From RPA to Enterprise Agents: The New Era of Workflow Automation

Making low-code, personalized agents accessible to non-technical users and scalable for enterprise needs

A Resurgence of RPA to Agentic Workflows

Robotic Process Automation, or RPA, emerged to automate high-volume, repetitive, and rule-based tasks typically performed by humans. The promise of RPA was to reduce manual work, improve cost savings, and accelerate workflows through rule-based tasks across systems. It originated from technologies like screen scraping and UI testing automation in the 1990s. As business needs grew more complex and demand for operational efficiency rose, RPA included integrations across enterprise systems, setting the stage for broader adoption.

For years, RPA companies like UiPath, Blue Prism, and Automation Anywhere dominated the enterprise automation landscape. However, these platforms have struggled to evolve beyond their core limitations: high implementation costs, brittle workflows tied to UI changes, and heavy reliance on non-intelligent scripts. At the end of the day, only large enterprises could afford to use RPA for select, tedious workflows.

The global RPA market was valued at $23 billion in 2024 and is projected to reach $21 billion by 2034, representing a CAGR of 25%. In the US alone, the market is expected to grow from $8 billion in 2024 to $64 billion by 2034. Given recent tech advancements, companies can now significantly reduce operational costs, making this an essential enterprise investment. However, the traditional mental model and terminology of “RPA” is now outdated, as AI agents will penetrate all use cases across enterprises. This article explores the history of RPA, new startups in the space, and where entrepreneurs can build robust horizontal or vertical agents. 

RPA 1.0: The Birth of Legacy Incumbents

In the early 2000s, traditional RPA vendors like UiPath built their platforms during the first wave of automation. Though revolutionary at the time, these solutions suffer from inherent brittleness. UiPath’s business was based on screen-scraping bots, mimicking users' clicks and extracting information. These companies created rule-based automation with predetermined steps for structured data, ultimately reaching limitations. 

RPA 2.0: Low Code, Cloud-Based Horizontal Platforms

The second generation in the mid 2010s brought cloud-native, low-code platforms exemplified by companies like Workato and Zapier. These platforms democratized automation by enabling business users to create workflows without extensive technical knowledge. These low-code platforms leveraged pre-built API integrations and webhooks for more stability. However, they still faced limitations when dealing with legacy systems that lacked modern APIs and integration capabilities.

Learnings on Why Traditional RPA Failed

RPA often fell short of expectations in many organizations due to how it was selected, deployed, and governed:

1. Deployments were too complex: Legacy RPA often requires significant infrastructure investment, training, and specialized developers to build and maintain bots. Any small UI or workflow change in underlying systems frequently breaks bots, requiring constant maintenance.
2. Limited Intelligence:  While RPA excels at structured, rule-based tasks, it struggles with unstructured data, decision-making, and exceptions, requiring human intervention or separate ML add-ons that increase complexity.
3. Siloed Implementations:  Many enterprises end up with fragmented RPA initiatives that are hard to scale, leading to limited ROI compared to the initial promise.

How AI and Browser Automation are Disrupting the Market

Recent advances in AI-powered browser automation leverage LLMs, computer vision, and agentic reasoning to handle tasks that traditional RPA cannot, offering net new capabilities:

1. Resilience to UI Changes at Scale: AI agents can “see” and “understand” web interfaces rather than relying on brittle selectors, making them more robust against layout changes and diverse interfaces.
2. Computer Vision: Solutions powered by computer vision can "see" and identify all screen elements (e.g., buttons, checkboxes, text fields, images) on any interface, even virtual desktop or legacy systems.
3. Handling Unstructured Data: LLM-powered systems can read, extract, and reason over emails, PDFs, and documents, enabling end-to-end automation where legacy RPA would stall.
4. Natural Language Instructions: Users can describe workflows in plain English, and AI agents can interpret and execute them across browser-based workflows, lowering the barrier to automation creation.
5. Scalable and Lightweight Implementation: Browser-based AI automation requires no heavy infrastructure, allowing faster iteration, cheaper deployment, and democratization of automation to non-technical teams.

RPA 3.0: New Agentic Browser Automation

The emergence of AI-native browser automation represents the third wave of “RPA”. Companies like OpenAI's ChatGPT Operator are pioneering browser automation for consumers that leverage multimodal intelligence and reasoning capabilities. Operator can perform complex web-based tasks by understanding visual interfaces and executing multi-step workflows, for example, booking restaurant reservations, making a shopping purchase, or uploading content to websites. Other newer startups like Orby use a novel "large action model" that learns from observing human behavior to generate automations automatically.

Whereas RPA specializes in automating well-defined, repetitive, rule-based tasks, this new wave of workflow automation embeds agentic features: autonomous systems that can perceive environments, make decisions, and take actions to achieve specific goals. AI agents bring cognitive abilities; they can adapt, learn, and make decisions in complex, unpredictable scenarios, enabling automation of processes RPA alone cannot handle.

In our thesis, we believe that automation needs the structure and customizability of traditional RPA, with some backstop deterministic nature and a hybrid human approach. This will help unlock new value for enterprises that have a higher bar for accuracy and compliance than fully allowing agentic systems to take over. This new generation of “AI RPA” will handle unstructured data, exceptions, and dynamic scenarios, alongside layered guardrails, rules, and bridges to legacy systems. We’re excited about the democratization of LLMs for non-technical users in work settings to create purpose-built agents for their own unique workflows. 

Market Landscape and Opportunities 

The Current State of the Market

We’ve already started to see a new wave of Horizontal RPA platforms like Luminai, Orby, Induced target a variety of industries across healthcare, financial services, government, and retail sectors. These startups can be used for a variety of internal efficiency use cases. 

Some example use cases:

• Document Processing - Uploading or importing documents to execute downstream tasks (filing claims or updating CRMs)
• Reconciliation - Automating accounts receivable reconciliation to improve finance team efficiency
• Data Cleaning - Migrating legacy data or contracts into new systems
• Validation - Validating claims data, customer policies, or customer requests

In addition to these horizontal platforms, Vertical Agents for industries (healthcare, finance, insurance, logistics) or functions (customer support, finance, engineering) have appeared. In financial services, companies like Hebbia can automate the extraction of earnings calls to surface strategic priorities and M&A recommendations for analysts. Compliance agents like Greenlite serve as copilots for financial crime analysts to automate manual lookups associated with enhanced customer due diligence and transaction monitoring. Within insurance, new startups like Kay and Further streamline submissions processing, policy generation, and quote comparisons. Within the healthcare sector, companies like Tennr help digitize legacy workflows like faxes sent for medical records or insurance information.

Where Startups Can Build and Differentiate

At Montage, we’re curious to find entrepreneurs who are building agentic workflows to meet the needs of enterprise-level business users across horizontal or vertical, regulated industries. 

A horizontal play: Allowing non-technical users to create their purpose-built agents

The horizontal automation market remains vast and underserved. Despite significant growth, Deloitte reports that only 3% of organizations have successfully scaled RPA to more than 50 robots. This gap represents an opportunity for platforms that can solve the scalability, maintenance, and integration challenges that plague current solutions.

While the RPA 2.0 incumbents are already re-framing their positioning as adding “AI agents” for flexible workflows and incorporating LLMs for browser automation, there is still a significant amount of UX improvement and design considerations. These platforms should provide a seamless user experience, bridging browser, desktop, and application workflows, with enterprise-grade governance and control. As of today, many browser automation workflows are still slightly slow with low latency. We expect over time for agentic processes to be able to control user environments, rapidly executing clicks and actions intelligently, taking into account goals and reflective learning.

Unlike vibe coding companies like Lovable or Bolt that are built for entrepreneurs or slightly technical users who can prompt their way to an outcome, enterprise-grade workflows require another level of customization, accuracy, and security. We’re excited about how non-technical users can prompt and fine-tune workflows tailored to their own needs, easily customizing their own off-the-shelf agents. 

The rise of vertical agents for regulated industries 

Vertical-specific automation solutions offer even more compelling opportunities due to their ability to deliver deep domain expertise and pre-built industry workflows. For example, across our main sectors at Montage, core use cases include:

• Healthcare: Healthcare revenue cycle management represents a particularly attractive vertical, with organizations reporting 30% reductions in claim denials and significant improvements in cash flow cycles through automation.
• Financial services: Automation in areas like regulatory reporting, customer onboarding, and risk management. Legacy institutions in these sectors are particularly motivated to invest in automation solutions that can modernize operations without requiring complete system overhauls.
• Insurance: Other promising verticals include insurance claims processing, quote comparison, and underwriting for brokers, MGAs, and carriers.

In contrast, what horizontal agents gain in generalizability, they often sacrifice consistency. Currently, most target simpler productivity or e-commerce use cases as they work towards enterprise-grade performance. Without the benefit of a more constrained problem space with appropriate data scaffolding and guardrails, more dependable browser agents must overcome key challenges, including managing complex action and observation spaces, maintaining context across multiple pages, and interpreting diverse web interfaces. The convergence of AI capabilities with enterprise automation needs creates compelling investment opportunities across verticals.

Tech Differentiation: Large Action Models

At Montage, we look for core differentiators in the team. We believe technical teams that can build both robust, best in class AI architecture along with distribution/data flywheels will win. Diving deeper into competitive differentiation, we’re looking for products and tech teams that prioritize the following pillars: 

• Large Action Models: Large Action Models are designed to translate user requests into executable actions, bridging the gap between understanding and doing. LAMs can process various types of input including text, images, voice commands, and user interactions. They use natural language processing techniques to extract key information and infer user intent from these diverse inputs, breaking the task down into smaller sub-actions. In the future, prior RPA processes won’t need to perform “scheduled runs” at times of the day; agents will be constantly monitoring and executing actions in the background. LAMs will continue to execute in real time, excel at action hierarchy and planning, and continually learn/adapt to improve performance over time.
• Explicit User Feedback and Intent: Teams must design systems that learn from the actions of the users to make well-rounded judgments in an autonomous, explainable manner. Agents can map out high or low confidence paths that prior RPA systems struggled with and needed binary flows. In addition, systems must have the ability to continuously learn from human feedback.

• Robust Reliability across Platforms: While many providers use Document Object Model (DOM) and computer vision to understand web pages at a pixel and structural level, more advanced platforms can navigate dynamic interfaces and JavaScript-heavy applications. However, browser automation requires significant computational resources and can be sensitive to network latency and performance variations. Overtime, we’re looking for consistent throughput from multi-step processes spanning browser, applications, databases, and systems.
• Dynamic Workflows and Intelligent Orchestration: The most advanced automation platforms combine multiple technologies—browser automation, API integration, AI reasoning, and workflow orchestration—into unified solutions. These systems can dynamically choose the optimal automation approach based on the specific task and available integration options. Dynamic workflow platforms excel at complex, multi-system processes that require decision-making and exception handling. They represent the future of enterprise automation but require significant technical sophistication to implement and maintain effectively.

Additional Considerations

• Becoming services-based to set up workflows: In prior RPA companies, employees would map out detailed workflows to automate, sometimes taking months of consulting services and setup. While the new generation of AI-native companies are creating system that can observe human interactions to create dynamic workflows, companies must strike the balance of consulting-oriented sales process to pinpoint suitable use cases, set up the right onboarding, and have human in the loop monitoring to ensure success. Streamlining the onboarding process and making the workflows as easy to set up via intuitive UI/UX is paramount.
• Production-level Quality at Scale: Productionizing workflows takes a level of quality and confidence for users. Productionizing a model that automates workflows in enterprises will likely be executing thousands of workflows a day. Reaching a high level of 90%+ accuracy is paramount, where the actions the model takes needs to be correct. Companies should create clear ROI stories that uncover accuracy against existing employees/solutions and existing models. 

Conclusion

Intelligence lies not merely in the answers produced, but the methodology and process in how they were developed. The market for the prior generation of RPA companies is being quickly disrupted by new functional and vertical agents. We believe there is still ample opportunity for vertical AI agents to serve regulated enterprises that require nuanced knowledge and integrations into existing workflows. We are also excited about the democratization of flexible, intelligent agents to serve non-technical, end business users across teams through seamless onboarding and mapping of workstreams. If you’re building in any of these areas, please reach out to connie@montageventures.com