Are LLM Agents the New RPA? Findings From Enterprise Experiments

Organizations face a critical automation decision: continue investing in proven RPA platforms or shift to AI agents that promise greater flexibility. Marketing pitches suggest AI agents will replace RPA entirely, but controlled experiments reveal a more nuanced reality. Each technology has distinct strengths, and understanding when to use which approach, or combine them, determines whether automation investments deliver reliability or create new operational risks.

Controlled experiments by Průcha and colleagues (2025) directly compared traditional RPA (UiPath) with AI agent-based automation (AACU) across three enterprise workflow categories: data entry, monitoring, and document extraction. The findings challenge simplistic narratives about one technology replacing the other.

The research demonstrates that RPA still outperforms AI agents in speed and reliability for repetitive, structured tasks, confirming that mature automation platforms excel where workflows are predictable. However, AI agents show clear advantages in development time (faster to build) and adaptability to dynamic interfaces (handling UI changes without reconfiguration). The promising frontier isn't choosing between technologies, but architecting hybrid systems that leverage RPA's reliability with AI agents' flexibility.

The Experimental Design: Comparing RPA and AI Agents Head-to-Head

Průcha and colleagues conducted controlled experiments across three enterprise workflow categories that represent common automation use cases: data entry (high-volume, repetitive tasks), monitoring (continuous observation and alerting), and document extraction (processing structured and semi-structured documents).

For each category, the research team implemented identical workflows using both traditional RPA (UiPath) and AI agent-based automation (AACU - AI Agent Computer Use). This head-to-head comparison isolated the impact of technology choice while controlling for workflow complexity and environmental factors.

The experimental approach measured four critical dimensions: execution speed (time to complete tasks), reliability (success rate and error frequency), development time (hours to build and configure), and adaptability (ability to handle UI changes or unexpected scenarios). These metrics reflect the real concerns of automation leaders deciding where to allocate resources.

Unlike observational studies that compare different automation projects across organizations, this controlled design ensures that performance differences stem from technology capabilities rather than implementation quality, organizational context, or workflow selection bias. The result is credible evidence about when each technology performs better. Understanding how AI adoption patterns evolve helps organizations make evidence-based technology choices.

When RPA Still Dominates: Speed and Reliability for Structured Work

The experiments revealed that RPA maintains clear advantages for high-volume, repetitive tasks with stable interfaces. In data entry scenarios, where automation bots input standardized information into fixed forms, RPA executed tasks faster and with higher success rates than AI agents.

This performance gap stems from RPA's architecture. Traditional RPA platforms script exact sequences of actions (click field A, type value B, press submit) that execute predictably when interfaces remain stable. There's no ambiguity, no interpretation, no decision-making, just deterministic execution at machine speed.

For mission-critical business processes where uptime and consistency matter most, claims processing, invoice entry, payroll updates, regulatory reporting, RPA's reliability advantage outweighs other considerations. When a workflow handles thousands of transactions daily and every failure creates downstream problems, the mature stability of RPA platforms becomes essential.

The research confirms what many automation leaders intuitively know: don't replace proven RPA workflows with experimental AI agents in systems where reliability drives business value. If a process runs successfully at scale today using RPA, the burden of proof for migration is high.

Where AI Agents Add Value: Development Speed and Dynamic Adaptation

While RPA excels at execution, AI agents demonstrate advantages during development and when handling interface changes. The experiments showed that AI agent workflows could be built significantly faster than equivalent RPA implementations, a critical factor when organizations need to prototype automations quickly or respond to changing business needs.

This development speed advantage comes from AI agents' ability to interpret interfaces and generate action sequences without explicit scripting. Instead of mapping every click and field precisely (as RPA requires), developers give AI agents higher-level instructions ("extract invoice totals from uploaded PDFs") and the agents figure out how to accomplish the task.

Even more significant is adaptability to UI changes. When software interfaces update, buttons move, fields rename, workflows reorganize, traditional RPA workflows break and require manual reconfiguration. The experiments demonstrated that AI agents can adapt to these changes automatically, continuing to function when RPA would fail.

This adaptability makes AI agents particularly valuable for automating systems with frequent updates: SaaS platforms that release new features monthly, internal tools undergoing active development, or workflows that interact with third-party websites beyond your control. In these contexts, the maintenance burden of constantly updating RPA scripts exceeds the value of automation itself.

For rapid prototyping and proof-of-concept automations, AI agents offer another advantage: you can demonstrate value quickly without investing weeks in RPA development. Once a workflow proves valuable and stabilizes, organizations can decide whether to migrate to RPA for production reliability or continue with AI agents if adaptability remains important.

Hybrid Architectures: The Promising Frontier

The research identifies hybrid RPA-AI agent architectures as the most promising approach for enterprise automation. Rather than choosing one technology to replace the other, hybrid systems route tasks to the best-suited automation type based on task characteristics.

One effective pattern: use RPA for core transaction processing and AI agents for exception handling. RPA handles the high-volume, predictable "happy path" at speed and scale. When RPA encounters an unexpected scenario, a form field it doesn't recognize, an attachment it can't process, a workflow state that breaks its script, it escalates to an AI agent that can interpret the situation and decide how to proceed.

This architecture leverages RPA's reliability for the 95% of transactions that follow standard patterns while using AI agents' adaptability for the 5% of edge cases that would otherwise require human intervention. The result is higher straight-through processing rates without sacrificing the stability of mission-critical workflows.

Another hybrid pattern: RPA for backend integrations and AI agents for frontend interactions. Many enterprise workflows involve both structured data exchanges (API calls, database updates, file transfers) and user interface interactions (web forms, desktop applications). Use RPA for the backend plumbing where reliability matters and AI agents for the frontend portions where interfaces change frequently.

Organizations implementing hybrid architectures need workflow orchestration systems that can route tasks intelligently. Simple rule-based logic works: if the task involves a stable interface and repetitive actions, route to RPA; if the task involves UI changes or unstructured inputs, route to AI agents. Tools like n8n, Zapier, or custom workflow engines can implement this routing with minimal complexity.

Practical Implementation: Building Your Automation Portfolio

For organizations building automation portfolios, the research suggests a deliberate approach to technology selection. Start by auditing existing and planned automations, categorizing them by two dimensions: interface stability (static vs. dynamic) and task variability (repetitive vs. adaptive).

High-volume, repetitive tasks with stable interfaces are clear RPA candidates. Think: daily batch processing, standard form completion, scheduled report generation, routine data transfers. These workflows benefit from RPA's speed and reliability, and the development investment pays off through consistent execution.

Tasks involving dynamic interfaces or frequent software updates belong in the AI agent category. Examples include: automating workflows in SaaS platforms that release monthly updates, scraping data from third-party websites that change layouts, processing documents with varying formats, interacting with applications undergoing active development.

For new automation projects, consider a maturation path: start with AI agents for rapid prototyping and proof of concept, then migrate to RPA once the workflow stabilizes and volume justifies the additional development effort. This approach balances speed-to-value with long-term reliability.

Establish governance protocols that recognize RPA and AI agents serve different use cases. Avoid blanket mandates to replace RPA with AI agents or vice versa. Instead, create decision criteria based on task characteristics and let teams choose the appropriate technology. Monitor automation portfolio performance and maintenance overhead by technology type to refine selection criteria over time. Organizations should also consider broader workforce implications when deploying AI automation at scale.

For teams without RPA expertise, AI agents offer a lower barrier to entry, you can build useful automations faster with less specialized knowledge. For teams with mature RPA capabilities, preserve that investment for mission-critical workflows while exploring AI agents for new use cases where adaptability matters more than raw speed. Success depends on matching technology capabilities to actual labor market needs.

Real-World Examples: How Organizations Navigate the RPA vs. AI Agent Decision

A global insurance company with 200+ RPA bots handling claims processing faced pressure to modernize when AI agents emerged. Instead of wholesale replacement, they conducted a workflow analysis: 80% of automations were high-volume, stable processes where RPA excelled. They preserved these but introduced AI agents for two specific use cases: handling claims with unstructured attachments (medical records, photos) where RPA struggled, and automating workflows in a new digital platform with frequent UI updates. The hybrid approach maintained 99.2% uptime for core RPA processes while reducing development time for new automations by 40%.

A 50-person e-commerce retailer needed to automate order processing but faced challenges with traditional RPA: supplier platforms changed UIs frequently, and they lacked RPA development expertise. They deployed AI agents for supplier data extraction (handling dynamic interfaces) and used simple API-based integrations for stable systems like Shopify. AI agents reduced setup time from an estimated 6 weeks (for RPA) to 2 weeks and automatically adapted when supplier platforms updated. For high-volume order processing, they eventually added a lightweight RPA layer once the workflow stabilized, demonstrating a pragmatic hybrid approach that matches technology to task maturity. Understanding how AI tools shift work patterns helps organizations design effective hybrid automation strategies. When planning deployments, consider adoption patterns and change management to maximize technology value.

Read the full research: Are LLM Agents the New RPA? A Comparative Study with RPA Across Enterprise Workflows (arXiv:2509.04198)