ISPOR 2026 in Philadelphia made one thing clear: the debate over whether AI belongs in HEOR is over. The harder questions — how we govern, validate, and hold ourselves accountable for AI-assisted evidence — are just getting started.
Key Takeaways
- AI is now the #1 HEOR trend for 2026–2027 — what matters isn’t the ranking, it’s that LLMs, causal ML, and AI-assisted workflows are maturing simultaneously for the first time
- The three hardest questions in AI-enabled HEOR — how to validate outputs, who carries accountability, and whether the workforce is ready — still have no consensus answers
- Human-in-the-loop is necessary but not sufficient — what HEOR actually requires is expert-in-the-loop design, where disciplinary judgment is built into each stage of the workflow, not bolted on at the end
What’s Actually Different This Year
AI earned its own dedicated track at ISPOR 2026 and topped the society’s 2026–2027 HEOR Trends report for the first time, up from third. But the ranking isn’t the story. What makes this moment different is that large language models, causal machine learning, and AI-assisted RWE workflows are all maturing simultaneously, and that convergence is new.
Previous cycles saw one method emerge at a time. Teams could adapt incrementally. What the sessions at ISPOR 2026 made clear is that the field is now facing adoption across multiple methods at once, each reinforcing demand for the others. The most-attended sessions weren’t introductory overviews, they were applied methodological debates: how to validate LLM-assisted literature syntheses, translate causal ML findings into submission-ready evidence packages, and build cost-effectiveness models with generative AI while satisfying HTA rigor. The field is no longer asking whether to adopt AI. It’s contending with how to govern it across several methods simultaneously, with standards that are still being written.
The Three Questions the Field Still Can’t Answer
This was, for me, the most honest thread running through ISPOR 2026, often just beneath the surface of the methodological debates, but unmistakable.
1. How do we validate AI-generated HEOR outputs?
There is no consensus. Whether the output is an LLM-assisted literature review or a set of transition probability matrices generated for a cost-effectiveness model, teams are building their own validation heuristics. Some are rigorous. Most aren’t. And none of it is portable, meaning the same question gets answered differently across organizations, sponsors, and reviewers. That won’t scale, and it won’t hold up under regulatory scrutiny. The ELEVATE-GenAI reporting guidelines are a meaningful start; operationalizing them into actual review infrastructure is the unfinished work.
2. Who is accountable when AI-assisted evidence informs a coverage decision?
The traditional accountability chain, principal investigator, sponsor, independent reviewer, was designed for a world where humans (experts) drove every step of the evidence. As AI embeds deeper into literature synthesis, model structure, and data analysis, that chain blurs. The PI who signs off on an AI-assisted model may not have seen the intermediate outputs the LLM generated. The reviewer may not know which transitions were human-specified and which were model-generated. This isn’t a hypothetical future problem. It’s arriving with current workflows. The field needs to address accountability directly, not assume legacy frameworks will hold under conditions they were never designed for.
3. Is the workforce keeping pace?
A real division is emerging between teams that have invested meaningfully in AI literacy and those that haven’t, and the gap is growing faster than most expected. The issue isn’t just knowing that a tool exists. It’s understanding what it can and cannot do, where its failure modes are, and how to interrogate its outputs rather than simply accept them. That kind of fluency takes time and deliberate investment. Knowing what’s possible and deploying it responsibly remain two very different things, and the distance between them is where errors will be made.
From Human-in-the-Loop to Expert-in-the-Loop
The session that most shaped my thinking at ISPOR 2026 wasn’t a keynote, it was a poster. Researchers had built a proof-of-concept framework using an LLM to automatically generate and execute cost-effectiveness models across multiple structures. The results were imperfect, but close enough to shift the question from if generative AI belongs in health economic modeling to how the field governs it.
The phrase “human-in-the-loop” is everywhere in HEOR right now. It’s doing a lot of work — and not always the right work. In a field built on multidisciplinary expertise, the more precise framing is expert-in-the-loop: not just a person present at review, but a trained analyst, economist, epidemiologist, clinician, statistician, exercising domain judgment at defined points in the workflow.
Bolting oversight on at the end is not expert-in-the-loop design. It’s expert-at-the-end-of-the-loop design — and those are fundamentally different things.
What the poster made visible is that when an LLM generates a model structure, the expert’s role isn’t to approve or reject the output. It’s to interrogate the assumptions embedded in it: which transition states were constructed, which comparators were selected, which data sources were drawn on, and where the model’s choices diverge from clinical reality. That kind of engagement can’t be retrofitted. It has to be built into each stage of the workflow as a defined point of analytical ownership, not a checkpoint, but an active intervention.
The expert’s role doesn’t shrink under that model, it sharpens. Less time constructing scaffolding, more time supplying the judgment that no model can generate.
What clinical medicine learned through the adoption of AI in radiology and pathology — and what took longer than it should have to internalize — is that expert oversight isn’t a transitional phase until the algorithms improve. It’s the condition under which high-stakes AI becomes trustworthy enough to use. HEOR is arriving at the same conclusion. The field that designs expert judgment from the start, rather than managing it retroactively, will be the one whose evidence holds.
What I’m Taking Forward
ISPOR 2026 was the society’s 30th annual meeting, back in the city where it began. The field has navigated every major methodological shift, and it will navigate this one. But only if standards development keeps pace with adoption. Based on what I saw in Philadelphia, I’m cautiously optimistic. That optimism has to be earned, guideline by guideline, accountability structure by accountability structure.
The question worth asking now isn’t whether your team is using AI. It’s whether you’ve designed the expert back in — deliberately, structurally, at every step that matters.
About the Author
Ashis Das, MD, is Senior Director of Evidence Intelligence at Komodo Health, where he works at the intersection of clinical medicine, AI, and real-world evidence, including Komodo’s Marmot platform. A physician and thought leader, he focuses on how the healthcare industry generates and applies evidence to drive better decisions across biopharma and payer stakeholders. Connect with Ashis.
