Intelligence should survive a difficult question.
A persuasive answer is easy to admire. A medicines team needs something harder: an answer whose source can be checked, whose uncertainty changes the next action and whose limitations remain visible when the work becomes busy. PHARMORIS development explores bounded assistance with medicines information and operational review. The commercial question is whether that assistance improves useful work after the cost of checking it has been counted.
PHARMORIS remains in controlled development. The approach below describes proposed evaluation questions for model-assisted capabilities; it does not claim independently verified AI performance, autonomous medicines decisions or a production clinical service.
Choose a task a responsible person can judge
Start with an observable task: identify missing fields in a service review, structure a permitted source for assessment or distinguish a stated assumption from an observed result. Specify what a correct output contains and which mistakes would matter. An open instruction to optimise medicines is too broad to establish a useful comparison. The intended user should be able to explain why the assistance is needed and what they would do if it were unavailable.
Make the checking work part of the result
A draft created quickly can still be expensive if every sentence needs reconstruction. A proposed evaluation would measure the full task: preparation, output review, correction and final acceptance. Reviewers would compare it with a competent existing approach using equivalent material. A useful result could be reduced rework or clearer identification of a missing source. It could equally be that assistance adds no net value for this task. Either finding is more informative than a demonstration timed only to the first response.
Show where the answer came from
The proposed reading experience should distinguish source statements, calculations and interpretations. A reviewer should be able to return to the relevant permitted material and understand its date, coverage and limitations. A source that changed after the review should not silently rewrite the original decision basis. The public material describes these expectations at the level a buyer can examine; it does not disclose confidential implementation details or imply every intended control has been verified.
Give uncertainty an operational consequence
Missing information, conflicting sources and a request outside the intended task are different problems. We would propose testing whether the workflow directs each to an appropriate response: request material, record disagreement, narrow the question or stop. The responsible reviewer needs time, information and authority to reject a result. A mandatory approval click alone would tell us little about whether meaningful oversight took place. Review effort and reasons for disagreement would therefore belong in the evaluation record.
Examine the examples that could break the case
The proposed artefact is a task evaluation card. It would identify the intended user, permitted inputs, comparator, error categories, checking burden and acceptance decision for one use case. Test material should include incomplete records, plausible but unsupported assertions and cases that should be declined. Different settings may expose different difficulties; a result from a central medicines team would not automatically establish suitability for a small pharmacy working with fewer review resources.
Treat a changed capability as a new question
A change in the task, model, source or user population may alter the earlier finding. A proposed assessment would record what was examined and identify which changes require further review. Where personal data processing is contemplated, the relevant data-protection assessment must be determined for the actual use. Intended medical functions also require an appropriate regulatory assessment. Calling a product an assistant does not settle those questions, and an evaluation agreement cannot remove responsibilities that apply to the intended use.
Define the evidence that would persuade your team
We welcome medicines leaders, evaluators and governance colleagues who can nominate one recurring task and a credible existing comparison. A useful next discussion would identify the errors you cannot accept, the evidence your reviewers need and the threshold at which further investigation would be worth funding. We would agree any commissioned evaluation separately, including permissions, responsibilities and the decision it should inform.
External guidance and the scope of this approach
The ICO's AI and data-protection resources inform questions about personal-data use. MHRA's software and AI guidance is relevant when assessing potential medical-device functions. These sources provide context; they do not approve PHARMORIS or establish its regulatory classification. The task card and proposed comparisons above are NEUVIOR evaluation proposals.
