AI Agents: what is getting real adoption now?

Are AI agents actually in production now, or are most companies still testing them?

AI agents are already in real production today, although scaled adoption still belongs to a minority of companies.

McKinsey's latest global AI survey gives us one of the cleanest broad measures. About two in ten respondents said their organizations were scaling AI agents across the company. Among businesses with more than $1 billion in annual revenue, 40% said agents had reached the scaling phase in at least one function, up from 27% the previous year. Smaller companies barely moved, staying around 22%.

That is a serious jump among large enterprises in one year. It also explains why the market can feel much further along if you work around major technology companies than if you look at the average business.

Cisco found an even bigger gap between interest and production when it surveyed major enterprise customers earlier in 2026. Some 85% were experimenting with agents in some form, while only 5% reported production deployment under Cisco's stricter definition.

The market has clearly moved beyond pure experimentation. Real production adoption exists at meaningful scale now, but ordinary companies have not suddenly turned into “agentic enterprises.”

Why do AI agent adoption surveys say 5%, 20% and 60% at the same time?

AI agent adoption numbers clash because researchers are calling very different levels of usage “adoption.”

Cisco's 5% figure counts broad production deployment among its surveyed enterprise customers. McKinsey gets roughly 20% when it asks whether organizations are scaling agents. Salesforce gets 66% when it looks specifically at customer-service organizations and asks whether they use AI service agents. All three numbers can be true at once.

The denominator changes too. Customer support is one of the easiest places to deploy an agent, so a survey of support teams should produce much higher adoption than a survey covering every department of every business. Surveys that deliberately interview AI leaders will naturally look more advanced again.

So a statement such as “60% of companies already use agents” is almost meaningless until we know what “use” actually means: testing one, giving employees access, running one in production or handing it a large share of a real workflow.

A survey says…	It could actually mean…
“We use agents”	At least one agent exists somewhere
“Agents are deployed”	The agent touches live work
“Agents are in production”	Real users depend on the agent
“Agents are scaling”	Usage has expanded beyond a small pilot
“Agents resolve 50% of cases”	Half of a real workload has moved to AI

Which AI agents are actually getting adopted fastest right now?

Right now, coding and customer-service agents have the clearest real adoption, with IT support next and security, sales and finance forming a second group.

We reached that ranking by giving more weight to completed work than to licenses, agent counts or vendor revenue. Coding has widespread repeated usage plus production outputs we can inspect. Customer service gives us unusually clean resolution numbers. IT support is starting to produce similarly large autonomous-resolution rates.

Security is already impressive in managed environments where permissions can be tightly controlled. Sales agents have generated real pipeline and closed revenue, especially from leads that humans would otherwise ignore. Finance deployments are saving substantial amounts of manual work, although they still concentrate on particular processes.

Legal and broader office work are moving fast these days, but human review remains more important there. Consumer browser agents are available to real users now, yet evidence that ordinary consumers routinely delegate web errands remains thin. Cross-functional fleets of agents are earlier again.

AI agent category	Real adoption now	What agents are actually doing
Software engineering	Very high	Writing code, fixing issues, testing, reviewing and opening pull requests
Customer service	Very high	Resolving conversations and carrying out account actions
IT support	High	Resolving tickets, provisioning access and handling routine incidents
Cybersecurity	Medium-high	Investigating alerts and taking approved response actions
Sales	Medium	Following up, qualifying and nurturing leads
Finance	Medium	Processing documents, accruals, reconciliations and workflow steps
Legal and knowledge work	Medium, growing fast	Research, drafting, analysis and controlled actions
Consumer browser agents	Low	Booking, forms, shopping and other web errands
Cross-functional multi-agent systems	Low	Coordinating several specialized agents across a workflow

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

Are coding agents really doing serious software work now?

Yes—coding agents are already doing serious software work today, with human review still covering most production changes.

JetBrains surveyed more than 15,000 professional developers between May and July 2026. Ninety percent said they were using some form of AI coding agent at work at least weekly, and 68% were using one every day. At that level, coding agents have clearly escaped the early-adopter niche.

The more interesting evidence comes from actual repositories. Microsoft's .NET team published ten months of data from GitHub's Copilot Coding Agent working on dotnet/runtime, one of Microsoft's most demanding open-source codebases. The agent produced 878 pull requests, equal to 14% of all pull requests during the period. Humans merged 535 of them. Only three of those merged agent PRs were later reverted, a 0.6% revert rate.

The agent was especially useful when engineers could define a clear task and let tests, CI and human reviewers decide whether the output deserved to ship. Larger architectural decisions still leaned heavily toward humans. That is already a pretty big job compared with autocomplete.

GitHub sees the same shift across thousands of enterprise accounts. Developers who had moved from its “code first” stage into an “agent first” stage averaged 78% more merged pull requests, while developers in its most agent-heavy cohort averaged 151% more than the code-first group. GitHub is measuring correlation here, so we cannot credit agents for every extra pull request, but 93.3% of accounts showed higher average throughput when comparing the first and most advanced stages.

The economic effect has started to leak outside engineering teams too. McKinsey found that 32% of respondents had already skipped buying at least one software product or feature because agentic coding tools let their company build it internally. Coding agents are now influencing build-versus-buy decisions, which is a much stronger form of adoption than developers saying they enjoy the tool.

Are customer-service agents actually resolving problems without humans now?

Yes—customer-service agents are currently closing huge volumes of real customer cases without a human taking over.

Salesforce's 2026 survey of 3,075 service professionals found AI-agent usage inside customer-service organizations rising from 39% to 66% in one year. Among organizations already using them, 70% said they saw measurable value within 60 days.

The workload figures are even stronger. Intercom says Fin is used by almost 8,000 customers, resolves close to two million customer queries every week and averages a 76% resolution rate. A resolution means the conversation was completed without a human support teammate needing to step in.

Those agents are also beginning to do more than search a help center. Engine, the business-travel platform, receives more than 800,000 customer-service requests per year. Its Agentforce agent now fully resolves 50% of incoming chat cases, including requests such as reservation cancellations that require the system to retrieve booking data and actually change something. Engine reports a 15% reduction in average handle time and a 16% increase in chat customer satisfaction.

Support has become such a strong agent market partly because the economics are obvious. There are lots of repetitive requests, the knowledge can usually be centralized, actions can be restricted to approved systems, and difficult cases can move straight to a person. An agent does not need to become a brilliant general-purpose employee to remove a very large amount of work from the queue.

Are AI agents already replacing Level 1 IT support?

AI agents are already replacing a large share of Level 1 IT work in the strongest deployments we found.

ServiceNow says agents currently handle 90% of employee IT-support requests inside its own 29,000-plus-person organization, with 96% resolution efficiency. The company says that allowed it to move 85% of the humans previously assigned to the service desk into other work. The remaining people deal with difficult cases, VIP support and oversight of the agents.

ServiceNow is an unusually favorable environment because the company owns the platform and has every reason to make its internal deployment excellent. A newer Leidos project shows where another large company thinks the same model can go. Leidos is deploying ServiceNow agents across an organization of roughly 50,000 employees and targets up to 60% autonomous IT-ticket resolution. If reached, the company estimates that around 80,000 tickets a year could disappear from the human queue, alongside more than $3 million in annual savings. Those are targets for now, rather than completed results.

IT help desks fit agents unusually well. Password resets, software-access requests and common incidents often have known procedures, structured system data and obvious escalation rules. We expect this category to keep moving quickly because the agent does not have to understand an entire company. It needs to become very good at a finite catalogue of recurring problems.

Are cybersecurity agents trusted enough to take action now?

Cybersecurity agents are already taking real defensive actions today when companies can tightly control what the agent is allowed to do.

Sophos has one of the clearest examples anywhere in the agent market. After a year of running an agentic model inside its managed detection and response service, the company says AI closes 52% of MDR cases end to end without a person intervening. For cases the system is authorized to resolve automatically, the average time from case creation to response is 89 seconds. The same operating model now covers about 40,000 MDR customers.

Those numbers deserve attention because Sophos is counting actual security cases rather than agents created or employees with access. Half of a live workload has moved to AI.

The autonomy still has boundaries. Sophos analysts decide what kinds of cases the AI may resolve, continually adjust those limits and remain responsible for the harder incidents. That setup looks much closer to what enterprises currently trust than giving an AI broad permission to change whatever it wants across a corporate network.

Governance remains a real constraint elsewhere. Deloitte surveyed 3,235 business and IT leaders directly involved with their organizations' AI programs and found only 21% reporting a mature governance model for agentic AI. Security agents can move very fast once permission is granted, so weak governance becomes a much bigger problem than it was with a chatbot that could only answer questions.

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

Are sales agents actually making money, or just sending more emails?

Sales agents are already producing real revenue, but their sweet spot today is relentless follow-up and qualification rather than running a difficult sales process alone.

SaaStr provides a useful example because the agent was given work its tiny human team was barely touching. Nearly 1,000 warm prospects had accumulated in its CRM while three employees focused on larger opportunities. Its Agentforce agent contacted more than 1,000 prospects with over 3,000 emails, recorded a 72% open rate and more than a 10% response rate. Salesforce says those conversations eventually contributed to $2.7 million in closed revenue and another $3.5 million in nurtured pipeline. Humans still stepped in when valuable opportunities became serious.

Equipter shows the same idea at a smaller company. A six-person sales team was receiving as many as 200 leads a week, with lower-quality social leads often waiting around eight hours for a response. After an agent began handling them within minutes, response rates rose from roughly 4% to almost 10%. About 2% of agent-contacted leads eventually became genuine sales opportunities.

Uber for Business has pushed the model further. Its onboarding team faced tens of thousands of unfinished self-service sign-ups each month. After introducing autonomous follow-up, Salesforce reports almost complete lead coverage, a roughly 60% increase in conversion and a 50% improvement in response rate. Uber expects a seven-figure incremental revenue effect.

All three examples come through Salesforce customer studies, so their conversion rates should not be treated as an industry average. They still show something commercially useful: sales teams can hand thousands of low-priority conversations to an agent and recover opportunities that were previously too expensive to chase.

Are finance agents doing real finance work yet?

Finance agents are doing real work now, particularly inside repetitive processes where the rules already exist and people can review exceptions.

Rivian recently described using agents built on Amazon Bedrock to help automate purchase-order accruals and other parts of its financial-close process. The company says the new workflow removes more than 15 days of manual work from each cycle.

The interesting part is why an agent helped. Rivian already had procedures describing how employees should deal with different accounting situations. Those instructions contained enough conditional logic and messy context to make conventional automation awkward. The agent can read the procedure, retrieve ERP information and work through the steps, while important decisions still go through human review.

A recent WNS deployment at a large Middle Eastern bank gives us another example from trade finance. AI agents were inserted into routine processing alongside workflow software and human specialists. WNS reports a 60–70% productivity improvement, 30% shorter turnaround times and 50% fewer compliance-related handoffs. People continue to handle exceptions and decisions that require judgment.

Finance looks more mature than the popular image of an “AI CFO” suggests. Accrual preparation, document checking, reconciliation-style work and transaction processing are already good agent jobs. Letting an agent independently make consequential accounting or capital-allocation decisions is a very different threshold, and we found far less evidence of that happening at scale.

Are AI agents spreading beyond developers now?

AI agents are spreading beyond developers extremely quickly these days, although those newer user groups are growing from much smaller bases.

OpenAI's latest enterprise usage data gives us a rare view based on actual product activity instead of a survey. Since February 2026, weekly active enterprise Codex users increased 108-fold in legal, 41-fold in sales, 41-fold in recruiting and 26-fold in marketing. Engineering grew fivefold over the same period.

Those multipliers need to be read correctly. Engineering started with a huge head start, while many lawyers or recruiters had never had a reason to open a coding agent before. A 108-fold increase therefore does not mean legal now has more agent users than engineering.

The change in behavior is still important. People are using the same agent-style environment for research, data work, document production, lightweight software, analysis and repetitive office processes. OpenAI's broader usage research also found non-developer adoption growing faster than developer adoption across individual and organizational users.

Software moved first because agents have unusually good feedback there: the code either passes tests or it does not. Office work often gives the agent fuzzier instructions and much fuzzier definitions of “finished.” The recent growth suggests users are gradually finding enough repeatable tasks outside engineering to make delegation worthwhile.

Are lawyers actually letting AI agents work on their own?

Lawyers are letting AI agents take action now, but supervised autonomy is still clearly the normal setup.

Deloitte surveyed 121 senior legal leaders around the world in April and May 2026. Sixty-one percent said their legal departments had reached some form of AI deployment, while 10% described AI as fully embedded in daily workflows. When Deloitte narrowed the question to agentic AI, 61% were still experimenting with or piloting it.

Icertis surveyed more than 1,000 U.S. in-house legal professionals and got a more precise look at autonomy. Some 46% said AI was mainly an assistive tool that did not act autonomously. Another 23% said AI occasionally handled tasks autonomously with people kept in the loop. For just under 10%, human review had already become the exception.

That distribution feels believable for legal work. An agent can search contracts, compare clauses, gather information, prepare a draft or move a workflow forward. The uncomfortable part begins when a mistake can create a binding obligation, waive a right or expose the company to liability.

Icertis found only 26% of legal professionals very confident that the AI their teams use is accurate enough for high-stakes business decisions. Nearly half said human judgment still had to be applied before they trusted the output.

Legal is a genuine agent-adoption market today, just one where “agent” usually means a system allowed to do a lot of the work before a lawyer checks the consequential part.

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

Are consumer browser agents becoming normal yet?

Consumer browser agents are usable now, but they have not become a normal mass-market behavior yet.

Google has already moved auto-browsing into Chrome on Android in the U.S. Gemini's agentic mode can perform errands such as booking parking, updating recurring orders and organizing travel. The agentic features are currently offered to paying AI Pro and Ultra users, while Google asks for confirmation before some sensitive actions.

Microsoft's Browse with Copilot can similarly open pages, search, book things and interact with websites. Microsoft explicitly tells users to monitor the agent and recommends avoiding banking, stock trading, credit-card details, government IDs, medical information and highly confidential data.

Those restrictions show where consumer agents currently sit. The products are useful enough to operate a browser for you, while their own developers still want a human close by when the downside gets serious.

Capability benchmarks show why. Stanford's 2026 AI Index reports that the best system on OSWorld, which tests agents across real computer tasks, reached 66.3% accuracy. That is an enormous jump from roughly 12% previously, but an agent failing around one task in three is still hard to trust with arbitrary online errands.

Consumer browser agents may end up becoming one of the largest agent categories simply because Chrome, Edge and major AI assistants already have enormous distribution. As of now, availability has moved much faster than proven habitual use.

Are companies really running teams of AI agents together?

Some companies are running multi-agent workflows for real work, but cross-functional fleets of agents are still early.

Deloitte's recent survey is particularly revealing because every respondent was directly involved in an organization's AI program. Only 15% said their organization had scaled orchestrated, cross-functional multi-agent adoption. Just 5% thought their business processes were highly prepared for agents in the first place.

There are working examples. ServiceNow says its own software-request process uses several specialized agents coordinated by an orchestrator. One agent can understand the employee's request, other agents deal with licensing or related systems, and the workflow completes without somebody manually passing the ticket from team to team.

That is a useful multi-agent system because the agents have clearly separated jobs and operate inside a well-defined process. Creating twenty named “AI employees” and hoping they figure out how to run a department together is a much shakier idea.

Huge agent counts can therefore be misleading. One organization with fifty tiny agents may have less meaningful adoption than another with one agent completing 70% of a major workflow. The amount of useful work moving between agents is the number we would rather see.

Do companies actually want fully autonomous AI agents?

Most companies currently appear to want selective autonomy: let the AI finish routine work on its own and keep people around for decisions where mistakes become expensive.

Deloitte's 2026 survey found 75% of executives agreeing that collaboration between humans and AI agents creates more value than automation by agents alone. Even looking four years ahead, executives largely imagine humans moving into oversight rather than disappearing from the process.

The production deployments we examined fit that preference remarkably well. Security agents can neutralize familiar threats inside permissions set by analysts. Finance agents work through routine procedures and route exceptions to people. Legal agents prepare and execute parts of a workflow while lawyers keep control over high-stakes decisions. Browser agents hand sensitive actions back to the user.

Those handoffs are economically useful, not embarrassing. If an agent can autonomously finish 60% of a workflow and make the remaining 40% dramatically easier for a person, a company can already get a large benefit.

We expect the autonomy boundary to keep moving as models improve. For now, companies are getting the best results by deciding exactly where an agent is allowed to act rather than starting with the assumption that maximum autonomy must be the goal.

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

Are AI agents already changing software budgets and headcount?

AI agents are already changing budgets and individual jobs, but broad company-wide headcount cuts have been much smaller than executives expected.

As seen above, McKinsey found that 32% of respondents had already decided against buying at least one piece of software or feature because agentic coding tools allowed the company to build it itself. That is an unusually concrete budget effect: money that might have gone to another software vendor stays inside the company.

Employment has moved more slowly. In the previous year's McKinsey survey, 32% of respondents expected AI to reduce their organization's headcount over the following year. In the latest survey, only 14% said AI had actually contributed to an overall workforce decline during that period. Two-thirds reported little or no AI-related change in total employment.

The financial picture is similarly mixed when we look at AI as a whole. Eighty percent of respondents say AI improves their own productivity, yet only 37% say it has contributed positively to company EBIT. Just around 6% qualify as McKinsey's AI “high performers,” meaning they report at least a 5% EBIT contribution alongside significant broader value.

Agents can already remove specific queues of work, change whether a company buys software and let a small team cover more volume. Those local effects are real. They have not yet added up to the enormous company-wide labor reduction that people predicted a year ago.

What is still stopping AI agents from spreading faster?

The biggest brakes on AI agent adoption now are messy company data, weak permissions and governance, difficult verification and the growing cost of letting agents run for a long time.

Data is probably the least glamorous problem and one of the largest. McKinsey reports that roughly eight in ten companies see data limitations as an obstacle to scaling agentic AI. An agent cannot reliably process a customer request or make a finance decision when the relevant information is scattered across old systems, contradictory databases and documents nobody has cleaned in years.

Governance has the same problem. Deloitte found only 21% of surveyed organizations with mature agentic-AI governance. A company suddenly needs rules for questions that barely existed with ordinary software: Which credentials can an agent use? How much money can it spend? Which customer records can it change? Who owns a mistake? How can security teams reconstruct what happened afterward?

Verification becomes harder as the work becomes less structured. Coding agents have tests and version control. Support agents have ticket outcomes. Finance systems have accounting rules. A strategy agent asked to “figure out what we should do in Asia” has much more room to be convincingly wrong.

Cost is appearing as another constraint. McKinsey's latest survey found roughly one in five organizations saying AI-related operating costs, including inference costs, had limited AI usage somewhere in the company. Long-running agents can consume far more computation than a short chatbot conversation because they repeatedly reason, search, call tools and check their own work.

That helps explain why adoption is clustering around workflows with clean inputs, clear actions and easy ways to tell whether the job was done correctly. The remaining market is much bigger, but also harder.

So which AI agents are getting real adoption now?

The AI agents getting real adoption now are coding agents, customer-service agents and IT-support agents; cybersecurity, sales and finance are real but narrower, while consumer browser agents and cross-functional agent teams are still early.

Software engineering is furthest along in repeated professional use. Coding agents are contributing production pull requests, completing whole development tasks and influencing whether companies buy software or build it themselves.

Customer service has gone furthest in autonomous execution. Agents can already own a large share of conversations from beginning to end, sometimes reaching into operational systems to cancel bookings, update accounts or complete other actions. Support also gives us something the broader agent market badly needs: a clean number showing what percentage of the workload the AI actually finished.

IT support follows closely because its most repetitive jobs look a lot like customer service. Security shows how far autonomy can go when permissions are tightly designed. Sales is finding real value in the thousands of leads humans never had time to chase. Finance is moving through repetitive, procedure-heavy back-office work. Legal and general knowledge work are growing quickly, although people still review more of the output.

Consumer agents currently have a different problem. The interfaces are arriving quickly, but everyday behavior has not caught up. Multi-agent organizations are earlier still.

Across everything we reviewed, the use cases winning today share a surprisingly practical set of characteristics. There is plenty of repetitive work. The agent can reach the systems it needs. The desired outcome is reasonably clear. Somebody can check whether the job succeeded. Mistakes can usually be corrected. Difficult cases have somewhere to go.

That explains why the first big agent markets look less futuristic than many predictions did. Companies currently get the most value from agents by giving them a well-defined job and enough authority to finish it. Each time models become more reliable, that job can get a little bigger.

Real AI-agent adoption is already here. It is spreading outward from coding, support and IT rather than arriving as one giant wave of autonomous digital employees.

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

Who is the author of this content?

NEW MARKET PITCH TEAM

We track new markets so founders and investors can move faster

We build living "market pitch" documents for emerging markets: AI, synthetic biology, new proteins, and more. Instead of outdated PDFs or hallucinated LLM answers, our clients get a clean, visual, always-updated view of what's really happening: key players, deals, regulations, and signals that matter. Learn more about us.