In 2010 I joined Liberty Mutual's claims operation: 200+ processes, 3,000+ users, and a knowledge base drifting apart at enterprise scale. We automated the highest-volume, most error-prone steps with RPA — screen scrapers and rules engines, the finest robots 2010 had to offer. Errors fell 15%. Cycle time fell 22%. The knowledge base served 3,000 people at four nines.
Sixteen years later I design agentic AI systems for a living, and here's my most useful professional secret: it's the same job. The tooling changed. The architecture didn't.
What RPA taught me that the agent era is relearning
1. Automate the redesigned process, not the broken one. The biggest RPA failures I saw were faithful automations of workflows that shouldn't have existed. We ran Kaizen first, fixed the flow, then automated. Today's version: pointing an agent at your inbox chaos doesn't organize the chaos — it accelerates it.
2. Knowledge architecture beats model quality. Our SOP base was versioned, governed and compliance-checked, and it's the reason automation held up under audit. Today that discipline is called context engineering and RAG hygiene, and it still decides more outcomes than the model card does. The robot intern reads whatever binder you hand it — make sure the binder is true.
3. Observability isn't optional. We replaced manual status collection with live metrics on cycle time and error rate — that instrumentation is what turned "improvement" from an anecdote into a number. If your agent fleet doesn't report completion rate, cost per task and escalation rate, you're not running a system. You're running a vibe.
What's genuinely new
Reach. RPA could only touch the structured world — forms, fields, screens. LLMs read contracts, drawings, emails, the unstructured 80% of a business that automation never could. That's why I can point an agent at takeoff analysis and submittal review at a construction company, work that would have been science fiction to 2010 me.
But reach without the old disciplines is just a bigger blast radius. The teams winning with agents in 2026 look suspiciously like the teams that won with RPA in 2012: process first, knowledge governed, everything measured, humans at the gate. The infrastructure underneath can even be humble — ask the beige box.
I've been doing "AI transformation" since before it had a marketing budget. The robots got smarter. The job description didn't change.