RPA is not dead yet - it's the rails under your agents

The pitch I hear from every CTO peer over coffee in Rome: "We're going to replace our RPA estate with LLM agents."

Some of them will. Most of them shouldn't.

What RPA actually is

UiPath and Blue Prism aren't just screen-scrapers. After ten years in production at a typical bank, an RPA estate is:

• A process catalogue - every legacy mainframe screen, every undocumented Excel ritual, mapped and named.
• A credentials vault - service accounts for SAP, Murex, Bloomberg, mainframe LU6.2 sessions - auditable, rotated, reviewed.
• An orchestration layer - queues, retries, SLAs, business-hour rules, alerts to the right ops team.
• An audit trail that survives external audits because it was designed to.

You don't replace that with three weeks of LangChain.

Where agents win

Agents are dramatically better at the parts that were always weakest in RPA:

• Anything unstructured - email triage, PDF extraction, customer-language classification.
• Anything that needs reasoning over policy text - "is this claim a fit for our 2024 reinsurance pool?"
• Anything that needs to call dozens of tools in a non-deterministic order.

The agent doesn't replace the RPA bot. The agent decides which RPA bot to run - and reads the output back into a human-readable answer.

The architecture I keep ending up at

User / channel
   │
   ▼
Agent (reasoning, policy, tool selection)
   │
   ▼
Tools = [ RPA process A, RPA process B, internal API, RAG over docs, … ]
   │
   ▼
Source systems (SAP, mainframe, CRM, …)

The RPA platform becomes a tool runtime. The agent becomes the policy layer. Old work isn't thrown out; it gets a new front door.

The temptation to greenfield

Greenfielding feels great in slideware. In production, the RPA estate is where most of the institutional knowledge about "how things actually work" already lives. Replace it slowly, process by process, with the agent reading the existing audit log and making smaller decisions. Don't burn the library to build a chatbot.