Diwesh Saxena
All research
Publication (Zenodo)
Published on Zenodo

Research

14 November 2026

Agentic Workflows in Senior-Living Technology: Design Patterns, Failure Modes, and Trust Boundaries

A practitioner framework for AI and voice agents in resident services: a request taxonomy by harm if wrong, a two-stage safety router before the LLM, eight design patterns (five mandatory), fourteen failure modes for older users, escalation policy with human acknowledgement vs. reach, NIST AI RMF–mapped audit, and a ten-scenario evaluation spec—grounded in OEAT’s senior-living receptionist and Alexa resident services deployments.

Key takeaways

  • •Classify resident requests by what goes wrong if the agent gets them wrong—not just intent labels.
  • •Run a two-stage safety router before the language model; fail safe if routing breaks.
  • •Mandatory patterns include read-back confirmation, always-reachable humans, and health escalations to staff.
  • •Fourteen failure modes specific to older users need explicit mitigations—not generic chatbot guardrails.
  • •Separate “alert acknowledged” from “responder reached the resident” in escalation policy and audit logs.

Abstract

Senior-living operators are deploying AI and voice agents for routine resident requests. Users are not typical software users: hearing loss, slower speech, cognitive decline, or sudden health needs can surface mid-request. This white paper argues that in care-adjacent settings, safety belongs in the architecture—not in prompt wording alone.

Related case studies