A plausible medicine recommendation does not establish a safe prescribing workflow. The service must know who the patient is, obtain sufficient relevant information, make an appropriate decision, issue it under the correct authority and arrange necessary communication and follow-up. Each stage can fail even when the selected medicine is a reasonable option in another context.
This article considers UK professional practice and proposed workflow safeguards as reviewed on 10 October 2026. It is not a dosing guide or a statement that an overseas pilot grants prescribing authority in the UK.
Define what the software actually does
A prescribing-related product might retrieve medicines information, suggest an option, prepare a draft, support an authorised prescriber or execute a narrowly permitted transaction. Those functions should not be collapsed into a single claim that "the AI prescribes".
The distinction determines what evidence and controls are needed. A drafting assistant must preserve the clinician's instruction accurately. A system recommending treatment must be assessed for the quality and applicability of that recommendation. Software capable of issuing an authorised order also needs controls around identity, permissions, destination and confirmation of the action.
A buyer should ask to see the exact transition between recommendation and commitment. Which person or process authorises it? What information is displayed? What can still be changed? An attractive demonstration of a completed screen does not answer these questions.
A blank answer is not a negative history
Consider a fictional patient seeking a first treatment for a skin problem through an online service. The assessment contains useful information, but a question relevant to treatment suitability has been left unanswered. A hypothetical system summarises the record as having "no relevant contraindications".
The immediate concern is the unsupported transformation of missing information into reassurance. The problem would remain even if the same treatment were frequently suitable for patients with that condition. Eligibility belongs to the individual assessment, not simply to the diagnosis label.
The GMC's guidance on deciding whether it is safe to prescribe, reviewed on 10 October 2026, requires sufficient, reliable knowledge of the patient's health and consideration of the consultation mode, dialogue and relevant information sharing. It specifically recognises that missing information can make prescribing unsafe.
The scenario does not allege that a particular service omits those checks. It illustrates a test that should be included before an automated workflow is trusted with consequential decisions.
Initial treatment and renewal require different questions
Starting treatment involves establishing the problem and selecting a suitable approach. Renewing an existing prescription involves checking whether continuation remains appropriate. Neither can be reduced to locating the medicine name in a record.
The GMC's repeat-prescribing guidance, reviewed on 10 October 2026, addresses ongoing need, appropriate monitoring and arrangements for review. An existing prescription is relevant evidence, not a permanent authorisation to continue unchanged.
A proposed renewal test could introduce a medicine listed as current in one source but discontinued in a later specialist letter. Another could leave required monitoring unavailable. The expected response is not a predetermined refusal in every case; it is recognition of the discrepancy and an appropriate route to resolution.
Evaluation should report the actual task. Evidence from a restricted renewal workflow cannot automatically validate new treatment initiation, and performance on a first-treatment pathway cannot establish safety across complex long-term prescribing.
Authorisation should bind the decision to the action
A clinician might approve an appropriate plan, but the system still has to carry that plan into the correct patient record and prescribing process. The review should make clear the patient, medicine, formulation, route, relevant instructions and intended destination without requiring the clinician to infer them from a general summary.
This article provides no clinical doses. The point is that language generation is not a substitute for verified inputs, appropriate numerical tools and the checks required by the prescribing system. A coherent explanation cannot validate a calculation merely by repeating its result.
The authorisation should apply to a defined action. If the patient context or proposed order changes afterwards, the workflow needs a rule for whether renewed review is required. A broad permission to "manage prescriptions" should not silently become permission to change treatment beyond the agreed scope.
These are proposed governance and engineering controls. Their implementation should be demonstrated in the actual environment rather than assumed from a company's general safety statement.
Dispensing and starting treatment are separate milestones
An issued prescription may not have reached the intended dispensing service. A medicine may be unavailable, a clarification may be required or the patient may not obtain it. Even successful dispensing does not prove that the patient understood the plan or started treatment.
A complete pathway should therefore distinguish preparation, authorisation, issue, receipt and any necessary follow-up. It should make uncertainty visible when a transaction times out rather than report success because the model intended to complete it.
The system also needs a safe response to retries. Repeating an uncertain action without checking its status could create a duplicate; abandoning it could leave the patient without the intended treatment. The appropriate response depends on the prescribing infrastructure and must not be improvised by conversational confidence alone.
Pharmacists and other professionals should have a clear route to raise discrepancies. A workflow that speeds the prescriber's task while making clarification harder for the dispensing team may transfer rather than reduce the safety burden.
Follow-up is part of prescribing, not an optional extra
The GMC's guidance on reviewing medicines, reviewed on 10 October 2026, addresses monitoring, follow-up and review even when a prescription is issued on a single occasion. An automated service should make those arrangements visible where they are clinically required.
A proposed assessment should ask who responds to adverse effects, treatment failure or new information. What happens when the patient does not answer a follow-up message? Can they reach an appropriate professional? Is a handover to another service agreed, or has responsibility merely been implied in a generated instruction?
The endpoint should reflect the care being offered. "Prescription sent" may be an honest transaction measure, but it is not a patient-outcome measure. Report appropriate treatment, necessary escalation, follow-up completion and unresolved problems separately.
Check medicines information without confusing reference with authority
As described by iatroX in October 2026, its clinical-reference offering includes linked guidance and SmPC information from emc. A clinician can use iatroX's medicines destination to explore a general medicines question and inspect the underlying source.
That is a different function from prescribing, authorising an order or monitoring an individual patient's treatment. This article is published by iatroX, and its answers require the same applicability checks as other reference tools. Source-grounded design is not proof that every output is correct.
For an information need, a reference workflow may be sufficient. For treatment initiation or continuation, the service must establish the patient-specific facts and professional authority. For automation beyond that decision, the evaluation must follow the action through dispensing, communication and any required review.
Frequently asked questions
Is choosing an appropriate medicine enough to validate an AI prescribing service?
No. Identity, patient information, authorisation, accurate execution, communication and necessary follow-up are separate parts of the safety assessment.
Is renewing a prescription just an administrative task?
Not always. Continued treatment may require review of ongoing need, monitoring, tolerance and changed circumstances, even when the medicine has been prescribed before.
Does a clinical-reference answer authorise a prescription?
No. A reference tool can support information checking, but prescribing authority and the patient-specific decision remain separate from receiving an answer.