RPA vs. AI agents: where the line actually is

Most vendors selling automation now call it "AI agents," whether or not anything in the product actually reasons about anything. The distinction is not a branding choice. It is architectural, and it determines what the system can and cannot handle before you buy it.

What RPA actually is

Robotic process automation executes a fixed script against a defined input: click this field, read this cell, copy this value into that system. The script has no understanding of what it is looking at. It follows coordinates and rules, not meaning.

That is precisely why RPA is reliable for stable, repetitive, high-volume work and why it breaks the moment the input changes: a relabeled field, a reformatted invoice, a moved button. The fix is always the same — a person rewrites the script.

What an AI agent does differently

An AI agent is given a goal, not a script. An LLM reasoning layer interprets the input, decides what steps to take, and can act differently across two runs if the circumstances differ.

This is the real capability gap: roughly 90% of enterprise data is unstructured — emails, scanned documents, chat logs, free-text fields — and that is exactly the category a scripted RPA bot cannot touch without a human pre-defining every possible variation in advance.

An agent can read a document it has never seen in that exact format before, extract what matters, and flag what does not fit rather than simply failing.

Source: McKinsey; Deloitte

Why "agents replace RPA" is not quite right

The honest version is less dramatic than the marketing version. RPA remains the cheaper, more predictable, easier-to-audit choice for steps that never vary — moving a fixed value from system A to system B, on a schedule, every time.

Rewriting that step as an agent adds cost and reasoning variability the task does not need. Agents earn their cost specifically where judgment, exceptions, or unstructured input dominate the process.

Agents are not a guaranteed upgrade either. Gartner projects that more than 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls.

Source: Gartner Newsroom, June 25, 2025

The realistic pattern industry analysts describe is hybrid, not sequential: RPA handles the deterministic legs of a process, an agent handles the judgment call in the middle, and RPA (or the agent itself) executes the resulting structured action. Treat agent adoption with the same due diligence as any other automation investment, not as a default replacement for a working RPA script.

A quick way to tell which one you are being sold

Ask one question: what happens when the input does not match the expected format? A system that fails, throws an error, or needs to be reprogrammed is RPA, whatever it is called on the pricing page.

A system that reasons about the mismatch, adapts, or intelligently flags the exception for review is an agent. Everything else in the sales conversation is downstream of that one answer.

This is also the line Lunnoa is built around. The platform is AI-native rather than a scripted-automation layer with AI features added on top, so the judgment-heavy steps — reading a document that does not match a template, deciding how to handle an exception — run through the same self-hosted reasoning layer as the rest of the workflow, on the organization's own infrastructure.