NICE draws a hard line around AI use in health economic evaluation
NICE has set out an interim position on AI methods in health economic evaluation, a part of health technology assessment that can absorb substantial staff time when analysts assemble evidence, build models, and test assumptions about value for money. The framework accepts that generative AI could make parts of this work faster, especially where teams are drafting code, estimating starting parameters, or replicating existing models. It does not treat that speed as a reason to hand over judgment to software.
The guidance places a consistent condition on every use case: AI may support the process, but it should not replace the human analyst. In practice, that means AI-generated code still needs verification before it runs, parameter suggestions still need expert scrutiny, and model outputs still need challenge against known clinical and economic expectations. The emphasis is not on automation for its own sake; it is on reducing repetitive work without weakening the evidentiary chain that underpins a reimbursement decision.
A model that runs quickly but cannot be trusted is useless in HTA.
Where generative AI can fit into economic modelling
The framework identifies several areas in which AI could be applied. In model conceptualisation, it may help explore the structure of a decision model and the assumptions that need to sit beneath it, alongside a human expert who determines whether the structure makes clinical sense. In parameter estimation, generative AI may help generate initial inputs for a de novo model, a task that can be laborious when data are sparse or scattered across sources.
The same logic extends to model construction, where AI could write code in languages such as R or Python for a human to review. That kind of support could shorten the time needed to move from concept to executable model, particularly when teams are translating a conceptual diagram into technical scripts. In model validation, AI could be used to probe whether outputs behave as expected when inputs change, a useful guard against hidden programming errors and brittle assumptions.
NICE also points to model adaptation, where existing country-specific models are adjusted rather than rebuilt from scratch, and model replication, where a published model is reconstructed for validation, reuse, or updating. Both are plausible targets for generative systems because they involve structured, repetitive work. Even there, the framework draws back from full delegation. AI can assist in identifying or updating key parameters, but it should not be allowed to substitute its own outputs automatically into a model.
Why the human analyst still sits at the center
The caution in the guidance reflects a basic feature of health economic modelling: a model is only as reliable as the assumptions encoded inside it. Errors can arise from mis-specified structure, unrealistic parameter values, or code that executes correctly but answers the wrong question. In economic evaluation, those failures are not cosmetic. They can shift cost-effectiveness estimates, alter incremental cost-effectiveness ratios, and change the apparent value of a technology to a health system.
Generative AI introduces a second layer of risk because it can create plausible-looking text or code without proving that the content is valid. A script may appear internally consistent while still embedding a flawed assumption, and a parameter estimate may look reasonable while being detached from the relevant evidence base. NICE therefore keeps the human in-the-loop not as a formality but as the primary quality control step. Analysts must decide whether the tool’s output is clinically and methodologically defensible before it influences a model used in decision making.
For health technology assessment teams, that position has immediate operational consequences. AI can reduce drudgery, but it cannot absorb responsibility. The framework treats support-only use as the acceptable boundary, and it expects reporting to make that boundary visible.
Reporting expectations and the need for traceability
The framework is explicit that reporting should show how AI was used for support only. That requirement matters because readers of an HTA submission need to understand what was generated by software, what was checked by a human, and what was actually adopted in the final model. Without that trail, it becomes difficult to judge reliability, reproducibility, or the extent to which AI influenced the result.
NICE’s interim position also draws on early best-practice work from its HTA Lab, which has explored how generative systems might be used across economic evaluation workflows. The guidance accepts that the evidence base is still immature and that methodological risk remains. In a field where uncertainty is already built into the analysis, another layer of opacity would only compound the problem.
One sentence in the guidance captures the operational standard clearly: AI should be used as a support tool only, with a human closely in the loop, not as a route to semi-automation or full automation. That leaves a narrow but practical role for the technology. It can help draft, test, and adapt, but the final model must still reflect expert review, transparent reporting, and a defensible line from evidence to estimate.
What this means for HTA practice now
The framework does not close the door on generative AI in health economic evaluation. It opens it cautiously, with conditions attached. Teams that want to use AI will need to show that the technology stays within the support function, that outputs are checked before use, and that model decisions remain under human control. In a discipline built on transparency and reproducibility, those safeguards are not optional extras; they are part of the method itself.
For now, the most important question is not whether AI can write faster code or generate a plausible parameter set. It is whether a health system can trust the result once that code or parameter has been examined, verified, and documented. NICE’s answer is clear enough: trust begins with human oversight, and any AI use that cannot be traced back to that principle falls outside acceptable practice.