AI agents vs. RPA.

RPA platforms have been selling "intelligent automation" since 2019. AI agent platforms have been selling "the new automation" since 2024. Both claim to do basically the same thing: automate repetitive business work. They aren't the same thing, and pretending they are leads to the worst kind of failed deployments: ones that the architecture was never going to support.

This piece is the version we wish more buyers had read before procurement. It's opinionated, because we've built both and watched both fail in characteristic ways.

What each one actually is

RPA (robotic process automation) is software that scripts a fixed sequence of UI actions: click here, type this, copy that value, paste it there, hit submit. It's deterministic by design. Given the same input and the same screen layout, it produces the same output every time. UiPath, Automation Anywhere, Blue Prism, Microsoft Power Automate are the major players.

AI agents are software that uses an LLM (Claude, GPT, Gemini, or open-weight) plus a set of tools plus a reasoning loop to accomplish goals. Non-deterministic by design. Given the same input, they might take different paths to the same outcome, or recover from a failure mode the original developer didn't anticipate.

That difference, deterministic vs. probabilistic, is the source of every other trade-off.

Where RPA wins

• Volume + stability. Processing 10,000 identical invoices through an unchanging SAP screen? RPA does it cheaper and more reliably than any agent.
• Deterministic compliance requirements. When a regulator wants to know the exact sequence of system actions, an RPA script is auditable down to the click. An agent's reasoning is auditable, but the "what would it do" question is harder to answer for unseen inputs.
• Cost at scale. Pennies per execution. Agents cost more per task (inference is real).
• Stable, well-defined legacy systems. A 30-year-old mainframe terminal whose screens haven't changed since 2008 is the perfect RPA target.

Where RPA loses

• Variability. The moment screens change layout, forms add fields, or inputs arrive in slightly different shapes, RPA breaks. Maintenance is the silent killer.
• Unstructured input. Reading a customer email and figuring out what they want? RPA can't. It needs structured fields.
• Judgment. "Is this urgent?" "Does this look like fraud?" "Should this be escalated?" RPA requires every decision to be coded explicitly. Agents read the situation.
• Long-tail variation. The 80% of cases an RPA handles is fine. The 20% it doesn't is where most of the failure modes live. And building exceptions for every variant is how RPA projects accumulate technical debt.

Where AI agents win

• Unstructured input. Emails, chat messages, PDFs, voice transcripts, conversational interfaces. The agent reads, summarizes, extracts intent.
• Judgment-required workflows. Triage, prioritization, escalation, qualification, routing, anywhere a human used to be the dispatcher.
• Variability tolerance. The screen changed? The form added a field? The customer asked the question differently? An agent generalizes; an RPA script breaks.
• Reasoning across systems. When the work requires pulling from CRM, checking inventory, comparing against contract terms, and drafting an answer that integrates all of it, an agent does this naturally. RPA needs a meta-script tying separate bots together.
• Adapting to system changes. Software updates that would break an RPA script often don't faze an agent, it reads what's there, not what was there last week.

Where AI agents lose

• Cost at very high volume. Inference per task is cheap but non-zero. At millions of executions per day on a simple task, RPA is dramatically cheaper.
• Hard determinism. When a regulator demands you guarantee the exact action sequence in advance, agents are a harder sell. Possible, but harder.
• Hallucination surface area. Agents can fabricate. Mitigations exist (grounding, citation, eval harnesses, low-confidence routing) but the failure mode is real, where RPA's failure mode is just "didn't work, crashed loudly."
• Latency for trivial tasks. An RPA bot can fire-and-forget a clicked-through workflow in 200 ms. An agent that has to plan and reason might take 5 seconds. For trivial work this matters.

A decision framework

Ask three questions about the workflow before you pick an architecture:

1. Is the input structured and stable? If yes, RPA. If no, agent.
2. Does the work require judgment, or is every decision explicitly rule-encodable? If rule-encodable, RPA. If judgment, agent.
3. Does the workflow change shape over time? If it's been stable for years, RPA's brittleness isn't a problem. If you're touching a system that ships updates quarterly, an agent is much cheaper to maintain.

Two out of three "yes" for RPA → use RPA. Two out of three "no" → use an agent. Mixed → look more carefully, and consider running both.

The hybrid pattern

The most successful enterprises we work with run both. RPA handles the mass-volume deterministic work where it's mature and cheap. Agents handle the judgment-required variable work where RPA would either fail or accumulate unmaintainable complexity. They share infrastructure (logging, monitoring, secrets management, identity) but are deployed independently.

A practical pattern: an agent at the front of a workflow does the "read intent, qualify, decide what to do" work, then dispatches to an RPA bot for the high-volume deterministic execution work. The agent brings judgment; the RPA brings cheap throughput. Together, they're stronger than either alone.

One real example from our practice: in our wealth management firm engagement, the agents on top use Playwright (which is closer to RPA in posture) to drive a legacy SaaS UI, while the reasoning layer uses an LLM to handle the variability of "what report does this VIP want, from which date range, for which client." Both architectures, one workflow. Use what each is good at.

What this means for procurement

If you're sitting in a room with an RPA vendor and an AI-agent vendor pitching the same workflow, they're often both right and both wrong. The honest answer is workflow-by-workflow. A vendor who tells you "we replace all your RPA with agents" or "agents are just RPA with extra steps" is selling you their product, not advising you on architecture.

The cleanest test: ask either vendor what their solution does when the input is unexpected. The RPA answer should be "logs an exception and stops." The agent answer should be "reads the unexpected input, decides what to do, either handles it or escalates with reasoning." If you hear an RPA vendor claim their bot can "handle anything," walk away. If you hear an agent vendor claim deterministic SLAs on novel inputs, also walk away.

When to talk to us

If you have a workflow that's stuck between an RPA project that keeps breaking and a manual process that keeps growing, that's the sweet spot for an agent. Bring it. We'll tell you honestly which architecture fits, including telling you to keep using RPA when that's the right answer.

More reading: What Is Agentic AI?, How Much Does an Enterprise AI Agent Cost?, Measuring AI Agent ROI. Also: Quantilus services and case studies.