Computer use agents: From robotic process automation (RPA) to agentic process automation (APA)

Key Highlights

• Computer use agents represent a new approach to enterprise automation, AI that operates software through the screen, the way a person does.
• They are especially promising in financial services, where older systems without APIs have kept large volumes of manual work out of reach of conventional automation.
• An expanding ecosystem of startups is developing these technologies across the model, infrastructure and application layers.

While much of the conversation around AI focuses on generating text, images and code, its greatest impact in financial services may come from something less visible but far more consequential: transforming operations. Across large enterprises, employees still spend significant time navigating legacy systems, vendor portals and disconnected applications that were not designed to work together. AI has the potential to streamline these workflows, reducing manual effort and helping information move seamlessly across organizations so teams can focus on higher-value work.

For two decades, Robotic Process Automation (RPA) promised to address this. The idea was simple: record a person's clicks and keystrokes and have a bot play them back. The market grew to $4 billion, and major companies deployed hundreds of bots to automate manual processes.

As with any new technology, there is a catch. RPA can be brittle, creating hidden costs. The majority of enterprise spending on RPA, it turns out, is not on the software licenses, but instead on the large number of engineers needed to fix the bots when they break down. Those maintenance costs can run into the millions for large enterprises. While RPA showed significant promise, its impact seems to be more incremental than transformative.

The automation upgrade enterprises have been waiting for

Now, a new crop of startups have a technology solution that is poised to deliver on RPA's original promise and go further: computer use agents.

Think of these agents as AI that operates a computer just like a person does. It uses advanced computer vision to see the screen, reason about what's on it and then act by clicking, typing and scrolling to complete a task. Unlike an RPA bot that follows a fixed script, a computer use agent can understand the context of a screen. If a button moves, the agent can find it. If the layout changes, it can adapt.

This is the new paradigm. Computer use agents represent a significant evolution beyond traditional RPA, combining the ability to interact with user interfaces dynamically with AI-driven decision-making. Rather than following only predefined rules and workflows, these agents can interpret the state of an application, adapt to changing interfaces and execute tasks across systems that were not designed to integrate. Their capabilities are powered by a combination of large language models (LLMs), which provide planning and reasoning, and vision-language models (VLMs), which enable them to understand and navigate graphical user interfaces by interpreting what is displayed on screen. This allows them to work more like a human operator, understanding context and taking actions based on what they observe.

Given the versatility of these agents, their disruption may not just affect RPA. The same technology is taking aim at the much larger, $227 billion Business Process Outsourcing (BPO) industry. For decades, the BPO model has relied on the simple math of labor arbitrage: hiring workers in lower-cost countries for repetitive, screen-based tasks. Computer use agents could pose a threat to make that math obsolete. Much of the BPO industry runs on the very work that agents are designed to do: data entry, invoice processing, order management and navigating the same legacy software that was never built for automation.

Modernization without migration

The flexibility of computer use agents could unlock the single biggest challenge for digital transformation in financial services: legacy systems. Some banks and insurance companies run on older core systems, many written in languages like COBOL that few programmers know today. These systems may hold records for the most critical transactions, but they were built long before the era of APIs. Modern AI, which relies on APIs to connect to data, cannot reach them.

The traditional answer has been a "core replacement" that can involve multi-year, high-risk and expensive projects to replace old systems. Computer use agents offer an alternative, a modernization without migration.

Since an agent interacts with the screen, the screen itself becomes the API. A 40-year-old terminal application suddenly becomes accessible to a state-of-the-art AI model without a single line of code being rewritten. This allows institutions to keep their stable, certified systems of record while incorporating modern AI capabilities. For an industry that has deferred core replacement for a decade, this is a game-changer.

Top use cases in financial services

The most immediate opportunities are in high-volume, repetitive workflows:

• Client onboarding and KYC: Pulling data from various registries and portals into case files
• Credit card processing: Validating documents, running credit checks and creating customer records
• Anti-money laundering (AML): Assembling evidence from multiple systems for an analyst to review
• Disputes and chargebacks: A multi-system, document-heavy process with tight deadlines

Computer use agents could extend beyond automating tasks, but for now the startups involved are helping shape the market for enterprise adoption of these solutions.

Ecosystem and value chain

The market has organized into four layers with very different economics, and the companies building in each layer face very different competitive dynamics.

1. Foundation models supply perception and reasoning. Major frontier model owners all ship computer use natively, priced per token and falling. Raw screen navigation looks like it is becoming a base-model feature rather than a product.
2. Infrastructure provides cloud browsers, virtual desktops and session management that the agents run on (e.g. Browserbase and Steel). This is a consumption-priced layer, and among the more durable positions as usage scales with adoption regardless of which agent wins.
3. Orchestration and frameworks handle planning loops, action abstraction and error recovery. Open-source projects including Browser Use and Skyvern have found real developer adoption, though much of what they do is converging with what the labs now provide natively.
4. Applications own a workflow end-to-end. Durable advantage is most likely to sit here, because the hard part is not navigating a screen but knowing how a particular institution works: its terminology; its escalation paths; its definition of a correct output.

The market is evolving from browser agents to full computer-use agents. Browser-focused platforms built by startups such as Yutori, MultiOn and Fellou excel at automating workflows within web applications. Meanwhile, startups like Simular and Zomma are extending automation beyond the browser, enabling agents to interact with desktop software, terminal applications, documents and legacy systems. This broader reach is particularly relevant for large enterprises, where critical processes often span both modern web platforms and older operational systems

The RPA incumbents are also moving toward the same territory: UiPath, Automation Anywhere, SS&C Blue Prism and Microsoft Power Automate have all repositioned around agentic automation. Meanwhile, AI-native entrants like Sola and Automat are also staking their claim, bringing purpose-built approaches that treat agentic process automation as a first principle rather than an afterthought.

Among the emerging companies, several illustrate distinct strategies:

• Simular was founded by former DeepMind researchers and is focused on what it calls "autonomous computers." Its agents can operate desktop applications, browsers and enterprise software environments, and the company has demonstrated leading performance on computer-use benchmarks. The flagship product, Sai, runs in a dedicated virtual environment and is designed to execute end-to-end workflows with human oversight and approval controls.
• Narada, built by researchers out of UC Berkeley, runs "large action models" - agentic AI systems that plan, orchestrate and execute complex business workflows across enterprise apps such as ServiceNow, SAP Concur, Salesforce and collaboration tools like Slack and Outlook. Narada enables enterprises to automate multi step tasks without scripting or configuration, using a combination of API calls and UI automation.
• HCompany, which was also founded by former DeepMind talent, is focused on "action-oriented" AI and recently launched managed computer-use agents built on its Holo model family, allowing enterprises to deploy agents that can interact directly with browser and desktop environments.
• Browserbase is a cloud platform that provides AI agents with scalable, programmable browser infrastructure to automate and interact with websites like a human.
• Sola could be the most direct threat to traditional RPA, letting ordinary business users record themselves doing a task and turning that recording into an editable and reliable agent.

The path forward

Computer use agents could shift the economics of automation. Processes that were once too variable or low-volume to justify the cost of brittle RPA could be prime candidates for this new, flexible approach.

As AI moves from advising to acting, the critical questions are no longer about capability, but about accountability. How do we identify a non-human actor? What is it authorized to do? How are its actions audited, and who is responsible when something goes wrong?

While enterprises work through these critical issues, the overall value of computer use agents is becoming clearer. At their core, these agents have the potential to transform enterprise automation, thanks to a host of startups that are developing the technology to enable wider adoption.

For more information, email Vibhor Rastogi at [email protected], Jelena Zec at [email protected] or Maria Bodiu at [email protected].

To learn more about Citi Ventures and our portfolio, click here.