Agentic RPA: When Bots Stop Following Instructions and Start Making Decisions
Traditional RPA has a reliability problem that the industry talked around for years. When it worked, it worked beautifully — fast, tireless, perfectly consistent. When it broke, it broke in the most frustrating way possible: silently, on an exception nobody had anticipated, often discovered only when a queue had been filling for hours and someone finally noticed.
The root cause was always the same. RPA bots execute instructions. They don't understand what they're trying to accomplish. Click this pixel, read this field, paste this value. If the button moves two pixels to the left in a UI update, the bot fails. If an invoice arrives in an unexpected format, the bot fails. If a rule needs context to apply correctly, the bot fails and waits for a human. Traditional RPA has project failure rates between 30–50%, with 70–75% of total cost going to maintenance rather than new automation. Those aren't numbers you can build a long-term automation strategy around.
Agentic RPA replaces the instruction-following model with intent-based execution, and it changes the calculus entirely.
What "intent-based" actually means in practice
Instead of "click at coordinate (247, 891)," an agentic RPA system receives "submit the invoice for approval." The LLM layer interprets the intent, identifies the relevant UI element semantically, and executes — regardless of where that element appears on screen or what it's called in the current version of the application. When a step fails, the agent doesn't stop and raise an exception. It reviews its action history, assesses what's visible, and tries a different approach. Self-healing isn't a feature — it's a consequence of the system understanding what it's trying to do rather than just how it was told to do it.
The exception-handling problem — historically the biggest single drain on RPA ROI — becomes tractable. Where a traditional bot would queue an invoice it couldn't parse, an agentic system reads the document, extracts the relevant fields using intelligent document processing, and routes it appropriately. UiPath put this directly: agents now handle the exceptions that bots couldn't, within the same orchestration platform the bots already run on. The governance infrastructure — audit trails, approval gates, human-in-the-loop controls — transfers directly. The execution logic becomes dramatically more capable.
What's already shipping
The market moved fast. Automation Anywhere renamed its core product to Agentic Process Automation, anchored by a Process Reasoning Engine that predicts workflow steps with 90% accuracy. Early production results are striking: Petrobras reported $120 million in savings within three weeks; Boston Children's Hospital cut administrative burden by 80%. AI bookings at Automation Anywhere grew 45% year-on-year through 2025, representing over 70% of total business by the end of the year.
UiPath launched its Platform for Agentic Automation in 2025, with Maestro as the orchestration layer coordinating humans, agents, and bots across enterprise systems. In February 2026, it added industry-specific healthcare agents for prior authorisations, claim denial prevention, and medical records summarisation — use cases where structured rules perpetually break on real-world complexity.
Both platforms now use MCP to maintain persistent process context across tools, enabling agents to recover from failures with full awareness of what happened earlier in the workflow. That state management capability is what separates production-ready agentic RPA from demos that look impressive and break under real conditions.
The governance question nobody should skip More autonomous systems require more deliberate oversight — not less. Governance-as-code is emerging as a standard pattern: audit trails, approval gates, and rollback capabilities built into the agent architecture rather than bolted on afterward. Layering intelligent agents onto outdated governance structures doesn't produce autonomous enterprise operations. It produces faster, harder-to-diagnose failures.
The organisations getting genuine ROI from agentic RPA in 2026 are the ones who treated governance as a design constraint from day one, not a compliance checkbox at the end.