Matching the AI to the Evidence: Three Approaches Coming to ISPOR Europe 2026

Whether to use AI in a literature review is still an important question. Increasingly, another question matters just as much: which type of AI is appropriate for which type of review?

At ISPOR Europe 2026 in Vienna, our team is presenting three posters that approach the question from different angles. Together, they argue that AI tools are not one-size-fits-all and their responsible use should complement, not replace, human experts. For higher-stakes reviews, such as those informing health technology assessment (HTA) or publication, the role of AI requires particularly careful consideration as expectations around AI use continue to evolve. For reviews where AI does have a role, both the type of AI and how it is used matter.

Start with the purpose of the evidence

Literature reviews serve different purposes, and this influences how much AI-related risk is acceptable. A scoping review or evidence map designed to shape internal strategy or generate hypotheses has a different purpose from a systematic literature review (SLR) that will inform a publication or HTA: they should not carry the same AI-associated risks.

One of our three posters proposes a risk-based workflow for responsible AI use in SLRs with different tiers based on how the evidence will be used. Each tier sets a different bar detailing how AI is implemented, validated, overseen, and reported, including how AI performance is measured against human reviewers, a benchmark that is itself imperfect, and whether the work can be reproduced. The workflow draws on converging guidance from ISPOR, Cochrane, NICE, and others, and helps teams decide what role, if any, each type of AI should play.

When you need scale: Generative AI for landscape and gap mapping

Not every literature review sets out to answer a narrow clinical question. Some aim to show where research is concentrated and where to identify the gaps. In this setting, generative AI may offer a way to perform work that would be impractical to do by hand.

Our second poster presents a workflow in which a large language model (LLM), a type of generative AI, assigns publications from a fast-growing evidence base to topics defined in advance by experts, synthesizes what each topic shows, and flags evidence gaps for expert review. Our poster is explicit about where human expertise is still needed.

When you need volume: Repeatable screening with a biomedical model that runs in-house

Generative AI is not always the right choice. It can yield different answers to the same question, be costly at scale, and be unsuitable when data must remain on your own computers or servers. Our third poster explores an alternative: a specialist biomedical model that analyses text rather than generating it. It learns from abstracts that experts have labelled as relevant or not, then ranks the rest from most to least likely to be relevant, so reviewers can start screening the most promising abstracts at the top of the list. It runs on standard computers and has been tested across multiple research questions. It trades some set-up effort, in the form of expert labelling, for repeatable results and transparency. The poster looks at how well that trade works in practice and what makes the difference.

Join us in Vienna

Taken together, the three posters point to the same conclusion: innovation in evidence review isn’t about using AI everywhere. It’s about matching the right tool to the research question, the purpose of the review, and the evidence available. We would welcome the chance to talk it through with you at ISPOR Europe 2026 (8–11 November, Vienna).

If you are weighing up which approach suits your next review, come and find us at the posters or at booth #1008.

Oxford PharmaGenesis, part of the Klick Health family