The medicines business | A question to investigate
A convincing summary is not enough
An evidence summary can read beautifully and still omit the study that changes the conclusion. That is the central challenge when investigating AI-assisted review in medicines: fluency must not be mistaken for completeness or accuracy. NEUVIOR is interested in a narrowly defined evaluation of evidence-preparation work, where a qualified reviewer can inspect the sources, challenge the output and decide whether any observed benefit survives the checking it requires.
In the conversation.
NEUVIOR and our founder’s perspectives have appeared across science, business and technology publications.
Reporting, commentary, letters and podcast appearances | Read each feature for its format and source.
Define the task before judging the tool
Literature retrieval, extraction of study characteristics, comparison of findings and drafting a narrative are different tasks. A system can perform well at one and poorly at another. A proposed evaluation would select a bounded question and specify exactly what the output must contain. It would also identify what the system must leave unresolved. This avoids a familiar failure: judging a polished answer against an expectation that was never written down.
The reference context matters
NIST’s Generative AI Profile addresses risks particular to generative systems (NIST, 2024). FDA’s January 2025 guidance on AI supporting drug regulatory decisions remains draft and describes context-specific credibility assessment (FDA, 2025). Neither establishes that a proposed evidence-review use is acceptable. The first study would remain a controlled preparation exercise, with any later regulatory use requiring its own assessment and appropriate expert engagement. Draft guidance must be presented as draft wherever it informs the discussion.
Build a comparison that can find omissions
A reference set should include relevant positive, negative and inconclusive findings, along with studies whose relevance requires judgement. Independent reviewers should agree the assessment criteria before seeing experimental outputs. They should distinguish an absent source from an inaccurate extraction and an unsupported inference from a disagreement about interpretation. A single overall score can conceal an error that matters. The evaluation needs enough detail to show where assistance helps and where it creates work or risk.
An error profile the team can use
The proposed deliverable would distinguish omissions, inaccurate extraction, unsupported interpretation and reviewer disagreement. It would show whether a narrow task performs differently from the wider narrative. That gives the team a concrete development choice: retain a useful component, change the workflow, repeat a focused test or reject the approach on the available evidence.
A public decision brief | An illustrative example
Open the question.
Inspect the reasoning.
Choose a business question, then look beneath its claim. See what a public source supports, what depends on the local setting and what a useful next step could be.
A medicines or implementation team considering an AI-supported evidence review.
Would a quicker review make the decision better?
The interesting question is the complete piece of work: finding relevant material, checking it, correcting the draft and reaching an accountable decision.
A proposition to investigate
Less preparation. The same ability to challenge.
An evidence-review approach could be useful if it reduces avoidable preparation while preserving source access, relevant qualifications and professional checking. That is the proposed value to test. No time reduction, accuracy improvement or customer result is asserted in this example.
The proposition is constructed for this public brief. It is not a published finding about PHARMORIS.
What the next brief would contain
A decision someone
can commission.
- The decision
- Whether one defined review task merits an agreed evaluation.
- The comparator
- The team's existing complete review process, including checking and corrections.
- The evidence
- Representative cases, source records, reviewer decisions, complete-task effort and exceptions.
- Stop or redesign
- Stop or redesign if source checking weakens, errors are unacceptable or the complete task offers insufficient value.
Constructed public example | Not PHARMORIS-generated software output or a measured customer result.
Keep this example briefA proposed PHARMORIS study
Inputs could include openly available publications, an agreed search boundary and a reviewer-authored reference brief. The comparator would be the current preparation workflow applied to the same question. An experimental evidence-linked brief would be assessed without influencing treatment, product claims or regulatory submissions. Reviewers would record the sources they checked and corrections they made. The purpose would be to investigate one workflow, not demonstrate clinical validation or broad autonomous research capability.
Count the cost of checking
Candidate measures include source fidelity, material omissions, incorrect numerical extraction, unsupported causal claims and total reviewer time. Time should include verification and correction, not just generation. The study should record whether users become more confident than the output warrants. A useful result might support a narrow extraction task while rejecting narrative synthesis. It might also show that a simpler non-AI organisation of sources delivers the same benefit with less uncertainty.
When the result is not good enough
The study should stop if fabricated references recur, if reviewers cannot trace material claims or if the task boundary expands beyond what was tested. It should pause after a substantive model or retrieval change until the comparison is repeated appropriately. No small evaluation should be marketed as proof that AI can replace a systematic review. A credible conclusion states the exact task, material tested, error pattern and unresolved limitations, including a negative result where appropriate.
Bring one review that consumes attention
A medical, scientific or regulatory team can bring a recurring preparation task and describe how its output is checked today. We can discuss whether a controlled comparison is feasible and what performance would justify further work. The potential commercial value lies in a reliable, inspectable contribution to a human review process, with evidence that the assistance earns its place rather than merely producing more persuasive text.
Sources and scope
NIST | Generative AI Profile | FDA | AI regulatory decision-making draft guidance
NIST (2024) Generative AI Profile. Available at: https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf (Accessed: 14 October 2026).
FDA (2025) AI regulatory decision-making draft guidance. Available at: https://www.fda.gov/regulatory-information/search-fda-guidance-documents/considerations-use-artificial-intelligence-support-regulatory-decision-making-drug-and-biological (Accessed: 14 October 2026).
Another useful angle.
- Which supply exception deserves attention first?
Explore a proposed wholesale exception-review evaluation that compares effort, escalation and operational consequences.
- Traceability has to work when something does not match
A proposed US medicines traceability exception-review study for manufacturers, distributors and dispensers.
- When quality becomes a capital decision
A proposed manufacturing-quality evaluation linking operational uncertainty, investment options and measurable decision evidence.