Agentic AI vs RPA: why classic robotization failed in healthcare
Fragility to interface changes, absence of reasoning, maintenance cost: the limits of RPA.
Classic RPA enjoyed real momentum in healthcare between 2018 and 2022, followed by a wave of disappointments. Many projects launched with strong promises were abandoned or reduced to symbolic scope.
The cause is structural, not cyclical. Classic RPA works by recording deterministic scenarios: the robot reproduces exactly the clicks and entries a human would make. This model is fragile as soon as the environment moves.
And healthcare portals move constantly. Insurer operators modify their interfaces several times a year. Public payer services evolve. Core systems update regularly. Each change breaks RPA scenarios, which must be re-recorded by hand.
Maintenance cost quickly becomes prohibitive. On a fleet of 30 to 50 robots covering insurer portals of a group, a full FTE is typically dedicated to maintenance. Initial ROI evaporates.
RPA also doesn't reason. Faced with an unforeseen case (an unusual error message, an extra field, a wording change), it stops or commits a silent error. This rigidity drastically limits exploitable use cases.
The consequence is limited coverage. Robots work on the 60-70% standard cases and leave the remaining 30-40% to humans. Yet these 30-40% often concentrate the bulk of administrative time.
Agentic AI works differently. It understands the objective, adapts to the interface it encounters, reasons on unforeseen cases, and asks for human help when uncertain.
This adaptability changes the game on three axes. Coverage goes from 60-70% to 85-95% of cases. Maintenance cost collapses, because the agent absorbs interface changes without intervention. And reliability increases, because the agent recognizes its own limits.
It is this difference in nature, not a simple precision gain, that makes AI agents viable in healthcare production. We are not talking about a slightly improved RPA, but a fundamentally different architecture.
For operators having already invested in RPA, transition to agentic happens progressively: replacing the most maintenance-costly robots first, expanding scope next. The RPA investment is not lost: it serves as a knowledge base to calibrate agents.