AI patient triage from a web form

When a patient wants an appointment, they fill in a web form describing their symptoms. That submission lands with a nurse who reads it, draws on clinical experience to judge how urgent the case is, picks the appropriate appointment type, and arranges the booking. If the case does not fit a clear pattern, it gets passed to a clinician for a second opinion. The whole cycle runs roughly 9,000 times a year, takes about seven minutes per case, and adds up to well over a thousand hours of nursing time annually. Right now that time is coordinated through email and spreadsheets alongside a practice management system whose API has never been connected to anything.

The process sits at the intersection of operational routine and genuine clinical judgment. The majority of submissions follow recognisable patterns that an experienced nurse handles quickly and consistently. A meaningful minority, somewhere between one case in ten and one in seven, involves symptoms that are ambiguous, incomplete, or outside standard patterns and requires a more careful human decision. The stakes are not abstract: a wrong urgency call is a patient-safety event, and all of the data involved is GDPR special-category health information subject to regulated retention rules.

The practice is small. The Practice Manager oversees the process, three clinicians provide the clinical safety net, and one part-time administrator handles overflow. There is no in-house developer and no existing automation infrastructure. The opportunity is real and the need is clear, but any solution has to work within those constraints from day one.

The right approach here is a Hybrid: an AI workflow that handles the predictable majority of triage submissions automatically, paired with a mandatory human review gate for every case that falls below a defined confidence threshold. This is not a compromise between automation and caution, it is the architecture that actually matches the shape of the process. Around 85 percent of submissions follow patterns that an AI classifier can handle reliably once it has been trained on real examples from this practice. The remaining 10 to 15 percent require the kind of nuanced judgment that no current AI model should be trusted to make alone in a patient-safety context. The Hybrid path serves both populations correctly rather than forcing a single approach onto a process that genuinely has two distinct case types.

The practice has no in-house developer and has expressed no platform preference, which shapes the recommendation in a direct way. A platform like Make or n8n, or a healthcare-oriented low-code equivalent, can connect to the practice software's published API without custom code, host the AI classification step, and route exceptions to a human review queue using built-in workflow logic. That keeps the build within reach of a consultant-led implementation and avoids creating a dependency on ongoing developer support. The practice software API is a meaningful asset here, even though it has never been used. Connecting the approved triage outcome directly to the patient record and booking workflow eliminates the current manual handoff through email and spreadsheets, which is where a significant amount of the friction and error risk lives today.

A Fully Automated AI Workflow was scored and considered seriously. The case against it is not that AI classification is unreliable in general, it is that this specific practice's 10 to 15 percent exception rate is too high to accept without a human gate, and that the consequences of a misclassified high-urgency case are severe enough to make the risk unacceptable at this stage of deployment. If the model's real-world accuracy is validated over time and the exception rate proves lower than expected in practice, the confidence threshold can be raised and the human review burden reduced without rebuilding anything. The Hybrid architecture is designed to evolve in that direction; full automation is a destination the practice could reach, not a starting point it can safely adopt now.

Traditional Code and RPA were both assessed and neither fits. A rule-based coded system cannot handle the natural language variability of free-text symptom descriptions, and maintaining it as clinical patterns change would require ongoing developer cost the practice does not have the structure to support. RPA is built for structured, UI-driven repetitive tasks and has no capability to interpret clinical language or make triage judgments. It could theoretically assist with downstream data entry but cannot perform the core classification step, making it irrelevant to the main problem. An autonomous AI Agent introduces unnecessary complexity and removes the human oversight that compliance and patient safety both require. Staying Manual is not the cautious option it might appear to be. Over a thousand hours of nursing time per year, coordinated through email and spreadsheets, with tacit knowledge concentrated in two or three people, is itself a fragile and expensive situation that grows riskier as volume increases.

One tradeoff worth naming honestly is that the human review queue is a permanent feature of this design, not a transitional one. The nurse or administrator reviewing flagged cases will still spend time on exceptions, and that time needs to be factored into the ROI model rather than treated as a cost that disappears. The projected 65 percent reduction in handling time accounts for this. The goal is not to remove people from triage entirely but to remove them from the routine majority so they can apply their judgment where it genuinely matters. That is a meaningful improvement in how clinical staff spend their working day, and it is the honest version of what this automation delivers.

The two most competitive paths here are the Hybrid approach and a fully Automated AI Workflow, and the gap between them is meaningful in this specific context. A pure AI Workflow without a mandatory human gate could theoretically handle the structured majority of cases well, but the 10 to 15 percent exception rate is not a rounding error. At 9,000 submissions per year, that is 900 to 1,350 cases annually where the AI would be operating outside the patterns it was trained on, and in a patient-safety environment a confident wrong answer is worse than a flagged uncertain one. The Hybrid path solves this directly: it auto-processes the routine majority while routing ambiguous, incomplete, or high-urgency cases to a nurse or clinician before any action is taken. The ROI difference between the two paths is modest, since the human gate is a permanent cost either way, but the Hybrid genuinely earns its keep on liability and compliance grounds alone.

Traditional Code was considered but scores poorly for this process. A rule-based system could codify the clearest triage patterns, but symptom descriptions arrive as free text from patients who do not use clinical terminology consistently. A rules engine cannot interpret that variability, and the practice has no in-house developer to maintain the brittle logic as symptom patterns or clinical criteria shift over time. The cost of keeping traditional code current would quickly erode any initial savings. RPA and an autonomous AI Agent were also assessed and both are poor fits: RPA has no ability to interpret natural language, and a fully autonomous AI Agent introduces unacceptable clinical risk without the defined human checkpoints that this process demands. The Hybrid path is not just the highest scorer; it is the only path that honestly addresses all three constraints at once, which are the compliance requirements, the judgment-heavy exceptions, and the no-developer operating environment.

The build starts with a discovery and data mapping phase before any platform is configured. The practice needs to pull 200 to 300 historical triage submissions and their recorded outcomes, anonymise them properly under GDPR before they touch any external system, and use that sample to understand the real distribution of case types and the symptom language patterns that nurses actually encounter. At the same time, the team maps the practice management software's published API to identify the endpoints needed for reading patient records and writing triage outcomes back in. This phase, which should take two to three weeks, produces both the validation dataset and the integration specification that everything else is built against.

With that foundation in place, the AI classification layer is configured on a low-code workflow platform. Make or n8n are both capable options given the no-developer constraint, and both can call REST APIs out of the box without custom code. The platform is configured to receive incoming form submissions, pass the symptom text through an AI classification model, and produce a suggested urgency level and appointment type alongside a confidence score. The critical configuration decision here is setting the confidence thresholds: cases above the threshold are auto-routed directly to the practice software, and cases below it are held and forwarded to the human review queue. A clinician must sign off on the threshold settings before any live patient data flows through the system, and the initial thresholds should be set conservatively.

The human review queue is the heart of the hybrid model and needs to be genuinely fast to use. Whether it is built as a lightweight web dashboard or delivered as a structured email with approve and override options, the nurse or administrator reviewing a flagged case should be able to see the original submission, the AI's suggested classification, and the reasoning behind it, then make a decision in under two minutes. Cases requiring clinical judgment beyond the nurse's remit escalate to a clinician through the same interface. Once a decision is made, the outcome is written to the practice software via API, triggering the booking workflow and eliminating the current manual handoff through email and spreadsheets.

The final phase is a supervised pilot running across at least four weeks of live volume. During the pilot, a clinician audits a random ten percent sample of auto-approved cases to validate that the confidence thresholds are performing as expected on this practice's specific patient population. Every decision, whether automated or human-reviewed, is logged with a timestamp and the identity of the reviewer, satisfying both the GDPR audit trail requirement and the regulated records retention obligation. Threshold adjustments based on pilot findings are made before full production go-live, and a monthly review cadence is established from day one so the model does not drift silently over time.