An AI answer can be clinically flawless and professionally unusable, because clinical plausibility and authority to act are separate questions and only one of them is in the model's gift. Whether an intervention is reasonable is a medicine question; whether this prescriber may initiate it is a function of registration, annotation, personal competence, employer authorisation and, for some, a clinical management plan, and no general system can infer that stack from a job title. The practical consequence is a rule this article builds toward: prescribing-support prompts should state profession, role, setting, scope and jurisdiction, and answers should be read as scoped to what was stated, never to what was hoped.
The authority stack, layer by layer
Four layers, each independently capable of stopping an otherwise sensible plan. Registration and annotation: independent prescribing is, in the HCPC's framing, prescribing by an accountable practitioner responsible for the assessment and clinical-management decisions, and the entitlement is specifically annotated on the register for eligible registrants; no annotation, no prescribing, whatever the answer says. Legal profession-specific limits: entitlements differ across professions in defined ways, controlled-drug provisions being the visible example, so identical clinical situations carry different lawful options for a nurse, pharmacist, paramedic, physiotherapist or podiatrist prescriber. Personal competence: the framework's requirement that prescribers work within their own demonstrated competence, which no register can annotate and no answer can assess. And organisational authorisation: employer scope agreements, local formularies and governance, the layer where "may in law" meets "may in this job". Supplementary prescribers add a fifth: the clinical management plan, inside which every decision must fit regardless of the wider evidence. The same case, an intermediate-complexity presentation with a clear guideline pathway, can therefore be: routine for the ANP whose scope covers it; appropriate for the pharmacist prescriber within their agreed area; lawful but outside agreed scope for the paramedic prescriber, whose correct action is referral; plan-dependent for the supplementary prescriber; and a defined-boundary question for the physiotherapist prescriber, whose entitlement maps to their practice area. One answer, five different right actions.
What this means for AI use, concretely
Three behaviours convert the analysis into practice. State the stack in the prompt: profession, annotation, role, setting, agreed scope and jurisdiction, one line, which lets a well-designed system frame its answer honestly and lets you notice when a generic one has ignored it. Read every answer as evidence, never as authorisation: access to a correct answer creates no authority to act on it, and the sentence deserves to be internalised exactly because fluent answers make actions feel routine, the automation-bias mechanism applied to scope. And run the escalation tree wherever the layers disagree: clinically indicated but outside your annotation, refer to a prescriber who holds it; within annotation but outside personal competence, the framework's answer is develop or refer, never improvise; within competence but outside organisational scope, the route is governance, not workaround; and within everything but outside the clinical management plan, the plan is renegotiated, not overridden. Evidence tools, this platform included, sit correctly on one side of this line: iatroX supports the retrieval and understanding of evidence, and nothing about a retrieved answer, however well sourced, alters what any individual prescriber is authorised to do with it.
Frequently asked questions
Should AI tools refuse to answer outside a user's scope?
Refusal is the wrong instrument, because scope is invisible to the tool and legitimate uses, learning, preparation for referral, team discussion, span every scope; honest framing plus the user's stated stack is the workable design.
How should scope questions be documented when AI was involved?
As ordinary professional reasoning: the clinical picture, the evidence consulted, the scope determination and the action, referral included, with the AI's role as evidence retrieval; a documented referral for scope reasons is competent practice, not an admission.
Does expanding my scope change any of this?
It changes the stack's contents through the proper machinery, training, assessment, agreement, and changes nothing about the method: the layers get re-checked, the prompt line gets updated, and the answers stay evidence.
Does stating my scope in prompts risk anchoring the answer too narrowly?
Ask both ways when learning: the unscoped answer shows the clinical landscape, the scoped answer shows your lawful corner of it, and the difference between them is precisely the referral conversation you may need to have.
How do employers' scope documents stay current with expanding roles?
Through the governance route the escalation tree names: role expansion is agreed, trained and documented before practice changes, and an AI answer describing what expanded practice would look like is preparation material for that conversation, not permission to begin.
Where do community practitioner formulary prescribers fit?
Inside the same stack with a narrower legal layer: their defined formulary is the profession-specific limit made explicit, and the method is unchanged, evidence read as evidence, actions bounded by the entitlement actually held.