Which AI agents are businesses paying for now?

Are businesses actually paying for AI agents now?

Businesses are already paying serious money for AI agents today, although the spending is concentrated in a handful of jobs where the agent can complete real work and the buyer can measure the result.

Salesforce gives us one of the cleanest numbers. In its latest quarterly results, Agentforce ARR exceeded $1.5 billion, up more than 240% year over year. Salesforce also reported 7 billion Agentic Work Units completed across Agentforce and Slack, including 3.2 billion during the quarter alone. Three quarters earlier, Salesforce was reporting $800 million of Agentforce ARR and 2.4 billion cumulative work units. Whatever definition we use for an “agent,” customers are clearly putting more money and much more workload through these products.

Microsoft shows the same shift from another angle. More than 30 million Microsoft 365 Copilot seats are now paid seats, while nearly 40 million agents have been registered in Agent 365 across tens of thousands of companies. Microsoft has also started charging some customers based on consumption rather than only seats; thousands were already paying for usage-based Cowork within weeks of launch.

The caution is that an enormous gap still separates buying AI from handing AI broad authority. UiPath said earlier this year that its agentic products were moving from pilots into production, while its overall platform ARR reached $1.90 billion. By the end of July, UiPath had 387 customers spending at least $1 million annually, up from 320 a year earlier. Much of that money still covers automation infrastructure alongside AI agents, which is how enterprise adoption usually looks in practice.

So the answer is yes. AI agents have a real enterprise market now. The interesting part is where companies have decided they are worth paying for.

Which AI agents are businesses spending the most money on today?

The strongest paid AI-agent categories today are customer service, software development, legal work and enterprise workflow automation, with sales agents and broader knowledge-work agents growing behind them.

We get a clearer picture when we look for several forms of evidence at once: standalone ARR, large paid deployments, repeated usage, consumption revenue and customers expanding after an initial purchase.

Customer service already has billion-dollar-scale platforms attaching agent pricing directly to resolutions and actions. Coding has tens of millions of users and increasingly large enterprise contracts. Harvey has taken specialist legal AI from a niche product to more than 2,400 customers and 80% of the Am Law 100. ServiceNow, UiPath, Microsoft and SAP are embedding agents into IT, HR, finance and operational workflows companies were already paying software to manage.

Sales is further along than it was a year ago, but the strongest results currently come from prospecting, qualification and follow-up rather than autonomous deal closing. Horizontal “AI coworkers” have huge distribution through Microsoft and other platforms, although the commercial evidence still mixes agent usage with broader assistant subscriptions.

That creates a fairly visible hierarchy today.

AI agent category	Best evidence that businesses pay	Where the market stands now
Customer service	Agentforce >$1.5B ARR; Zendesk reached ~$200M AI ARR	Already a large production market
Coding	GitHub Copilot at 50M users; enterprise consumption revenue growing	Large and rapidly deepening
Legal	Harvey at 2,400+ customers; 80% of Am Law 100	Strong specialist market
IT and workflow automation	ServiceNow, UiPath, Microsoft and SAP production deployments	Large, often bundled into existing platforms
Sales	HubSpot reports 80% more meetings among Prospecting Agent users	Commercially real but less mature
Finance and HR	Growing inside SAP, UiPath and Microsoft workflows	Production use with tighter controls
General knowledge work	30M+ paid Microsoft 365 Copilot seats	Huge distribution; autonomous work is newer

Why are businesses paying so much for customer-service AI agents?

Customer-service AI agents are currently the easiest agents to sell because companies can see exactly how many conversations they resolve and what each avoided human interaction is worth.

Salesforce's latest Agentforce numbers cover several products, but sales and service remain central to its commercial push. The company says bookings for premium editions built around sales, service and agentic functionality more than doubled quarter over quarter in its latest results. Microsoft also says customer service is leading its move toward consumption pricing, with usage-based credit consumption in the category rising fourfold in one quarter.

Zendesk provides an unusually useful comparison because its AI revenue is disclosed separately. The company said it entered 2026 with approximately $200 million in AI ARR and more than 20,000 customers using its AI products. Zendesk is now pushing further toward pricing based on completed resolutions, which makes commercial sense: a company can compare the cost of an automated resolution with the cost of putting the same case through a human support team.

HubSpot released fresh customer data this week that points in the same direction. Its Customer Agent is resolving more than 70% of conversations automatically among the deployments highlighted by HubSpot, and Professional and Enterprise customers using it recorded a 2.3-times higher support-conversation resolution rate than customers without it during the period HubSpot measured. Those are vendor-reported comparisons rather than controlled experiments, but the metric itself tells us what buyers care about: resolved conversations.

That is why support has moved so quickly. A customer-service agent can handle repetitive questions, check a knowledge base, look up an order, perform a permitted action and hand the complicated cases to a person. Even 50% or 70% automation can create an obvious return without requiring 100% reliability.

The buyer also knows the baseline before the agent arrives. Ticket volumes, handling times, staffing costs, escalation rates and resolution rates are already measured. Agents enter an environment where their work can be priced almost immediately.

Are businesses really paying coding agents to write software?

Businesses are paying heavily for coding agents now, and software development has moved further into genuine agentic work than almost any other white-collar profession.

Microsoft's latest numbers are much larger than they were earlier in the year. GitHub Copilot now has 50 million users, Copilot revenue accelerated more than 60% quarter over quarter after usage-based billing was introduced, and Microsoft says one in three pull requests on GitHub now involves an agent. More than 90% of the Fortune 500 use GitHub.

That trajectory has been remarkably fast. Earlier this year, Microsoft was reporting 4.7 million paid GitHub Copilot subscribers, up 75% year over year, with Siemens deploying it to 30,000 developers. One quarter later, nearly 140,000 organizations were using GitHub Copilot and enterprise subscriptions had almost tripled from the previous year. The newest figures show the market moving beyond paid autocomplete toward agents that consume more compute and work for longer periods.

OpenAI's Codex data reinforces the point. Weekly Codex usage went from more than 3 million developers to more than 4 million within two weeks earlier this year. Cisco uses Codex to reason across large interconnected repositories, Ramp uses it for code review, Notion uses it to build features and Virgin Atlantic uses it for testing and technical-debt work. OpenAI later reported that enterprise agent usage had spread far beyond engineering as well.

Coding has one huge advantage over most knowledge work: we can test a lot of the output. Code can compile or fail. Tests can pass or fail. Linters catch some mistakes. A reviewer can inspect a pull request. Production systems can roll back a bad deployment.

That verification layer makes companies much more comfortable giving an agent a long task. A developer can ask the agent to inspect a repository, change several files, run tests and return with a proposed patch. The human moves from typing every line toward assigning and reviewing work.

The economics help too. When the employee using the agent is expensive, even a modest gain can cover a substantial AI bill. Coding agents therefore have room to move from $20-or-$40-per-seat tools into products that also charge for the compute used by longer autonomous sessions.

Why has legal AI become one of the biggest agent markets?

Legal AI agents are making serious money because a relatively small number of expensive professionals can support a very large software business when the agent saves hours of work on high-value matters.

Harvey is the clearest case. At the beginning of 2024, Harvey had around 40 customers. Today it says more than 2,400 legal organizations use the platform across over 70 countries, including 80% of the Am Law 100. Its customer base includes law firms as well as corporate legal teams.

The financial progression has been just as fast. Harvey finished 2025 at roughly $190 million ARR according to CEO figures reported by Forbes. This summer, the company said it had added more than $100 million of ARR during a single quarter. Days ago, Harvey raised another $550 million at a $15.5 billion valuation. That valuation is an investor judgment rather than revenue, but the repeated fundraising would mean much less without the underlying expansion in paying customers and recurring usage.

The agent layer is already material inside the product. Harvey disclosed more than 25,000 custom agents running across customers earlier this year. Those agents can be configured for workflows such as due diligence, contract analysis, compliance and litigation rather than simply answering generic legal questions.

Usage also appears sticky. An independent study conducted by RSGI for Harvey found a 92% average monthly usage rate among the customers surveyed. We should treat commissioned research cautiously, but a figure that high is more useful than a launch announcement because it asks whether lawyers keep returning after deployment.

The broader economics explain why legal has moved so quickly. Harvey estimates that M&A due diligence alone sits inside a global process costing tens of billions of dollars a year. When lawyers charge hundreds of dollars per hour and a transaction can involve thousands of documents, shaving hours from review has a much larger dollar value than helping someone write emails a little faster.

Harvey's expansion	Earlier point	Latest available evidence
Customers	~40 at start of 2024	2,400+
Am Law 100 adoption	Early concentration among top firms	80%
Custom agents	Previously a minor part of the story	25,000+ disclosed
Valuation	$11B earlier this year	$15.5B after latest financing

Are companies paying AI agents to run IT and internal workflows?

Companies are already paying AI agents to handle IT and operational workflows, especially when those agents sit inside software the company already trusts to manage permissions, data and business processes.

This market looks different from legal or coding because the agent revenue is often bundled into broader platforms.

UiPath is a good example. The company reached $1.94 billion ARR by the end of July, with 387 customers spending at least $1 million annually. UiPath says its agentic products are moving from pilot projects into production, but we should not call that $1.94 billion “agent revenue.” Customers are buying a wider automation platform that increasingly mixes AI agents with traditional robotic process automation and workflow orchestration.

ServiceNow is taking a similar route. Its partnership with Accenture now offers more than 300 pre-built agent skills and agentic workflows designed to move projects from experiments into production. The reason ServiceNow has an advantage here is straightforward: companies already run IT requests, employee workflows, security processes and operational tickets through its platform. An agent can act inside a system with established permissions rather than starting from scratch.

Microsoft is opening the same door through Dynamics 365. Its agents can now access more than 650,000 actions across sales, finance, supply chain, HR and customer service while inheriting the existing data models, business rules, permissions and audit trails of those systems.

This is probably closer to how agent adoption will look inside large companies than the popular image of thousands of independent digital employees. Businesses already have software that controls what workers can see and change. Agents are being added inside those boundaries.

Are sales AI agents actually making businesses money?

Sales AI agents have become commercially useful, especially for prospect research, lead qualification and follow-up, but companies are still keeping humans close to the parts of selling that require judgment and negotiation.

Fresh HubSpot data gives us one of the better measurements available. Professional and Enterprise customers using its Prospecting Agent booked 80% more meetings than customers without it during the period HubSpot studied. HubSpot says the agent researches prospects, identifies the right people, monitors buying activity and prepares personalized outreach.

The comparison comes from HubSpot's own customer population, so we should avoid reading it as proof that adding the agent causes an 80% increase for every sales team. Companies choosing the agent could already be more sophisticated users. Even with that limitation, we are looking at real customers using an agent inside a live sales process and measuring the outcome in meetings rather than generated emails.

Microsoft has another useful example. Sandvik uses its Sales Qualification Agent across tens of thousands of potential customers, automating a stage that would otherwise consume sales-team time. Microsoft has also reported more than $450 million in annualized revenue from agentic LinkedIn products that automate sourcing, screening and message drafting for recruiters. Recruiting is a different market, but the workflow resembles sales: find the right person, research them, contact them and decide whether a human should take over.

The strongest use case today is fairly specific. AI can watch a large lead pool continuously, research accounts, decide who deserves attention and prepare the next action. That is much easier to automate than a complicated enterprise negotiation involving politics, pricing, objections and several decision-makers.

We see far less convincing evidence that companies are handing the entire sales cycle to AI. The money is currently gathering around the repetitive work that happens before and between human conversations.

Are companies paying AI agents for finance, HR and other back-office work?

Finance, HR and back-office AI agents are moving into production, although companies generally put more controls around these agents because mistakes can affect money, employees or compliance.

The attraction is obvious. Back-office teams spend huge amounts of time moving information between systems, reading documents, checking requests and preparing routine transactions. Much of that work sits between old-fashioned automation and human judgment, which is exactly where modern agents can help.

Microsoft now exposes hundreds of thousands of Dynamics actions to agents across finance, HR and supply-chain systems. UiPath is building agents into the same enterprise processes where its software already automates invoices, procurement, accounting and administrative work. SAP is doing the same through Joule across its business applications.

The deployment pattern matters more than the label. A finance agent might read an invoice, investigate a discrepancy and prepare an action while a rule checks the numbers and a person approves the payment. An HR agent can answer routine employee questions or prepare a transaction while sensitive cases go to a specialist. Procurement agents can review documents and flag risks before a buyer commits the company.

We should expect this category to become very large because the underlying work is enormous. We should also expect slower autonomy than in coding or customer support. A slightly wrong support answer can be escalated. A slightly wrong payment or payroll action can create a much bigger problem.

Are voice AI agents becoming a real business market?

Voice AI agents are becoming a paid business category now because call-center work has a clear cost per conversation, although voice remains harder to automate reliably than text support.

The commercial logic is compelling. A contact center knows how many calls arrive, how long they last and how much human coverage costs. Once a voice agent can authenticate someone, understand the request, access the right system and perform an action, the company can calculate the savings almost immediately.

The major service platforms have therefore started folding voice into their agent businesses. Salesforce, Microsoft and Zendesk increasingly treat voice as another interface into the same customer-service workflows rather than a separate chatbot product. Salesforce's own usage pricing assigns more credits to voice actions than standard text actions, reflecting the additional compute and complexity.

Voice still has more ways to fail. People interrupt each other. Accents vary. Customers call when they are angry or confused. Regulated industries may require exact wording. Authentication becomes harder when the agent cannot simply rely on a logged-in web session.

That naturally pushes adoption toward bounded jobs such as scheduling, status checks, simple qualification and common support requests. We are seeing a real market, but text-based customer-service agents remain further ahead today.

Are businesses paying for general-purpose AI coworkers?

Businesses are spending enormous amounts on general workplace AI, while the newer autonomous coworker products are only beginning to separate themselves commercially from the broader Copilot market.

Microsoft now has more than 30 million paid Microsoft 365 Copilot seats. NHS England is rolling Copilot out to 505,000 staff, KPMG is expanding it across more than 276,000 professionals, HSBC committed to 200,000 seats, and EY purchased Microsoft's broader E7 package for 400,000 employees.

Those numbers prove that companies will pay to put AI across almost an entire workforce. They do not tell us that 30 million people have autonomous agents doing their jobs. Microsoft 365 Copilot includes chat, search, writing, meetings, document work and increasingly agentic features inside the same commercial relationship.

What has changed lately is the product behavior. Microsoft now describes Cowork as software that can complete multi-step work using company data, and its newer Autopilots can run longer autonomous tasks. Thousands of customers were already paying usage charges for Cowork shortly after usage billing launched.

OpenAI's enterprise data points in the same direction. By June, Codex accounted for 64% of the combined Codex and ChatGPT output tokens generated by enterprise customers in OpenAI's dataset. Since February, weekly enterprise Codex users had risen 108-fold in legal, 41-fold in sales, 41-fold in recruiting and 26-fold in marketing, compared with fivefold growth in engineering.

Those percentages tell us that agentic work is spreading quickly beyond developers. They still measure usage rather than standalone agent revenue, so the commercial conclusion should stay narrower: companies have clearly bought horizontal AI access, and autonomous task execution is starting to consume a larger part of that budget.

Why are businesses choosing specialized AI agents instead of one super-agent?

Businesses are currently building fleets of specialized AI agents because narrow permissions, clear context and measurable jobs make agents much easier to trust.

Microsoft says nearly 40 million agents are already registered through Agent 365 across tens of thousands of organizations. Harvey disclosed more than 25,000 custom legal agents months ago. Salesforce is processing billions of Agentic Work Units across its customer base. The interesting pattern is the multiplication of agents inside each company rather than the arrival of one universal company-wide agent.

That fits the way companies already operate.

A support agent needs customer history and permission to issue certain refunds. A coding agent needs repository access and perhaps a sandbox, while production access stays restricted. A recruiting agent needs candidate and job data. A legal agent works with privileged documents. A finance agent may inspect transactions but require approval before money moves.

Giving one general agent all of those permissions would create a huge security and governance problem. Splitting the work also makes performance easier to measure. We can tell whether the support agent resolved the case, whether the coding agent passed its tests and whether the recruiting agent produced qualified candidates.

Agent specialization reflects practical enterprise constraints more than a shortage of model intelligence. The model can be the same underneath several agents. What changes is the context, tools, permissions and definition of success.

How autonomous are the AI agents businesses are paying for?

Most paid enterprise AI agents today have meaningful autonomy inside a bounded workflow, while humans or deterministic software still control the points where mistakes become expensive.

Customer-service agents can resolve many conversations from beginning to end and escalate the uncertain ones. Coding agents can inspect a repository, edit files and run tests before a developer reviews the proposed change. Legal agents can work through large document sets while a lawyer remains responsible for the advice. Finance agents can investigate a discrepancy while payment rules and approvals stay in place.

That mixed model shows up in the platforms businesses are buying. UiPath deliberately combines reasoning agents with conventional automation. Microsoft gives agents access to existing Dynamics permissions, rules and audit trails. ServiceNow is putting agent workflows inside a platform already designed to control enterprise processes.

The distinction becomes important when we hear claims about “AI employees.” A company does not need an agent to imitate a human worker from morning to evening to get a return. If an agent independently finishes thousands of narrow tasks that previously required ten minutes of employee time, that can already justify a substantial budget.

Autonomy is arriving task by task. The highest-confidence tasks get delegated first, and the boundary expands as companies learn where the agent performs reliably.

Why are AI-agent companies starting to charge for completed work?

AI-agent pricing is moving toward usage and outcomes because charging only for human seats becomes awkward when the software itself starts doing more of the work.

This change is now visible across several large vendors. Salesforce uses Flex Credits tied to agent actions and also offers pricing structures linked to conversations and resolutions. Microsoft has introduced consumption billing across parts of its agent stack, including Cowork, Dynamics and GitHub. Zendesk has pushed toward charging around automated resolutions. GitHub added usage-based billing as coding agents began consuming much more compute than autocomplete.

The shift tells us something important about what customers are buying.

Traditional SaaS economics assumed that more employees using the software meant more value and therefore more seats. An agent can produce the opposite outcome. If one AI system allows five people to handle work that previously required seven, the vendor cannot rely forever on charging for seven human users.

Completed work offers a more natural unit. A support platform can charge for a resolved conversation. A coding platform can charge for longer agent runs. A workflow platform can meter actions. A company buying the product can then compare the bill with the human or operational cost that disappeared.

We should not assume every agent business will end up with pure outcome pricing. Some jobs are difficult to define cleanly, and enterprise contracts will continue combining licenses, capacity and consumption. Still, the direction is unusually consistent across vendors.

The software market is slowly learning how to price labor performed by software.

So which AI agents are businesses paying for now?

Businesses are paying most confidently for customer-service agents, coding agents, legal agents and agents embedded in IT or operational workflows; sales, back-office, voice and general knowledge-work agents are growing quickly but have less uniform proof of autonomous production use.

Customer service currently has the cleanest economics. Salesforce has pushed Agentforce ARR above $1.5 billion, Zendesk entered the year with roughly $200 million of AI ARR, Microsoft says customer-service consumption is growing rapidly, and HubSpot now reports more than 70% automatic resolution for Customer Agent in the deployments it highlights. The unit of value is easy to understand: one more customer problem resolved without taking a human agent's time.

Coding has arguably moved furthest into genuine delegation. GitHub Copilot now reaches 50 million users, coding-agent consumption is producing incremental revenue, and one in three GitHub pull requests involves an agent. Codex is being deployed across companies such as Cisco, Ramp, Notion, Virgin Atlantic and Samsung. Software gives agents something most office work cannot: fast, relatively objective feedback about whether the work functions.

Legal proves that a much smaller professional market can still create a huge agent business. Harvey now serves more than 2,400 organizations, reaches 80% of the Am Law 100 and recently raised money at a $15.5 billion valuation after adding more than $100 million of ARR in a single quarter. Its customers were already running more than 25,000 custom legal agents earlier this year.

The next layer is enterprise operations. ServiceNow, UiPath, Microsoft and SAP are putting agents inside IT, finance, HR, supply-chain and other systems where companies already have data, permissions and workflows. Sales agents are showing measurable results in prospecting and qualification. Voice agents are entering contact centers. General workplace agents now have distribution through tens of millions of paid Copilot seats, and the amount of work being delegated to them is rising quickly.

Across all of those categories, the same commercial pattern keeps appearing. Companies spend fastest when an agent has a well-defined job, enough context to do it, permission to take the required actions, a clear way to check the result and an obvious economic cost to compare against.

The businesses paying for AI agents today are mostly buying support resolutions, code changes, legal workflows, qualified leads and completed enterprise processes.

For now, the biggest agent businesses are being built around work we can count.