Computer use agents Market Outlook 2026–2032: Growth Drivers, Key Companies, Strategic Developments & Adoption Forecast — A Meticulous Next™ Foresight Brief

This Meticulous Next™ brief examines computer use agents, artificial intelligence systems that look at a screen, move a pointer, type and complete tasks in software that has no application programming interface. It tests how far these agents move from demonstrations to a managed layer of enterprise automation by 2032, and what that means for robotic process automation vendors, platform owners, chief information officers and the firms that integrate them.

Capability has risen quickly on standard tests. The original OSWorld benchmark reported a state of the art of 12.24% in 2024 against a human baseline of 72.36% on 369 tasks, and by August 2026 leading models were reported at about 72.5% to 72.7% on the verified version, with higher self-reported figures. The conclusion is that agents have reached the human baseline on short, well defined tasks, and the remaining gap lies in length, reliability and cost.

Long tasks remain hard. In OS-Marathon, a benchmark of 100 repetitive workflows, the best agent reached 68.29% of fields correct but completed only 11.53% of whole tasks end to end, against 82.62% for human evaluators, and a UiPath agent scaffold with a leading model completed 0.56%. A separate study from the University of California, San Diego found that the best agents take 2.7 to 4.3 times more steps than necessary. The conclusion is that errors compound over long workflows, and enterprises will keep people in the loop for the next several years.

Supply is consolidating. OpenAI shut down Operator on 31 August 2025 and folded it into ChatGPT agent, Anthropic launched a computer use research preview for macOS on 24 March 2026, and Google closed Project Mariner on 4 May 2026, citing compute cost, slow performance and form errors. The conclusion is that model providers are moving computer use into general assistants and desktop apps, and standalone products struggle on cost.

The brief treats computer use agents as a high-impact signal with a six year horizon to 2032. It identifies which model providers, automation vendors and integrators capture value as agents move from trials to managed use. It is a focused 108-page decision brief that presents an indicative trajectory rather than a segmented market model.

The agent works from the screen. It receives images of the display, decides where to click or type and repeats until the task ends, which lets it use any software that a person can use, including systems with no application programming interface. The conclusion is that legacy applications, remote desktops and vendor portals can be automated without integration work, which is the main appeal for chief information officers.

Scripted automation is the incumbent. Robotic process automation records fixed steps and breaks when a screen changes, while an agent reads the screen and adapts. The conclusion is that agents threaten the maintenance and exception handling work that automation vendors charge for, and vendors respond by adding agents to their own platforms.

Cost and speed set the economics. Agents use roughly four times the tokens of a chat interaction, secondary estimates put the cost at USD 0.03 to 0.13 per task and each action takes a fraction of a second to several seconds. Zocdoc reported USD 2 per successfully booked appointment with a voice agent. The conclusion is that agents pay back where a task is valuable and repeated, and compute cost limits use on low value tasks.

Security is unresolved. Anthropic reported that attacks succeeded 23.6% of the time without mitigations and 11.2% with them in autonomous mode for its Chrome agent, and a study of adaptive attacks bypassed twelve published defences with success rates mostly above 90%. A study by the National Institute of Standards and Technology found that raising attempts from one to 25 lifted average attack success from 57% to 80%. The conclusion is that defences reduce risk but do not remove it, so enterprises will limit scope and require approval for sensitive actions.

The limits are long task reliability, compute cost, prompt injection, the quality of benchmarks and the need for human approval. The brief tests each limit against the evidence to 2032.

The computer use agents market, covering agent platforms, hosted virtual desktop and browser environments and enterprise licences that let artificial intelligence agents operate graphical interfaces, and excluding application programming interface based workflow automation, scripted robotic process automation, chat assistants and coding agents, is valued at USD 0.60–1.00 billion in 2026 and is projected to reach USD 4.14–6.91 billion by 2032, at a CAGR of 38%.

The 2026 base rests on agent platforms and subscriptions, hosted environments for running agents and enterprise licences for agent features inside automation and workflow platforms. Growth then follows the number of workflows that agents complete without help, the cost per task, the share of enterprises that approve agent use and the spread of hosted desktops and browsers.

The forecast is sensitive to gains in long task reliability, the cost of compute and the success of defences against hijacking.

Scenarios.

The base case assumes long task completion rises steadily, compute cost per task falls by half by 2029 and enterprises approve agents under standard controls by 2029, bringing the inflection to about 2029 and a 2032 value near the middle of the range. An accelerated case assumes a sharp gain in long task reliability and strong defences, moving the inflection to 2028 and lifting the 2032 value to the top of the range. A delayed case assumes slow gains in reliability and a serious hijacking incident, moving the inflection to 2031 and holding growth at the bottom of the range.

Growth drivers

• Rapid gains in benchmark scores for screen operation.
• Large stock of legacy software without interfaces.
• Pressure on scripted automation and its maintenance cost.
• Launches of agent features by model providers and automation vendors.

Enablers

• Capable vision and language models.
• Hosted virtual desktops and browsers.
• Agent features in existing enterprise platforms.
• Open source agent frameworks.

Restraints and barriers

• Low end to end completion of long workflows.
• High compute cost and latency per task.
• Prompt injection and hijacking risk.
• Benchmarks that can be gamed and do not match real work.

The direction of computer use agents is set by the pace of reliability gains on long tasks, the cost of compute, the strength of defences against hijacking, the response of automation vendors and the willingness of enterprises to approve agents in sensitive systems.

The brief assesses policy and security evidence, provider and investor funding, model and benchmark readiness, platform and hosting readiness, and enterprise readiness, and rates each force for direction, speed and confidence.

The key participants in computer use agents and software automation without application programming interfaces include Alphabet Inc., Microsoft Corporation, Amazon.com, Inc., Salesforce, Inc., ServiceNow, Inc., SAP SE, Oracle Corporation, International Business Machines Corporation, UiPath, Inc., Automation Anywhere, Inc., Appian Corporation, Pegasystems Inc., SS&C Technologies Holdings, Inc., Workday, Inc., NVIDIA Corporation, Accenture plc, Infosys Limited, Tata Consultancy Services Limited, Cognizant Technology Solutions Corporation, Capgemini SE, OpenAI, Anthropic, browser-use, Browserbase, Magnitude, Zocdoc, Wiz, GenseeAI, the University of California, San Diego, the University of California, Berkeley, the University of Oxford, the Georgia Institute of Technology, the University of Surrey, the University of Washington and the National Institute of Standards and Technology.

The brief profiles representative participants in each archetype and assesses which are positioned to lead as agents move from trials to managed use.

Alphabet Inc., Microsoft Corporation and Amazon.com, Inc. lead among cloud and platform owners that host agents, UiPath, Inc. and Automation Anywhere, Inc. lead among automation vendors, and ServiceNow, Inc., Salesforce, Inc. and SAP SE lead among enterprise platforms that add agent features. Competitive intensity is high among model providers and rising among automation vendors.

Where value migrates.

In 2026 value sits with model providers and agent specialists that supply the capability and with hosting providers that run it. By 2028 it moves toward automation vendors and enterprise application platforms that embed agents with approval, logging and exception handling, because buyers want agents inside systems they already trust. By 2032 it settles with suppliers that combine reliable long task completion, low cost per task and strong controls, where integration, trust and distribution carry the margin. Margins hold for platforms with large customer bases and workflow data and narrow for suppliers that only offer a thin agent layer over a public model.

Rapid benchmark gains, a large base of legacy software, pressure on scripted automation and launches by large providers support growth, while low long task completion, compute cost, hijacking risk and doubts over benchmarks hold back the pace.

A gain in long task reliability, a fall in compute cost, or a serious hijacking incident could shift the pace. The brief therefore tracks four indicators: end to end completion on long task benchmarks, cost per completed task, share of enterprises with approved agents and reported hijacking incidents.

Human in the loop pilots first, then reliable long workflows, then agents as the default for legacy software is the expected order of change.