Performance, Scalability and Maintenance: An Experimental Study of RPA vs. Autonomous Agents in the Healthcare Domain
B · Useful, limited A University of Santa Cruz do Sul design-science study with confusion matrices, cost formulas and a test-retest; clear and checkable, but one portal, 69 patients and one small model limit how far it travels.
Summary based on the full text·Original: about 30 minutes to read
In short
An arXiv paper (2 November 2026) by Daniel Schulz and colleagues at UNISC, Brazil, with a mastology clinic. Both a Playwright RPA script and a browser agent (browser-use, LangGraph, Gemini 2.5 Flash-lite) retrieved mammography and breast ultrasound reports from a laboratory portal. On a stable interface, RPA was more accurate, faster, cheaper and more consistent.
Key numbers
- 98.90% vs 96.49%accuracy of the RPA script vs the browser agent across two rounds
- 16.07 s vs 207.21 smean execution time per patient, about 13 times faster for RPA
- USD 0.030 vs 0.187cost per report, about 6.2 times higher for the agent
- 0.44 vs 3.51 pointschange in accuracy between rounds a week apart, RPA vs agent
As reported by the source.
English·Web page·opens arxiv.org
What it reports
- 69 patients (71 exams) processed in five daily batches, run twice, at T and T+7.
- Accuracy 98.90% vs 96.49%; precision 96.60% vs 92.57%; recall 100% vs 96.48%; F1 98.27% vs 94.48%.
- Mean time per patient 16.07 s vs 207.21 s; cost per report USD 0.030 vs 0.187.
- Both reached 100% login, patient identification and complete download rates for the 65 registered patients.
- Test-retest variation: RPA under 1.5 points on all metrics; agent 3.51 (accuracy) to 7.04 (recall). Agent recall was 92.96% at T and 100% at T+7.
- RPA false positives came from an iframe selection bug, fixed afterwards.
- The agent design kept filtering, extraction and duplicate control in deterministic tools and a lean prompt.
How they got it
Design Science Research: both solutions built as artifacts and run on the same patients, batches and gold standard, with confusion matrices summed over two rounds and cost computed from infrastructure plus LLM tokens.
Limits
- One portal whose layout did not change during the study, so the agent's claimed advantage on changing interfaces was not tested.
- Small sample; one low-cost model for the agent.
- The cost split between API calls and infrastructure is left for future work.
Why it matters for companies
For stable, repetitive web tasks, a deterministic script still beats an agent on every measure here, including consistency between runs. Agents earn their cost where interfaces change often or vary across sites, which this study did not test. Even when an agent is used, keeping the critical checks in deterministic code reduced its errors and token use.
Summary based on the full text. Read the original for the full argument and data.Read the original on arxiv.org