Faster but Not Truer, Part 2: The Regulatory Record, the Economics, and the TRACE-HEOR Reporting Standard

Background Part 1 of this series found that measured gains from artificial intelligence (AI) in health economics and outcomes research (HEOR) concentrate in tasks with observable ground truth and disappear where the quantity of interest is a counterfactual. What regulators, health technology assessment (HTA) bodies, and sponsors have said about that division has not been assembled in one place. Methods We assembled the primary regulatory, HTA, industry, and journal policy record through September 2026, requiring for each claim a named primary source and, where a quantity was reported, its provenance and whether the reporting party had a commercial interest. We downloaded the Food and Drug Administration (FDA) AI-enabled medical device list in full, tabulated published economic evaluations of AI, and set reported industry efficiency gains against disclosed use in submissions. We then evaluated the field against the task-based productivity framework of Nobel laureate Daron Acemoglu and derived a reporting standard from the documented failure modes. Results The FDA has authorized 1614 AI-enabled devices, 1230 (76.2%) in radiology, rising from 18 in 2016 to 335 in 2025, and has received more than 500 drug and biologic submissions containing AI components since 2016; its January 2025 credibility assessment guidance remains a draft and excludes discovery and operational uses. Industry sources report time savings of 40% to 97% in regulatory writing and approximately 80% in payer dossier drafting, but every one of these figures comes from a single-arm comparison without a concurrent control and was reported by the company that deployed or sold the tool. AI use was disclosed in 3 of 1391 HTA documents (0.2%) and 0.6% of one journal’s submissions, against surveys reporting use between 25% and 81%. Nineteen economic evaluations of AI have been published against 1614 authorized devices. No regulator or HTA body has been documented accepting or rejecting an AI-derived endpoint as such. Conclusions Current disclosure instruments ask whether AI wrote the text. Reviewers need to know whether AI selected the evidence, defined the cohort, derived the endpoint, or generated the code, and whether the recommendation would change without it.