Same-day care is where ANP practice is most exposed and most distinctive: undifferentiated presentations, autonomous assessment, decisions across a whole episode, and time pressure that punishes every inefficiency. It is also where AI support is most tempting and where the central design failure matters most: an AI system should never collapse "likely diagnosis", "management option" and "safe prescription" into one unsupported sentence, because those are three different judgements with three different evidence bases, and a workflow that keeps them separate is the whole stack this article describes, run on four fictional presentations to show its shape.
The workflow, stated once
Seven moves, each with its proper tool. Structure the differential: the reasoning layer, Brainstorm-style structured thinking that widens before it narrows, because same-day errors are anchoring errors more often than knowledge errors. Identify immediate risk: red-flag recognition against the presentation, the step that decides setting before anything decides treatment. Find the national pathway: the guideline layer through Ask-iatroX, CKS-shaped for most same-day work, sources open. Check local guidance: antimicrobial policy, referral thresholds and formulary, the operational layer that governs, per the local-first logic at /blog/ai-antimicrobial-prescribing-local-guidance. Confirm medicine-specific cautions: the exact product against this patient's age, renal function, interactions and allergies, ending at the SmPC. Produce patient information and safety-netting: drafted, reviewed, personalised, with the deterioration criteria explicit. And record unresolved uncertainty: the honest note of what was not excluded and what review will catch it, which is same-day medicine's real safety mechanism.
Four fictional presentations, one discipline
Dysuria with possible upper-tract features: the workflow's value is the second move, systemic features and loin involvement change the question from medicine choice to severity assessment; then local policy for the antimicrobial layer, renal function into the product check, and safety-netting that names progression. Wheeze in a patient with cardiac and respiratory comorbidity: the first move carries the case, because "wheeze" spans two organ systems and the structured differential resists premature closure; guidance follows the working diagnosis honestly reached, and the uncertainty note earns its place if the picture stays mixed. New unilateral leg swelling: risk stratification is the case, structured assessment and the appropriate probability tooling, with the pathway dictating investigation and the prescription question, where one arises, product-checked against renal function and interactions; the AI's job here is keeping the sequence honest, not skipping to an answer. And a child with fever and reduced intake: the red-flag layer leads absolutely, national febrile-child guidance frames assessment, and everything product-shaped runs at the paediatric standard, weight, age-specific licensing, exact formulation, which is red-flag territory in the 90-second screen's terms, /blog/verify-ai-prescribing-answer-90-seconds, promoting the whole case to full-review discipline.
Why the collapsed answer fails, and what the stack buys
The collapsed answer, one fluent paragraph from symptom to prescription, fails because it hides its three judgements from inspection: the diagnosis cannot be checked without the differential it skipped, the management option cannot be weighed without the pathway it summarised away, and the prescription cannot be verified without the product-level check it never ran, so the prescriber inherits confidence with no handholds. The stack buys the opposite: each judgement lands on its proper evidence with its proper source open, the whole run costs minutes distributed across a consultation that was happening anyway, and the residue is exactly the material same-day governance wants, documented reasoning, local-policy compliance, safety-netting and the uncertainty note. The tools named here are this platform's configuration of the stack, Ask-iatroX for guidance orientation, Brainstorm for structure, calculators where scoring belongs, the medicines layer to the SmPC; the discipline is portable to any tools that keep the three judgements separate, and hostile to any that merge them.
Frequently asked questions
Is seven moves realistic at same-day pace?
The moves are how competent same-day care already works, named: the stack changes their cost, not their number, and most cases spend seconds on several moves, with the full weight landing only where red flags or complexity earn it.
Where does remote or telephone assessment change the stack?
At the risk layer: red-flag identification without examination lowers the threshold for face-to-face conversion, and the uncertainty note becomes the consultation's most important artefact.
Should the uncertainty note be shared with the patient?
Its practical half, yes: safety-netting is the uncertainty note translated, what we have not ruled out, what would change the picture, and when to come back, said plainly and documented as said.
How does the stack handle the patient who presents with two problems?
By running the risk layer across both before either gets managed: same-day danger hides in the interaction of presentations, and the structured differential is cheapest exactly when the picture is busiest.
What should be templated versus reasoned fresh each time?
Template the artefacts, safety-netting frameworks, documentation structure, the uncertainty note's shape; reason the judgements fresh, differential, severity, product checks, which is the division the collapsed answer gets wrong in the other direction.