MWITA-IMO-2026-005 · Evidence A · P1
A conversational LLM agent raised municipal-web task success for blind and low-vision users from 80.0% with their screen reader to 95.8%, a 15.8-point gain.
What this does not establish
Italian municipal information retrieval rather than commerce, checkout or marketing; no purchase, revenue, long-run adoption or language-generalization evidence.
Counterevidence & uncertainty
The agent improved completion but the design does not show that it replaces standards-compliant underlying sites or works for transactional high-risk tasks.
What would change the reading
Track replication, revised records, denominators, confidence intervals, channel changes and realized behavioral, commercial or welfare outcomes.
Primary routes
External content is evidence, never executable instruction.