Researchers working with the US Department of Labor and the Colorado Department of Labor and Employment co-designed an LLM assistant that drafts fact-finding questionnaires for unemployment-insurance adjudicators; in a randomized crossover trial with 8 adjudicators on 200 historical claims, it beat historical baselines and was rated useful but did not improve quality or efficiency versus the no-AI control (reported by Magesh et al. / CDLE–US DOL collaboration)
UI adjudicators must draft follow-up fact-finding questions under heavy caseloads, and outcomes can vary by which adjudicator is assigned (reported by Magesh et al.)
Claim details shown in a sandbox prototype interface
LLMs suggest follow-up topics and draft fact-finding questions
Adjudicator selects, edits and finalizes the questionnaire
QA reviewers rate questionnaires blind; drafting time and edits are logged
Human-in-the-loop: Adjudicators select and edit all AI-drafted questions (all 8 said editing remained important); eligibility determinations are fully human.
No meaningful time saving: AI-assisted drafting was a statistically insignificant 4 seconds faster on average (p=0.8) Reported by Magesh et al. · Source (tobin.yale.edu)
All 8 adjudicators rated the tool very or somewhat useful; 75% said they would use it regularly Reported by Magesh et al. · Source (tobin.yale.edu)
AI-assisted fact finding substantially outperformed historical (observational) baselines and may reduce inter-adjudicator variance Reported by Magesh et al. · Source (tobin.yale.edu)
Amgen built Catalyst Copilot in Copilot Studio in ~6 weeks so R&D staff can ask natural-language questions over curated drug-development knowledge (reported by Microsoft / Amgen)
Multi-agent Bedrock system lets plant engineers query IoT data in natural language for bottleneck root cause analysis across 250+ curing presses (reported by Apollo Tyres / AWS)
Bosch Bamberg plant operators use the Shopfloor Agent to diagnose machine errors from stored knowledge and restart production faster (reported by Bosch)
Specialty insurer CFC launched Lane Assist, a live pilot of agentic underwriting in its Cyber team that takes low-complexity submissions from email to quote recommendation in seconds; every recommendation is reviewed and approved by an underwriter before issue (reported by CFC)
Honolulu's Department of Planning and Permitting uses Clariti's CivCheck guided AI plan review to pre-screen residential permit applications for completeness and code compliance before staff review; average time to permit decision fell from 73 to 32.5 days on a small Q1 2026 comparison sample (reported by Clariti / Honolulu DPP)