Evidence

Introduces HTA-Bench, an expert-grounded benchmark evaluating end-to-end AI generation of authentic German AMNOG and European HTA deliverables in Word templates. Regulaido achieves the Pareto-optimal #1 aggregate score of 72.8 (outperforming OpenAI Codex CLI at 67.3 and Anthropic Claude Code at 51.5) while completing runs in 26 minutes (vs. 50 min and 34 min) at $0.61 per run (7.6× lower cost than Codex CLI at $4.63 and 3.6× lower than Claude Code at $2.20).

Benchmarked against IQWiG working paper GA23-02 (132 confounder variables, 62 person-days manual effort), Regulaido achieved the deepest granular recall (63%, 83/132 variables) with 0% citation fabrication in under 5 minutes—outperforming general-purpose frontier LLMs (Claude Opus 4.5 at 45–46% and Gemini 3.1 Pro at 29–40%, which exhibited up to 36–40% DOI fabrication without corpus constraints).

Survey of 27 medical writers shows Perceived Usefulness is the only significant driver of AI adoption (β=0.522, p=0.028), while trust in AI does not predict professional identity threat (β=-0.032, p=0.901). With 96% using Regulaido and strong regulatory compliance confidence (M=6.22/7), successful GenAI integration depends on clear regulatory guidance and hybrid workflows that preserve strategic judgment while leveraging AI for routine tasks.

Analysis of 12,500 messages from 32 HTA writers revealed that AI assistants function as specialized toolboxes—primarily used for Information Extraction & Retrieval (27.2%) and Text Generation & Drafting (17.2%), focusing on Clinical Study Documentation (22.5%). User engagement follows intermittent, project-based patterns rather than daily use, suggesting AI tools should prioritize task-specific interfaces over generic chat functionality.