AI patient triage from a web form

This process begins when a patient submits their symptoms through the practice's web form. A nurse then reads the submission, applies clinical experience to judge urgency, assigns an urgency level (acute or can wait), selects the appropriate appointment type, and notifies the patient or progresses the booking. Cases that do not fit standard patterns, contain data entry errors, or require senior clinical judgment are handled as exceptions and reworked manually. The whole cycle runs multiple times each day, accumulating around 9,000 submissions per year.

The process is largely routine for the 85 to 90 percent of cases that match recognisable symptom patterns, but the remaining 10 to 15 percent demand genuine clinical reasoning that cannot be reduced to simple rules. Because the inputs arrive as free-text descriptions rather than structured fields, the task requires natural language understanding, not just data routing. The primary system is the practice's vendor software, which exposes an API that has never been connected to anything. Day-to-day, the team relies on email and spreadsheets to fill the gaps.

The compliance context is demanding. Patient symptom data is GDPR special-category health data, records retention is regulated, and a wrong triage decision carries direct patient-safety implications. The practice has no in-house developer, so any solution must be implementable and maintainable without one. At seven minutes per case and 9,000 cases per year, the process consumes roughly 1,050 staff hours annually, almost all of it nurse time that could otherwise go to direct clinical care.

The recommended path is Hybrid: an AI workflow that classifies and routes the routine majority of cases automatically, with a mandatory human review queue that holds every high-urgency flag and every case the classifier is not confident about until a clinician confirms the decision. This is not a compromise between automation and safety. In a patient-safety context, the human gate is a design requirement, and any architecture that removes it in the name of efficiency would be irresponsible regardless of how well the AI performs in testing.

The case for automating the routine 85 to 90 percent is straightforward. Those cases follow recognisable patterns that experienced clinicians apply consistently, which means a well-calibrated AI classifier can replicate that judgment reliably. Freeing nurses from the repetitive majority lets them concentrate on the exception segment, where their clinical skill genuinely matters, and on direct patient care. The ROI is strong: even a conservative 70 percent reduction in manual handling time represents around 735 hours saved per year against a build cost that the practice should recover within nine months.

The practice has no developer resource and no stated platform preference, so the implementation should use a low-code AI workflow platform. Make or n8n are both credible choices here. Either can receive web form submissions as a trigger, call an LLM-based triage classifier, apply a confidence threshold to decide whether to action the result automatically or route it to the human review queue, and pass confirmed decisions through to the practice software via its existing API, all without writing custom code. This is an important practical constraint and the recommendation is shaped by it. A bespoke-coded solution or a fully autonomous AI agent would require developer capability the practice does not have and cannot easily acquire, making those paths non-starters regardless of their theoretical merits.

Fully Automated AI Workflow scored lower than Hybrid precisely because removing the mandatory human gate is not a tuning decision but a safety decision. The classifier will occasionally be wrong, and in a triage context a false negative, an acute case classified as routine, is a potential patient harm event. Keeping a clinician in the loop for all high-urgency and low-confidence cases does not undermine the value of the automation; it makes the automation safe enough to trust. AI Agent was not seriously considered for this process. An autonomous agent capable of acting on triage decisions without human confirmation would introduce serious liability, and the practice has neither the technical resource nor the governance infrastructure to operate one responsibly. RPA and Traditional Code were both ruled out early: RPA cannot interpret free-text symptom descriptions, and a coded rules engine would require a developer, would struggle with natural-language variability, and would need constant maintenance as clinical guidelines change.

The honest tradeoff in the Hybrid approach is that it does not fully automate the process. A clinician will still spend time on exceptions and on reviewing high-urgency flags. The system will need quarterly governance cycles, a named clinical safety owner, and periodic retraining as guidelines evolve. These are not arguments against the recommendation; they are the conditions under which the recommendation works safely. A practice that treats the human review step as a temporary inconvenience to be optimised away over time would be misusing the system. The quarterly governance cycle and the conservative confidence threshold at launch are non-negotiable parts of the design, not optional extras.

The two most credible paths for this process are Hybrid and AI Workflow, and the gap between them comes down to one question: what happens when the system gets it wrong? An AI Workflow operating without a mandatory human gate could handle the structured majority of submissions effectively, and on a pure throughput measure it looks attractive. But in a patient-safety context, an unreviewed false-negative triage decision, where an acute case is classified as routine and the patient waits, is not an acceptable failure mode. The AI Workflow path scores well on efficiency and falls down on risk governance, which in a regulated healthcare setting is disqualifying on its own.

The Hybrid path resolves this by treating the human review gate not as a workaround but as an intentional part of the design. Routine cases, the reliable 85 to 90 percent that follow recognisable symptom patterns, are classified and routed automatically. Any case that carries a high-urgency flag or falls below the classifier's confidence threshold is held in a queue for a clinician to confirm before any action is taken. This means the system is fast where speed is safe, and cautious where caution is clinically necessary. That is exactly the right trade-off for this practice.

The AI Agent path can be set aside quickly. An autonomous agent making and acting on triage decisions without human confirmation introduces liability that no general practice should accept, and the practice has no developer resource to build, test, or govern one safely. RPA and Traditional Code both fail at the first hurdle: neither can interpret the free-text symptom descriptions that arrive from the web form, so they would automate the data-routing shell of the process while leaving the actual triage judgment untouched. Education alone does not reduce the 1,050 hours of annual staff time consumed by a process the team already understands well. Hybrid is the clear recommendation not because the other paths are close, but because it is the only path that delivers the efficiency gain without compromising the patient-safety requirement.

The implementation begins with a focused discovery exercise before any technology is selected or configured. Over two to three weeks, a clinician and the practice manager work through six months of historical web form submissions, mapping the symptom language patients actually use to the urgency levels and appointment types that were assigned. This produces the labelled examples the AI classifier needs and, just as importantly, surfaces the boundary conditions: the symptom combinations that experienced nurses treat as routine versus the ones that prompt a second opinion. Any recurring data quality issues in the web form, missing fields, ambiguous symptom descriptions, are identified and corrected at the source where possible.

With that mapping in hand, the team selects a low-code AI workflow platform. Make or n8n are both practical choices here: both offer webhook-based triggers that can receive form submissions directly, both have connectors to common practice management APIs, and neither requires a developer to configure standard workflow logic. The web form is connected to the platform so that every new submission fires the workflow automatically. Before any patient data touches the system, the GDPR obligations are addressed: a Data Processing Agreement is put in place with the platform provider and with the LLM provider used for classification, a Data Protection Impact Assessment is completed, and the workflow is configured with access controls, data minimisation, and a tamper-evident audit log that records every triage decision and its outcome.

The AI classifier sits at the heart of the workflow. It receives the structured form fields alongside the free-text symptom description and outputs two things: an urgency level, either acute or can wait, and a confidence score. The confidence threshold is set conservatively at launch, with clinical input, not optimised for throughput. Any case that scores below the threshold, and any case where the classifier assigns an acute urgency, is automatically placed in a human review queue rather than actioned. A clinician opens the queue item, sees the AI suggestion alongside the original submission, and either confirms or overrides it. On confirmation, the workflow calls the practice software API to pass the urgency level and appointment type through, eliminating the manual re-entry step that currently ties up the administrator.

Before go-live, the hybrid system runs in parallel with the existing manual process for four weeks. Every AI decision is compared against the nurse's independent assessment on the same case. Disagreements are reviewed with the clinical team to refine classifier thresholds and, where needed, add new examples to the training set. Once agreement on routine cases reaches the agreed clinical threshold, recommended at 95 percent or above, the parallel manual step is retired and the system goes live as the primary triage pathway. The validation evidence from the parallel run is documented and retained as the regulatory record of the clinical safety assessment.