Clinical AI systems can be built around two genuinely different starting points: beginning with the published literature and synthesising it, or beginning with authoritative clinical guidelines and using research to clarify gaps or uncertainty within them. Both are defensible design choices, and each carries risks the other structurally avoids.
Defining paper-first clinical AI
A paper-first system searches individual studies and reviews directly, synthesises what it retrieves, and, increasingly, grades the resulting evidence for a clinician to interpret.
Defining guideline-first clinical AI
A guideline-first system begins instead with authoritative clinical recommendations, identifies the currently recommended pathway for a given question, and turns to primary or secondary research specifically to clarify areas of genuine uncertainty or cover newer developments the guideline has not yet incorporated.
The genuine strengths of paper-first retrieval
New evidence becomes visible before any guideline committee has had the chance to review and incorporate it, which matters for genuinely fast-moving areas of medicine. Coverage of rare or emerging questions, ones no guideline has yet been written to address at all, is inherently better under this approach. And scientific uncertainty itself, disagreement or genuine gaps in the literature, tends to be more visible, since a paper-first system is not filtered through a single, already-synthesised recommendation.
The genuine strengths of guideline-first retrieval
Recommendations produced through a guideline process have already integrated evidence quality with safety, feasibility, cost and implementation considerations, work a paper-first system would otherwise have to attempt fresh for every single question. This gives guideline-first answers greater relevance to the specific local healthcare system a clinician is working within, and makes them considerably more directly actionable for the large majority of routine point-of-care questions, which are not at the cutting edge of active research uncertainty.
The risks each approach carries
A paper-first system risks producing an answer that is academically accurate but operationally inappropriate, technically correct about what a trial showed while missing what a clinician working within a specific healthcare system can actually do with that information. A guideline-first system risks lagging behind genuinely important new evidence that has emerged since the guideline's last formal review, presenting a recommendation as current when newer data has already begun to shift expert opinion.
The iatroX hybrid approach
iatroX's working structure prioritises UK guidance as the starting point, favours systematic reviews and meta-analyses as the next layer of evidence, considers high-quality newer randomised trials specifically where guidance may not have caught up, uses appropriate observational evidence where the clinical question calls for it, and, importantly, states explicitly where current guidance and newer evidence appear to diverge, rather than silently picking one over the other.
Which questions suit each approach best
Genuinely novel, fast-moving, or rare clinical questions, where no guideline yet exists or the evidence base is actively shifting, are better served by paper-first retrieval's greater visibility of uncertainty and emerging evidence. Routine, well-established point-of-care questions, the overwhelming majority of everyday clinical practice, are better served by guideline-first retrieval's greater direct actionability and system relevance.
iatroX as a bridge between synthesis and implementation
The genuinely useful positioning for a UK-native platform is as a bridge between evidence synthesis and clinical implementation, using the strengths of each approach where they are actually strongest, rather than committing exclusively to either paper-first or guideline-first logic across every kind of clinical question.