agents can now participate directly in purchases, giving platforms a clean event to monetize.

OpenAI already charges merchants a fee on purchases completed through ChatGPT's Instant Checkout. The exact public wording is simply a small merchant fee, while Etsy has described its payment to OpenAI as performance-based and similar to an affiliate channel. An IMF review of early agentic-payment experiments cited a 4% fee for autonomous agent-led conversions.

That structure is economically interesting because the platform only gets paid when money changes hands. A $100 purchase has an observable value in a way that "AI productivity" does not.

Google is building parallel infrastructure. Its Universal Commerce Protocol now provides merchants with a standard way to connect checkout directly to AI surfaces such as Search and Gemini. Google's current implementation guides support native checkout, account linking and transaction updates, while its Agent Payments Protocol handles authorization for agent-led payments.

The infrastructure is moving quickly. Google's UCP documentation has continued to receive new implementation updates, and the system is already available to selected merchants in several markets.

What remains unclear is who eventually keeps the economics. AI platforms want to become the discovery and transaction interface. Merchants want to preserve their customer relationship. Shopify, Stripe, card networks and other payment companies already sit in the transaction. Google explicitly keeps merchants as seller of record, while OpenAI's commerce design also leaves fulfillment and customer service with the merchant.

Agentic commerce is becoming technically real before its final profit pool has been decided.

The upside is still large. If AI agents become an important place where people discover and buy products, transaction fees could eventually rival subscription revenue for some consumer agent platforms. Today, however, this model remains far smaller and less proven than enterprise agent monetization.

Do AI agents have a gross-margin problem?

AI agents currently have a genuine gross-margin problem because the best products often become more expensive to run as users delegate more work to them.

Traditional SaaS benefits from very low incremental computing costs. Agentic products can behave differently. An agent may reason for minutes, search multiple sources, run tools, call expensive models repeatedly and keep working after the user has moved on.

Cursor's economics show what that can do. TechCrunch reported that the company had been operating with negative gross margins before greater use of its own Composer model and cheaper external models improved the picture. Large enterprise accounts had reportedly become gross-margin positive while some individual accounts were still losing money.

Cursor's latest product decisions look like a direct response. Teams now receive much larger allowances for first-party models, while third-party API consumption sits in a separate pool. Bugbot moved to usage billing. Heavy agent users can be placed on a $120 Premium seat instead of costing the same as a light user.

Harvey's recent 14.5-trillion-token monthly consumption gives us another useful scale marker. Six months earlier it had been around 1 trillion. Agent usage can therefore grow far faster than seat counts when products become persistent and autonomous.

Inference efficiency is becoming part of the business model itself. Agent companies are routing tasks to cheaper models, building smaller proprietary models, caching results, limiting unnecessary reasoning and separating expensive model usage from cheaper in-house capacity.

Outcome pricing can make this especially attractive if the vendor gets the equation right. A company might charge $1 for a completed task that initially costs $0.30 to produce. If model and orchestration improvements cut the cost to $0.05 while the value of the outcome remains unchanged, the margin expands sharply.

The reverse is ugly. If agents keep using more compute while prices stay fixed, a fast-growing agent company can end up with impressive revenue and mediocre economics.

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

So which agentic AI business models are actually working?

The main agentic AI business models working today are hybrid subscriptions, usage-based pricing and outcome pricing, with software-plus-services and agent infrastructure also producing real businesses. Marketplace commissions and transaction fees are credible but still much earlier.

Subscription plus usage is probably the strongest general-purpose model right now. Cursor shows why. Customers get familiar per-user pricing, while heavy agent activity moves into larger usage pools, premium seats or overage spending. The vendor gets predictable recurring revenue without promising unlimited expensive inference.

Consumption pricing works best for platforms. Salesforce, Microsoft and agent infrastructure providers can charge for actions, credits or compute because their customers are building many different workflows and the cost of each workflow varies.

Outcome pricing has the most interesting long-term economics. Intercom's $0.99 support outcomes, Salesforce's $2 Help Agent resolutions and Sierra's enterprise contracts show that customers will pay for completed work when success can be defined cleanly. In these cases, agent software starts competing with the cost of labor rather than another SaaS seat.

Software-plus-services remains important because many enterprise agents still need considerable help to deploy. That can support large contracts today, although the better companies will steadily turn repeated implementation work into reusable software.

Infrastructure is becoming its own layer as companies need tracing, evaluation, governance, identity, security and cost control for growing populations of agents.

Marketplace and transaction models sit further out. Salesforce is already building an ecosystem around agent components, while OpenAI and Google are creating real commerce rails. Those models could eventually become huge because they monetize third-party economic activity, but today's revenue evidence is much stronger in coding, customer service and enterprise workflows.

The broader picture is fairly clear. Agentic AI is producing several business models because agents create value in fundamentally different ways. Some help one employee, some consume variable computing resources, some finish measurable tasks and some participate directly in transactions.

The companies with the best economics will be the ones that charge for a unit customers immediately understand while making the cost of producing that unit fall over time.

Business model	Commercial maturity today	Best use case	Our judgment
Subscription + usage	High	Professional and employee agents	Strongest general model
Usage / credits	High	Agent platforms and infrastructure	Already becoming standard
Outcome pricing	High in narrow workflows	Support, qualification, collections	Most important new model
Software + services	High	Complex enterprise deployments	Necessary today, margins depend on productization
Agent infrastructure	Growing quickly	Observability, governance, runtime	Likely durable
Marketplace commission	Early	Agent skills and integrations	Real opportunity, little proof at scale yet
Transaction fee	Early	Shopping, booking, payments	Large upside, economics still unsettled