Like It or Not, It’s Here to Stay: AI Applications in HEOR

“Whether you like it or not, this technology is here to stay” was the general feeling emanating from ISPOR 2025 on the topic of AI. The sense is, it’s not if or when, but how do we adapt to a new way of working and implement the ever-advancing AI technologies into our workflows in the health economic and outcomes research (HEOR) community, to maximise efficiency and improve results, while maintaining trust and validity in research outputs?

To answer this question, the 2025 International conference highlighted a shift toward more autonomous, agentic systems that could fundamentally transform evidence generation and decision-making processes. In the literature reviews field, presentations focused on the use of AI in data extraction and search strategy development, as well as current published guidelines on these uses (covered in AI in Literature Reviews, unpacking the concept of this being the ‘low-hanging fruit’ in HEOR). A further key theme was ensuring we maintain trust and validity through the implementation of checklists and development of benchmarking appropriate to the tool as these evolve. Finally, it was interesting to see discussions around how organisations should develop a roadmap towards AI adoption in HEOR and gain leadership buy-in.

From Supporting Tools to Autonomous Agents

A key discussion topic was the move toward AI systems capable of functioning as independent, multi-agent workflows. An issue panel on AI agents and guardrails in HEOR showcased a move towards this agentic approach, utilising AI systems that continuously update evidence, incorporating real-time data with possible future applications in generating dynamic, living health technology assessments (HTAs).¹ Such agentic approaches, where multiple AI agents collaboratively review literature, extract evidence, and synthesise findings, aim to dramatically accelerate evidence refresh rates, reduce manual workloads and improve robustness through multi-review mechanisms.

Customised AI tools built on HEOR-specific large language models (LLMs) were also discussed across the conference, with a feeling that these are going to produce more accurate and trustworthy outputs than generic LLMs such as ChatGPT. As an example, one session compared the use of ChatGPT 4.o Turbo with a custom HEOR AI tool called Alde for developing a disease burden report.² It was found that ChatGPT, as the generalist, needed more specific prompting than Alde, and referenced only journal articles, many of which were irrelevant. The custom tool was able to produce results with less information specified in the prompts, was able to identify more relevant references than ChatGPT, and look at other sources such as web articles and press releases.

Other use cases for AI in HEOR at the conference included testing an LLM to reproduce a published economic model, first from the manuscript and then from a technical report (with improved outputs achieved in the latter approach), and the development of AI chatbots, which can access existing internal materials on a product to answer questions on topics such as clinical trial design and economic models.³

In real-world evidence (RWE) discussions, a key AI-based use case included extracting information from unstructured clinical notes to identify the presence or absence of key symptoms, observations or treatments. By mapping these to relevant structured fields or codes at extremely large scales, data quality can be improved and the volume of evidence enriched. Other sessions highlighted research findings that machine learning outperformed traditional modelling approaches used to predict health outcomes, and case studies which employed AI tools to conduct analyses and generate synthesised results.

Risks and Challenges

However, with these exciting advances come risks and caveats: appropriate boundaries must be set for AI autonomy, and insightful reporting traces developed for human review. With non-agentic approaches, where humans develop the prompts for AI models to act upon, these prompts can be reported on and evaluated externally, but this is not possible with the agentic approach where the LLM is deciding which external sources to consult and which actions to execute.¹ Therefore, the HEOR community must work to develop benchmarks for evaluating these agents in a continuous way.

Even with appropriate boundaries set and human in the loop review, there may be some more intrinsic risks to using AI in HEOR, which we must be aware of and guard against:

• Will increased automation lead to larger and more sophisticated HTA submissions? Will this leave human-focused teams overwhelmed with reviewing huge HTA submissions looking at every approach, which may not be better, just more complex to review?¹
• AI agents will raise the bar in terms of consistently strong outputs, but will this also result in a cookie cutter approach that limits submissions and their evaluation by artificially reducing options? Is there a benefit in having heterogeneity of human expertise rather than using LLMs which have all been trained on the same materials?¹

Acceptance and Transparency

In an attempt to consider mitigations to these risks and challenges, the importance of transparency and validation was still top of mind at ISPOR 2025. The sessions elaborated on the necessity of ongoing benchmarking, lifecycle evaluation, and reporting standards. Participants discussed the importance of transparent documentation and reporting when AI contributes significantly to evidence and submissions. Recommendations included clearly stating the use of AI tools in dossiers, supported by frameworks like the ELEVATE checklist, which offers a structured approach to reporting AI’s role, especially when its outputs fluctuate or evolve over time.² These measures are viewed as essential for regulatory acceptance and for maintaining the credibility of AI-generated evidence in HTA.

Organisational Strategies and Practical Implementation

So, accepting the benefits, risks and mitigation strategies for using AI in HEOR, how do we adopt AI at scale into many existing and complex workflows? One session outlined a roadmap to AI adoption in HEOR, including selected targeted use cases; building agile, cross-functional teams including HEOR experts, legal, and technical specialists; and creating sandbox environments for testing.² To ensure buy-in at all levels, it will be important to measure the impacts of AI incorporation, such as time savings and improved report accuracy, and tie AI projects to measurable business goals.

Looking Ahead: The Future of AI in HEOR

The consensus is that within the next year we will be living in a very different world. AI’s role will continue expanding, moving toward continuous evidence updates, real-time evidence synthesis, and scalable workflows. The potential for “living” evidence and AI-driven automation to reshape pharmaceutical research and development and healthcare decision-making was a prominent theme. However, responsible implementation, through validation, transparency, and stakeholder engagement, remains paramount.

It is hard to predict how AI will shape the pharmaceutical industry, but it feels inevitable now that it will be disruptive, and therefore those working in it must (cautiously?) embrace AI or risk being left behind.