Why Agentic RPA Outpaces Pure AI Agents in Enterprise Adoption

UiPath released a report in January 2025 surveying 252 U.S. IT executives. These executives work at companies with over $1 billion in annual revenue. The report contains some interesting numbers.

90% of respondents believe their business processes could be improved through agentic AI. 77% say they are prepared to invest in agentic AI this year. 37% are already using agentic AI, and 93% express strong interest in exploring the technology.

These numbers sound optimistic. But another set of data tells a different story.

According to Gigster's research, the full deployment rate of enterprise agentic AI is only 11%. Despite pilot projects surging from 37% in Q4 2024 to 65% in Q1 2025, the percentage of actual production deployments has barely changed.

McKinsey's research is even more direct. 92% of CIOs have initiated generative AI projects, but few have realized a return on investment. In McKinsey partner Brendan Gaffey's words: AI is everywhere except in the P&L.

This is the real state of enterprise AI automation today. High enthusiasm, difficult implementation.

Why Pilots Succeed But Production Fails

Between pilot and production, there are three major obstacles.

The first is system integration. Enterprises average over 175 business systems. Many are legacy systems without modern APIs. AI agents need to interact seamlessly with these systems, but most simply aren't ready for AI.

IBM MIT AI Lab research scientist Shae Khan puts it clearly: AI tools may eventually replace some RPA deployments, but RPA can be cheaper and faster to deploy, while being less prone to errors than most AI tools.

The second is security and access control. AI agents need sufficient permissions to complete tasks, but enterprises are understandably cautious about granting AI access to sensitive data and critical systems. Reuters Institute found that 48% of major news websites are blocking OpenAI's crawlers. Cloudflare released tools last year allowing websites to easily block AI agents.

The third is agent operations. Unlike traditional software, AI agents require continuous monitoring, debugging, and validation. The industry lacks mature frameworks and best practices for managing AI agents at scale. An agent that runs perfectly in a test environment may encounter various issues in production.

Determinism: The Core Reason Enterprises Choose Agentic RPA

Max Ioffe, Director of the Global Intelligent Automation Center of Excellence at Wesco Distribution, made a key statement: For larger scale processes, you need clear orchestration and governance, and that means a deterministic technology like RPA.

This gets to the heart of the issue.

The problem with pure AI agents is they're not deterministic enough. They can hallucinate. They can go off track. OpenAI's Operator demonstrated this in actual testing. When you're using AI agents to handle bank transfers or medical records, "might make mistakes" is unacceptable.

RPA is different. It executes according to predefined scripts, working the same way every time. Arjun Bali, staff data scientist at Rocket Mortgage, summarizes it well: RPA is still relevant for automating rule-based, repetitive, and redundant tasks, especially in industries where there is a big downside for an error like banking, insurance, and healthcare.

This is why agentic RPA has become the enterprise choice. It combines RPA's determinism with AI's intelligence.

Barclays in Practice: Minimum 30% Benefit

Sundar Ganesh, Director of Global Process Automation at Barclays, shared their experience.

Barclays established its Process Automation Center of Excellence (PACE) in 2019, providing end-to-end delivery of RPA, machine learning, and workflow tools. Ganesh says they've been able to operate on a minimum 30% benefit, including effort reduction and customer experience improvements.

Now Barclays is working with UiPath to explore how agents can unify fragmented workflows, particularly in complex processes like mortgages.

Ganesh emphasizes a key point: The differentiator is not the infrastructure, it's the amount of automation we can deliver and deploy.

They developed an approach called "process parties," different from traditional hackathons. First there's a period of deciding which processes enter the "party," then a few weeks of preparation working with governance, operations, and tech teams, then three days to execute, and finally moving to production.

Ganesh says: This is a real problem with real capabilities that we have. It's not an ideation, but to move things into production.

Fiserv Case: 12,000 Hours Saved, 98% Straight-Through Processing

Fiserv is a leading global fintech company. They used UiPath's orchestrated approach to streamline merchant category code determination and validation.

This process was previously done manually and was very time-consuming. They used UiPath to orchestrate robots for simple data extraction and web search cross-checking, then deployed generative AI and agents to update and select the correct codes.

The results: 12,000 hours saved through these automations, 98% straight-through processing rate. The remaining 2% where agents weren't confident were handled by humans checking and updating results.

This case demonstrates the correct way to use agentic RPA: robots handle repetitive work, agents handle reasoning, humans focus on oversight and decisions requiring trust.

Controlled Agency: Adjustable Autonomy Levels

UiPath CPO Graham Sheldon introduced a concept called "Controlled Agency."

The core idea: orchestration allows organizations to deploy agents where they're confident in the outcomes, while maintaining the ability to "dial up or dial back the amount of agency" based on performance and trust levels.

This is completely different from pure AI agents' "full autonomy" philosophy. What enterprises need isn't a fully autonomous AI, but a system where autonomy levels can be controlled.

In mortgage approval processes, this layered approach works like this: UiPath's BPMN-modeled agents and robots handle document extraction, eligibility validation, task execution, and system updates. Humans only get involved when expert judgment is needed.

This approach makes workflows look less like handoffs between systems and more like a coordinated digital team.

IDC Forecast: RPA Spending to Double

UiPath's report cites IDC's forecast: RPA spending will more than double between 2024 and 2028 to reach $8.2 billion.

This number shows that RPA isn't dying. It's growing.

UiPath public sector CTO Chris Radich says: RPA isn't dying — it's evolving. We've tested various AI solutions for process automation, but when you need something to work the same way every single time — without exceptions, without interpretations — RPA remains unmatched.

He predicts agents will eventually control RPA bots, with various robotic processes in a toolbox for agents to choose from. Today, we build separate RPA workflows for different scenarios. Tomorrow, with agentic capabilities, an agent will evaluate an incoming request and determine whether it needs RPA for data processing, API calls for system integration, or human handoff for complex decisions.

The Right Architecture for Agentic RPA

Based on the information I've gathered, the right architecture for agentic RPA should look like this.

The base layer is RPA robots handling structured, repetitive tasks. The middle layer is AI agents handling reasoning and decision-making. The top layer is an orchestration system coordinating robots, agents, and humans.

SleekFlow CTO Lei Gao predicts: RPA will likely be the foundation layer — still valuable, but increasingly invisible, embedded within AI-powered orchestration systems.

McKinsey's Gaffey breaks down this evolution into stages. The first stage is basic individual support tools like copilots, improving efficiency by 10-20%. The second stage is more organized task and workflow automation, with discrete processes seeing around 30% ROI. The third stage is automating whole domains like HR or finance, unlocking bigger benefits through cross-silo consolidation.

The real transformation happens when automation spans the full customer or business journey. This is where orchestration tools and AI agents can make their most significant impact.

Selection Criteria: When to Use RPA vs AI Agents

RPA2AI Research CEO Kashyap Kompella provides selection criteria.

Use RPA when: the process is repetitive and rule-based; inputs are structured and predictable; no decision-making is required.

Use AI agents when: the task involves unstructured data; ongoing learning and adaptation is needed; judgment calls are required.

Use both when: the process includes structured and unstructured elements; RPA can handle execution while AI manages analysis.

Specific examples include: For data extraction, RPA handles structured documents while AI agents process unstructured ones. For automated email responses, RPA manages templated replies while AI agents personalize content. For customer onboarding, RPA performs data entry while AI agents make decisions based on customer information. For IT support, RPA resets passwords while AI agents diagnose more complex issues.

Key Factors to Consider

When choosing a technology path, several key factors need consideration.

Regarding maturity, RPA has been around for 15 years and is well established in enterprises. AI agents are more experimental technology with few examples of large-scale business deployments.

Regarding reliability, AI agents' higher autonomy means they can hallucinate and go off track when attempting tasks. RPA bots stick strictly to scripts.

Regarding compute cost, AI agents are quite computationally expensive, while RPA software is comparatively lightweight. AI agents can also introduce latency due to LLM inference time.

Regarding API dependency, AI agents depend on API availability for system access. In environments with limited API access, RPA can interact directly with UIs, while AI agents struggle.

Regarding security governance, Gartner predicts that companies with robust governance will experience 40% fewer ethical incidents by 2028. This means governance isn't optional — it's essential.

Final Thoughts

The future of pure AI agents is indeed appealing. Fully autonomous, capable of reasoning, adapting, learning. But for critical enterprise processes, reliability matters far more than creativity.

UiPath CEO Daniel Dines says: As AI systems become more autonomous, enterprises must strike a balance between autonomy and human oversight to prevent unintended consequences and guarantee that AI-driven actions align with ethical, compliance, and legal standards.

This is why agentic RPA has become the enterprise choice. It's not about choosing between RPA and AI agents, but combining both: using RPA's determinism as the foundation, AI's intelligence for enhancement, and orchestration systems for coordination.

If you're considering enterprise automation, I suggest starting from existing RPA investments and gradually introducing AI capabilities. Don't try to achieve fully autonomous AI agents in one step. In banking, insurance, healthcare, and other industries with zero tolerance for errors, determinism matters far more than creativity.

RPA isn't dying. It's evolving.

Further Reading

AI Automation Series

How AI Can Fully Automate Tasks For You: How to use Playwright MCP to let AI control browsers for automation tasks.

This One Tool Alone Can Revolutionize 80% of Your Workflow: Detailed Playwright MCP configuration and practical cases.

Vibe Marketing: Automate Your Marketing Workflow with AI: From Vibe Coding to Vibe Marketing, how to use AI tools for marketing automation.

AI Cognition Series

The E^F Law: Why do some people get poorer the busier they are, while others produce effortlessly? The answer lies in the exponent.

Guide to Creating Demons: How to become a Maxwell's demon in the information age.

Tutorial Series

Prompts as Products, Earn Dollars Going Global: Complete tutorial series on building global products with AI from scratch.