Close that last mile in automation with computer-using agents

For years, enterprise automation has hit a wall when a process crosses applications that do not share a clean application programming interface (API) and work falls back to people pushing transactions forward one screen at a time. That is where computer-using agents (CUAs) are changing the equation.

RPA is now history, so close that last-mile automation gap with computer-using agents

Instead of waiting for every application to expose a modern interface, CUAs can interpret screens, navigate browsers and desktop apps, click, type, and complete multi-step work across the same visual layer employees use every day. The significance for enterprise automation leaders is not better screen scraping. It gives them a faster route to business outcomes, shorter transformation cycles, lower manual effort, and more automation coverage across the applications that still do not talk to each other.

From task automation to system orchestration, the visual layer may become the new enterprise integration layer

Human middleware, which involves people toggling between applications, rekeying information, reconciling screens, and pushing work forward one click at a time, is the final frontier of the automation challenge. CUAs address this problem directly by interpreting screens, moving across applications, clicking buttons, entering data, and adapting to the visual context of the task (see Exhibit 1).

Exhibit 1: CUAs are the missing execution layer in enterprise automation

Source: HFS Research, 2026

Traditional robotic process automation (RPA) succeeded because it brought structure to repetitive work through detailed scripting against known selectors and screens. The challenge is that small interface changes could break entire processes and force automation teams into constant maintenance. CUAs promise a more adaptive and resilient model for solving this problem by interpreting screen context and adjusting layouts, fields, or workflows when they change. They still need prompts, policies, and workflow orchestration, but the interaction model is closer to assigning work to a capable operator than coding every click in advance.

CUAs do not replace APIs, integration platforms, or workflow systems. They extend them by creating a practical execution layer for the long tail of enterprise work that still depends on screens, clicks, context switching, and human judgment. The opportunity for CXOs is not just better task automation. It is a shift toward system orchestration across the full software estate, including the legacy, third-party, and black-box environments that have historically resisted automation.

The category is maturing quickly from curiosity to disciplined, pilotable offerings

Most enterprise automation remains siloed by system or function. Human workers, by contrast, operate horizontally. They read a message, check a claim in one system, validate details in another, consult a knowledge source, and then take action. Anthropic’s Cowork is the clearest example of a CUA and is framed as a system that executes multi-step knowledge work on the user’s behalf across their computer, local files, and applications.

The ability of CUAs to mimic that cross-system behavior in one coordinated flow is what sets them apart from other automation tools. That creates potential for a more horizontal digital workforce, one capable of stitching together fragmented work across the enterprise. In enterprise terms, it is a new path to automate work across the software estate that APIs, classic RPA, and manual operations have not fully solved. The enterprise question becomes less of “can the model click?” and more of “can it finish the work?”

Similarly, OpenAI (Operator), Microsoft (Copilot Cowork), and UiPath (Screenplay) are pushing in the same direction. The important signal is that CUAs are moving from isolated demos to productized capabilities for real enterprise work as models, tooling, orchestration patterns, and governance controls are advancing simultaneously.

CUAs can help close the last mile of productivity

CUAs can automate across third-party portals, partner ecosystems, government interfaces, and legacy applications without waiting for a full integration program or custom API build. They promise a more adaptive and resilient model for solving the last-mile RPA problem by interpreting screen context and adjusting layouts, fields, or workflows when they change. For automation leaders, that means:

1. Breaking the integration backlog. The first business impact is speed. You do not need to wait for every acquired platform, partner portal, or legacy desktop application to be modernized before automating the process around it. CUAs provide a practical route to execute work through the existing interface, shortening the distance between business demand and delivery, and reducing reliance on long integration queues.
2. Removing swivel-chair work that drains productivity. The second business impact is labor efficiency. Many operations teams still spend a large share of the day moving between inboxes, ERPs, claims systems, service desks, finance tools, and external websites. CUAs do more than automate a single repetitive task; they can reduce the cross-application effort that forces skilled employees to act as human middleware. That creates capacity for higher-value work without waiting for a full-scale application overhaul.
3. Moving automation from isolated tasks to horizontal workflows. The third business impact is orchestration. Employees rarely work on one application at a time. They investigate an issue in one tool, validate it in another, search for context in a collaboration platform, and then take action somewhere else. CUAs can begin to automate those horizontal flows rather than optimizing each application in isolation.

The right response is a disciplined pilot strategy tied to operating friction

It is clear that the autonomous enterprise will be built through a coordinated workforce of humans and agents operating across the full application estate, including the environments that were never designed to integrate cleanly. The next phase of enterprise transformation will not come only from better APIs, but from orchestrating work across the visual layers where key processes still run.

However, there are no standardized click-through processes, which is precisely why RPA has failed in these scenarios. Success in real-life applications will depend on how trainable the CUA is and how reliably and transparently it can bring coordinated execution across many apps. That is why a pragmatic approach to experimentation and scaling is critical.

CUAs should be used selectively for the visual steps that the rest of the automation stack can’t reach. Start with workflows where employees spend a disproportionate amount of time moving data between systems that don’t talk to each other. Claims operations, order management, revenue-cycle processes, finance operations, field service coordination, compliance checks, and service desk workflows are obvious candidates. The ideal pilot is not the flashiest use case; it is the one with fragmented interfaces, clear operational pain, and measurable manual effort.

The next step is to build human-in-the-loop oversight into the design as supervised autonomy, which lets the agent gather context, navigate systems, and prepare actions while routing high-stakes decisions or final approvals to a human. This captures much of the efficiency benefit while controlling operational risk.

More importantly, treat CUA adoption as how work gets executed across humans, applications, and agents. Organizations that move first with discipline will not just automate tasks faster. They will redesign how operational work flows across the full software stack.

The Bottom Line: CUAs bring AI into the visual layer of enterprise work, where much of the real operational friction still lives. Use them selectively, govern them rigorously, and apply them where visual interaction is the shortest path to operational value.