AI Agents: what is actually working now?
In our agentic AI market deck, you will find everything you need to understand the market
SUMMARY
AI agents are genuinely working now in coding, customer support and tightly structured enterprise workflows, while general-purpose digital workers still fall well short of dependable production use.
The biggest change is task ownership. Companies are increasingly letting AI inspect systems, make decisions, call tools and carry work through several steps instead of stopping at a draft or recommendation.
Production adoption is already substantial inside large companies, but the depth of that adoption is easy to exaggerate. Around 40% of large enterprises report scaling agents somewhere, while only 15% of companies in Deloitte’s agentic-AI sample had scaled orchestrated multi-agent systems across functions.
Structure appears to matter more than task length. Coding agents can sometimes complete work measured in days or weeks of human effort because code provides tests, errors and clear success criteria, while the best system on OSWorld 2.0 fully completed only 20.6% of messy computer workflows that take humans roughly 1.6 hours.
Coding is currently the strongest proof that agents can create serious economic value. Measured productivity gains are appearing alongside unusually large migration projects, longer autonomous tasks and even decisions by companies to build software internally rather than buy it.
Customer support is the clearest non-coding success. Several real deployments now resolve roughly half or more of eligible conversations without human takeover, although vendor definitions differ enough that the percentages should not be compared as if they came from one benchmark.
The economics look much stronger at the workflow level than at company level. Individual engineering and support deployments can show excellent returns, while only 37% of respondents in McKinsey’s broader survey reported any positive EBIT impact from AI and roughly 6% qualified as AI high performers.
Job change is also arriving before mass job elimination. McKinsey found that 32% of companies had expected AI-related head-count reductions, but only 14% later reported an actual decline; meanwhile, engineers, support staff and other knowledge workers are already handing more execution work to agents.
The main bottleneck increasingly sits around the agent rather than inside the model. Data access, integration, permissions, undocumented exceptions and governance decide whether an impressive pilot becomes a production workflow, and only 5% of Deloitte respondents considered their processes highly prepared for agents.
The practical model today is controlled autonomy. Agents work best when the objective is clear, the information is accessible, actions are bounded, progress can be checked and expensive mistakes are caught before they propagate. Multi-agent systems can help when work divides cleanly, but adding more agents does not fix a badly structured process.
This market map, featured in our agentic AI market deck, highlights top companies and startups in the agentic AI market
Why are AI agents suddenly worth taking seriously?
AI agents are worth taking seriously now because people are increasingly handing them whole tasks rather than simply asking them for answers.
That change is showing up in actual usage. OpenAI reported that by June 2026, Codex generated 64% of the combined output tokens produced by Codex and ChatGPT among its enterprise customers. Longer agent tasks naturally use more tokens, so 64% does not mean most enterprise AI tasks are agentic. Still, it shows how quickly usage has shifted toward AI that executes work.
Anthropic found the same pattern while studying roughly 400,000 Claude Code sessions between October 2025 and April 2026. Debugging fell from 33% to 19% of sessions. Operating software rose from 14% to 21%, while writing and data analysis roughly doubled from about 10% to 20%. Anthropic also estimated that the economic value of the typical task increased 27% over those seven months.
Microsoft gives us a broader enterprise view. Active agents inside Microsoft 365 grew 15-fold in a year and 18-fold inside large companies. Meanwhile, McKinsey's latest global AI survey found that 40% of respondents at companies with more than $1 billion in annual revenue were scaling agents in at least one function, up from 27% a year earlier.
The shift is real. The open question is how much meaningful work companies can safely hand over, because the answer varies enormously depending on the job.
If you want more recent data on this point, please see our latest agentic AI market report.
What should we call an AI agent that actually works?
An AI agent is actually working when we can give it an outcome, let it execute several meaningful steps, and get reliable enough results that using it still makes economic sense after supervision and mistakes.
That definition removes a lot of inflated "agent" adoption.
A chatbot that retrieves a help-center article may technically contain an agent somewhere in the architecture. From the customer's perspective, very little has changed. The same goes for an AI coding tool that suggests the next ten lines while a developer remains responsible for every action.
The more interesting boundary appears when the AI starts carrying the work forward. A coding agent can receive an issue, inspect a repository, edit several files, run tests and prepare a pull request. A support agent can inspect an account, understand a cancellation request, call the relevant system, cancel the subscription and process a refund.
Autonomy alone tells us surprisingly little about usefulness. An agent that works alone for an hour and succeeds 40% of the time may create more work than one that handles a ten-minute process correctly 95% of the time.
So the test here combines four things: how much of the task the agent handles, how often it finishes correctly, whether companies use it repeatedly in production, and whether the economics still look good once human review is included.
| Type of AI use | What the system actually does | Where it stands today |
|---|---|---|
| Assistant | Suggests, drafts, summarizes | Widespread |
| Task agent | Completes a bounded multi-step job | Working well in several areas |
| Workflow agent | Moves through systems and completes a process | Working in controlled environments |
| General digital worker | Handles changing, open-ended work with little supervision | Still unreliable |
| Multi-agent workforce | Several agents coordinate broad business processes | Early |
As this chart shows, and as featured in our agentic AI market deck, search interest in AI agents has been rising rapidly
How much AI-agent adoption has actually reached production?
AI-agent adoption has reached real production scale, especially inside large companies, but deep company-wide autonomy is still much rarer than the number of pilots makes it look.
McKinsey's latest survey gives us a useful baseline. Around two in ten respondents said their companies were scaling AI agents. Among businesses with more than $1 billion in annual revenue, the figure reached 40%. Smaller companies stayed around 22%.
Microsoft's telemetry points in the same direction without relying on survey answers. Active agents in Microsoft 365 grew 15-fold year over year and 18-fold inside large enterprises.
Salesforce offers another measure from inside a major agent platform. More than 25,000 companies have built and deployed Agentforce, according to Salesforce. In its latest reported quarter, those agents completed 3.2 billion "agentic work units," up 97% from the previous quarter. Salesforce defines a work unit as an action such as making a decision, updating a record or triggering a workflow, so 3.2 billion work units should not be read as 3.2 billion complete human jobs. The useful part is the pace: quarterly activity almost doubled.
Then we hit a much lower number when we ask about deeper automation. Deloitte recently surveyed 501 US executives at organizations that were already at least piloting agentic AI. Only 15% had scaled orchestrated, cross-functional multi-agent systems. Just 5% thought their business processes were highly prepared for agents.
Companies are putting agents into production now, mostly one workflow at a time.
| Current evidence | Result | What we can reasonably conclude |
|---|---|---|
| Large companies scaling agents, McKinsey | 40% | Enterprise adoption is substantial |
| Smaller companies scaling agents, McKinsey | 22% | Adoption is much less broad below large-enterprise level |
| Microsoft 365 active-agent growth | 15× YoY | Actual usage is growing very fast |
| Agentforce Q2 work-unit growth | +97% QoQ | Production activity is accelerating |
| Scaled cross-functional multi-agent systems, Deloitte | 15% | Deep orchestration remains uncommon |
| Processes rated highly prepared for agents, Deloitte | 5% | Companies are less ready than the technology hype suggests |
Can AI agents actually replace software engineers?
AI agents can already replace large pieces of software-engineering work, and coding is currently the strongest proof that agents can handle economically valuable tasks at scale.
McKinsey found that around two in ten companies were already scaling coding agents, rising to 31% among businesses with more than $1 billion in revenue. Even more interestingly, 32% of respondents said their organizations had skipped buying at least one software product or feature because coding agents let them build it internally.
That is starting to change software budgets, not merely developer productivity.
Recent company data makes the case stronger. 1Password measured a 20.9% productivity improvement among a core group of engineers using Codex and a 10.9% reduction in median pull-request cycle time. During one difficult production incident involving more than ten microservices, investigation time fell from roughly two hours to between five and twenty minutes.
Asana provides the extreme example. The company needed to remove Enzyme, an old testing framework embedded across its frontend codebase. Its previous staffing plan estimated at least five years of work and around $6 million in cost. With several Codex agents working in parallel and engineers reviewing the changes, Asana completed the migration in roughly two calendar weeks for about $12,000 in model and infrastructure spending.
We should not turn that Asana result into a general productivity multiplier. The migration contained large amounts of repetitive, parallelizable work, exactly the kind of project where agents can look extraordinary.
METR shows how far individual coding tasks have stretched. Its benchmarks measure difficulty using the amount of time a skilled human would need to complete the same software task. In early 2026, leading agents pushed beyond two full working days at the 50% success threshold on METR's main software suite, although the researchers warned that estimates above 16 hours were increasingly uncertain as the benchmark became saturated.
A separate METR and Epoch AI experiment called MirrorCode went further. Frontier agents successfully rebuilt software corresponding to weeks of human development, including a codebase of roughly 16,000 lines.
Large parts of an engineer's job remain outside those benchmarks. The benchmark provides the task, environment and success criteria. A real engineer also has to deal with unclear product decisions, undocumented company history, security constraints and changing stakeholder priorities.
Anthropic's Claude Code research gives us a good picture of the current split. Humans still make most planning decisions about what should be done. Claude handles more of the execution decisions about how to do it.
Software engineering is already being compressed heavily at the task level. Full replacement of the engineer remains much less convincing.
If you want more recent data on this point, please see our latest agentic AI market report.
This chart, included in our agentic AI market deck, illustrates yearly VC funding for agentic AI startups
Are customer-service AI agents really resolving problems now?
Customer-service AI agents are already resolving a large share of real support conversations, making customer service the strongest non-coding agent use case today.
Intercom provides the clearest aggregate figure. More than 7,000 teams now use Fin, and Intercom says its average resolution rate reached 76% across customers in June 2026. The rate has risen roughly one percentage point per month over the previous two years, according to the company.
We have to be precise about what 76% means. Intercom's resolution rate measures conversations handled by Fin that were resolved without a human teammate taking over. It differs from the percentage of all incoming support conversations automated because Fin may not participate in every conversation.
The individual deployments are still impressive. Anthropic reached 50.8% resolution while putting Fin into 96% of eligible conversations. Lightspeed reports 65–72% resolution across more than 43,000 monthly conversations with 99% involvement.
Zendesk customers show similar results through a different platform. BritBox now automates 47% of ticket volume, has reduced full-resolution time by 27% and reports an 86% satisfaction score for interactions handled by its AI agent.
Salesforce says Engine, the travel platform, automatically resolves 50% of incoming inquiries while reducing human handle time by 15%. Grupo Falabella reports 60% autonomous resolution over WhatsApp.
Different vendors count resolution differently, so putting every percentage in the same league table would be misleading. Across platforms, though, agents are clearly handling enough customer volume to change how support teams operate.
Do AI agents still fall apart when the job gets messy?
AI agents still struggle badly with long, messy computer work where information changes, instructions are incomplete and the next step cannot be checked automatically.
OSWorld 2.0 exposes the problem particularly well. The benchmark contains 108 realistic computer workflows covering everyday and professional work. A human needs a median of about 1.6 hours to finish one. Agents may need hundreds of tool calls and have to move across applications, remember earlier constraints, interpret changing information and recover hidden state.
The best tested system completed only 20.6% of the workflows under the benchmark's main full-completion measure. GPT-5.5 plateaued at roughly 13%.
Those numbers look almost bizarre beside the coding results. METR can observe agents finishing some software tasks measured in days or even weeks of human effort, yet a 1.6-hour office workflow can still defeat them.
Software can provide tests, compiler errors, version control and explicit task specifications. A real office workflow may send an updated email halfway through the job, hide an important condition in a spreadsheet, require information from another app and give no automatic warning when the agent chooses the wrong interpretation.
OSWorld's researchers found exactly those failure modes. Agents lost constraints, missed new information, guessed instead of asking for clarification and sometimes finished without checking whether the result was actually correct.
Structure currently matters much more than raw task duration.
| Type of work | What current agents can do | Current limitation |
|---|---|---|
| Well-specified coding | Some tasks equivalent to days or weeks of human work | Results still need validation |
| Customer support | Resolve a large share of conversations automatically | Depends heavily on good knowledge, data and rules |
| Structured enterprise workflows | Read systems, update records and trigger actions | Permissions and integration constrain scope |
| Long realistic computer workflows | Best OSWorld 2.0 system completes about 21% fully | Agents lose context and mishandle changing information |
| Broad open-ended jobs | Useful pieces can be delegated | End-to-end reliability remains too low |
This chart, included in our agentic AI market deck, shows how Cognition is positioned in agentic AI
Are AI agents useful outside coding and customer support?
AI agents are becoming genuinely useful in research, data analysis, sales, legal work and recruiting, although we have much less hard evidence of end-to-end automation in those jobs.
OpenAI's enterprise data shows how quickly agentic use is spreading outside engineering. Since February, 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 giant multiples partly reflect much smaller starting points outside engineering. They show direction more clearly than absolute adoption. Lawyers have not suddenly become heavier coding-agent users than programmers.
Anthropic found the same widening of use cases inside Claude Code. Writing and data analysis roughly doubled to around 20% of observed sessions combined. Management, sales and legal occupations were among its fastest-growing non-software user groups.
Research is a particularly natural extension. An agent can search, open sources, extract data, compare documents, calculate figures, edit spreadsheets and assemble a draft.
The weak point is verification. Code can fail a test and tell us immediately. Research can omit the decisive source while producing a perfectly confident conclusion. Sales presents a similar issue: preparing an account brief and updating a CRM are easy to evaluate, while deciding how to handle a politically sensitive enterprise negotiation is much harder.
The useful unit outside coding and support is still usually the task rather than the whole job.
If you want more recent data on this point, please see our latest agentic AI market report.
Are multi-agent systems actually working, or are they mostly hype?
Multi-agent systems are working in some specialized setups, but companies currently have far more success scaling individual agents than building whole teams of agents that coordinate across a business.
Deloitte's recent survey makes the gap unusually visible. Every company in the sample had already reached at least the pilot stage with agentic AI, yet only 15% had scaled orchestrated multi-agent systems across functions.
And some of that 15% was still being used for relatively low-risk, low-return work, according to Deloitte.
There are good reasons to use several agents. A coding system can let one agent implement a change and another review it. A research system can send different agents down separate research paths and have another compare the findings. Different agents can also hold different permissions, which limits the damage one component can cause.
The problems start when orchestration becomes more complicated than the job itself. Each handoff creates another opportunity to lose context. Someone — or another model — must decide which agent acts next, determine whether its work is correct and reconcile disagreements.
Asana's recent Enzyme migration is useful here. Up to four coding agents ran in parallel, each in its own copy of the codebase, but humans still checked progress twice a day and reviewed every proposed change. The architecture worked because the work could be split cleanly and verified.
For now, multi-agent systems make the most sense when there is an obvious reason to divide the work.
This chart, included in our agentic AI market deck, illustrates yearly funding for agentic AI startups
Are AI agents actually delivering a good return on investment?
AI agents can deliver exceptional ROI on the right workflow today, while the average company is still struggling to turn local productivity gains into visible profit.
1Password offers one of the freshest measured examples. Among a core cohort of engineers using Codex, the company recorded a 20.9% productivity gain and a 10.9% reduction in pull-request cycle time.
1Password then modeled the economics for 50 consistent users. Using a fully loaded developer cost of $250,000, attributing 40% of the measured productivity gain to Codex and assuming that 75% of the resulting capacity could actually be used, the company estimated about $784,000 of annual engineering capacity and a 553% ROI.
The assumptions matter, so the 553% figure should be treated as a company model rather than audited profit. The underlying operational improvements were measured separately, which makes the example much more useful than a vague claim about hours saved.
Asana's roughly $12,000 Enzyme migration provides a second extreme case. The company's old staffing plan had been estimated at around $6 million. We would not extrapolate that ratio across ordinary software development, but it reveals how strong the economics can become on large repetitive migrations.
Customer support gives us a different kind of ROI. BritBox reached a 47% AI automation rate and reported payback in under a year. Zendesk says Docebo automates 35% of chat resolution and estimates annual savings above $250,000.
The enterprise-wide numbers are much more modest. McKinsey's latest survey found that 37% of respondents attributed at least some positive EBIT impact to AI, essentially unchanged from the previous year despite much broader deployment. Only about 6% qualified as AI high performers, meaning they attributed at least 5% of EBIT to AI and described its impact as significant.
One in five companies also said AI operating costs were already constraining usage somewhere in the organization.
Some agent workflows already have excellent economics. Most companies are still figuring out how to make those gains large enough to show up clearly in overall profit.
If you want more recent data on this point, please see our latest agentic AI market report.
Are AI agents already replacing jobs?
AI agents are replacing pieces of people's jobs today much faster than they are eliminating whole jobs.
The latest McKinsey survey gives us a rare opportunity to compare last year's predictions with what actually happened.
In 2025, 32% of respondents expected AI to reduce their organization's total head count during the following year. When McKinsey returned to the question, only 14% reported an actual AI-related head-count decline over the previous year. Two-thirds said AI had caused little or no change in total employment.
Expectations are running ahead again. Now 39% expect AI to reduce head count over the coming year.
Maybe they will eventually be right. For now, the repeated tendency to predict more immediate job destruction than companies actually report should make us cautious about claims that agents are already removing huge numbers of office workers.
The work itself is changing much more visibly.
Anthropic's Claude Code research found that humans make most planning decisions while Claude handles more execution decisions. In customer support, agents absorb routine conversations and leave people with more exceptions and unusual cases. In software engineering, developers increasingly specify, review and coordinate work that agents execute.
That can still reduce hiring. A five-person team that suddenly produces what once required eight people does not need layoffs to change the labor market; it can simply hire fewer additional employees.
This chart, included in our agentic AI market deck, compares the main business model options for autonomous AI agent platforms
Why do so many AI-agent pilots still fail to scale?
AI-agent pilots usually get stuck because the surrounding company process is too messy, too poorly connected or too risky for the agent to run reliably.
The latest Deloitte survey makes the readiness gap hard to miss. Only 5% of organizations already experimenting with agents considered their business processes highly prepared for them. Seventy-two percent said unified and accessible data was a problem, 70% cited difficulty trusting and governing agents, and 67% said integration was too expensive or complex.
McKinsey has found something similar from the data side. In earlier research on agentic scaling, eight in ten companies identified data limitations as a roadblock.
The companies getting the strongest financial results from AI tend to behave differently. In McKinsey's latest survey, nearly three-quarters of its small group of AI high performers said they had fundamentally redesigned workflows around AI. Among everyone else, only around one-quarter had done so.
Customer support shows why workflow design changes the outcome. An agent performs better when the company has clean help content, accessible account data, specific procedures, clear escalation rules and a small set of approved actions. Intercom even analyzes failed conversations and recommends which missing content, customer data or actions companies should add.
Coding has similar built-in advantages. Tests tell the agent when something broke. Git records every change. Pull requests create a review point. A failed experiment can often be rolled back.
Many office processes have accumulated years of undocumented exceptions, manual handoffs and information spread across different systems. Giving an agent access to that mess does not clean it up.
How much autonomy can companies safely give AI agents today?
Companies can safely give AI agents substantial freedom inside a narrow lane today; giving them broad authority across sensitive systems is still difficult to justify.
The technical evidence gives us the reason. On OSWorld 2.0's realistic long workflows, even the best tested agent fully completed only around one task in five. An error rate of that size is manageable when the consequence is a draft or a pull request waiting for review. It becomes dangerous when the same agent can move money, delete customer data or approve a legal commitment.
Companies are therefore building graduated authority into production systems.
Low-risk actions can run automatically. More consequential operations can require extra checks. High-impact decisions go to a person. Coding agents create pull requests instead of quietly pushing every change into production. Customer-service agents can receive tightly defined procedures for refunds and account changes.
Governance research suggests many companies still have work to do here. Deloitte found that 70% of executives in its agentic-AI sample were worried about their ability to trust and govern agents.
Gartner recently went further and predicted that 40% of enterprises will demote or decommission autonomous agents by 2027 after governance problems emerge in production. A Gartner forecast is obviously a prediction rather than observed failure data, so the percentage deserves some caution. Its underlying point fits what we can already see: permissions become more important as agents gain the ability to act.
Controlled autonomy is the model that currently fits the evidence best.
This chart, featured in our agentic AI market deck, shows the share of revenue generated by each customer segment in the agentic AI market
So what AI agents are actually working now?
AI agents are genuinely working now in coding, customer support and tightly structured business workflows; broader knowledge-work agents are becoming useful fast, while general digital employees and autonomous multi-agent organizations still fall well short of dependable production work.
Coding sits furthest ahead. We now have enterprise adoption data, measured productivity improvements, tasks stretching into days of human-equivalent work and real companies changing make-versus-buy decisions because coding agents make internal development cheaper.
Customer support is the other clear winner. Several large deployments automate roughly half of eligible conversations or more, while current platforms can increasingly execute cancellations, refunds and account changes instead of stopping after an answer.
Structured operational work comes next. Agents can update CRM records, trigger workflows, investigate systems and handle repetitive processes when companies provide reliable data, specific tools and clear permission boundaries.
Research, analysis, sales, recruiting and legal work are moving quickly. Usage is rising sharply, and agents already carry meaningful pieces of those jobs. We have much less evidence that companies can hand over the complete outcome and walk away.
The dividing line becomes obvious when we compare these successes with realistic computer-use benchmarks. Current agents can sometimes finish software projects measured in days or weeks of human work, yet the best system tested on OSWorld 2.0 completed only 20.6% of messy 1.6-hour computer workflows from beginning to end.
Agents work best when the outcome is clear, the relevant information is available, the allowed actions are defined, progress can be checked and failure can be caught before it causes serious damage.
For now, AI agents have crossed the line from interesting demos to useful workers on specific tasks. They have not crossed the much harder line into dependable replacements for whole human jobs.
If you want more recent data on this point, please see our latest agentic AI market report.
OUR METHODOLOGY
This analysis tests the question “AI Agents: what is actually working now?” by separating several things that are often mixed together: adoption, task capability, production reliability, economics, changes to human work and the amount of autonomy companies can safely allow.
For each part, we prioritized the freshest measurable evidence available and gave the most weight to first-hand sources: product telemetry, original benchmark results, production deployments, operational performance data and large enterprise surveys. Usage growth carried more weight when it was supported by evidence that agents were also completing real work or producing measurable outcomes.
We kept different measurements in their proper context. METR's time-horizon work measures coding-task difficulty using the time a skilled human would need to complete the same task, while MirrorCode tests larger software reimplementation projects. OSWorld 2.0 examines a different problem: long, stateful computer workflows where information, applications and constraints can change. Those results help explain why an agent can succeed on a long coding task and still fail on a shorter office workflow.
Production case studies were used to show what already works in specific environments rather than to estimate a universal market average. The 1Password and Asana deployments provide measured engineering examples, while Intercom, Zendesk and Salesforce provide customer-support deployments. Resolution, involvement and automation rates differ by platform, so we use them to establish the scale of real autonomous work rather than rank vendors against one another.
We treated measured operational improvements separately from modeled economics. Productivity changes, pull-request cycle times, resolution rates and support automation are operational results. ROI estimates such as 1Password's 553% figure depend on company assumptions. McKinsey's broader EBIT findings provide a separate check on whether those local gains are already large enough to become visible across an entire organization.
Key sources used for this analysis include OpenAI's enterprise usage data, Anthropic's Claude Code research, Microsoft's Work Trend Index, McKinsey's global State of AI survey, McKinsey's comparison of expected and actual AI-related job losses, Salesforce's Agentforce metrics, Deloitte's agentic-AI readiness survey, OpenAI's 1Password deployment study, OpenAI's Asana deployment study, METR's Task-Completion Time Horizons, METR's updated benchmark methodology, Epoch AI and METR's MirrorCode research, the original OSWorld 2.0 paper, Intercom's Fin deployment data, Intercom's resolution methodology, Zendesk's BritBox case study, Zendesk's Docebo case study, Salesforce's Engine case study, Salesforce's Grupo Falabella case study, and Gartner's autonomous-agent governance forecast.
This chart, included in our agentic AI market deck, shows how autonomous AI agent platform technology has evolved over time
Related blog posts
- AI Agents: what are the biggest challenges now?
- AI Agents: what’s changing now?
- AI Agents: what are the biggest unsolved problems?
- AI Agents: what is getting real adoption now?
- What are the latest funding developments in the agentic AI market?
