Autonomous triage may reduce the work of reviewing and routing requests, and GP Triage's early practice reports describe substantial reductions. They do not yet establish a general, independently verified saving for every practice. The critical distinction is between work no longer performed, time estimated from activity and money actually released from a budget.
At Langton Medical Group, the supplier reports about 30 GP triage hours released each week. At Swanscombe, it calculates time savings from automated booking volumes. Those are different forms of evidence and should remain different throughout an appraisal. Langton, October 2025 to February 2026; Swanscombe, February to March 2026
Start with the work, not the saving
Before assessing a claim, describe the previous pathway. Who opened a request? Who interpreted it? Who decided on urgency? Who contacted the patient or booked the appointment? Did every request reach a GP, or only selected ones?
The opportunity for automation depends on that baseline. Removing a GP review step from a heavily clinician-led process is not the same intervention as replacing an already brief administrative step. Applying the same assumed time saving to both would hide the difference.
Also identify the unit. A patient can submit several requests. One request can generate several contacts. One booked appointment can involve work by more than one person. The evaluation needs to know which of those units its activity count represents.
These are proposed measurement principles, not an assertion about how either practice's unpublished records were collected.
Langton's headline needs a visible costing model
GP Triage's Langton case study reports approximately 30 hours of GP-led triage removed weekly and around £163,000 in annualised savings. It attributes the figures to platform and practice reporting during October 2025 to February 2026, with annual savings confirmed by practice partners. That is useful attribution, but not an external financial audit. Published case study
An illustrative reverse calculation shows the missing assumption. Using 52 weeks solely for this calculation, 30 hours each week produces 1,560 hours. Dividing the advertised annual figure by those hours gives approximately £104.49 per hour. This is our arithmetic on the published estimates, not a reported salary rate, invoice or validated economic model.
The distinction between cash and capacity remains essential. A practice may redirect staff time into care without reducing its payroll. That can be valuable, but it is not the same as cancelling a recurring payment. Conversely, reducing paid additional sessions may produce a cash effect that can be checked against expenditure.
The case study also describes four sessions and includes a statement about two to three hours a day. The published page does not provide enough detail to reconcile all these descriptions. Session length and the number of staff involved should be confirmed rather than invented.
Swanscombe's hours are calculated, not directly observed
For the four weeks beginning 23 February 2026, GP Triage reports 5,068 autonomous bookings at Swanscombe. Multiplying that count by its five-minute assumption gives 25,340 minutes, or approximately 422.3 hours. The calculation explains the headline; it does not establish that every booking would otherwise have consumed five minutes of staff time. Swanscombe case study and calculation basis
The page describes the assumption as an NHS benchmark, but does not link the underlying benchmark source. Until that basis is supplied, it should be reported as the assumption used in the supplier's model, not as an independently verified national standard.
Nor are all of those hours necessarily GP hours. The case study describes clinical and administrative time together. Converting the total into GP sessions or a GP-only saving would change the meaning of the published figure.
The model is still useful as a starting hypothesis. A practice could sample its own manual process, measure the remaining automated-pathway work and replace the assumption with locally observed values.
A net workload ledger is more useful than an automation counter
A prospective evaluation could use the following ledger. It is an original suggested framework, not a description of data already collected at either practice.
| Work category | What to record | Why it belongs in the calculation |
|---|---|---|
| Manual assessment avoided | Time by staff role and request type | Establishes which work genuinely disappeared |
| Booking and navigation avoided | Time for the previous booking process | Separates clinical from administrative work |
| Exceptions retained | Reviews, failed bookings and rerouted requests | Prevents gross automation from being mistaken for net reduction |
| Patient support | Telephone help, reception help and repeated submissions | Captures work needed to keep access usable |
| Clinical recontact | Follow-up generated by an incomplete or unsuitable route | Identifies work displaced later in the journey |
| Implementation and oversight | Training, configuration, monitoring and incident review | Captures resources required to operate the service |
The resulting calculation is conceptually simple: work avoided minus work retained or introduced. The difficult part is collecting both sides consistently.
Measure separately by role before placing a monetary value on the result. A combined total can obscure a reduction in one team's workload alongside an increase in another's. That may still be a worthwhile redesign, but it should be visible.
Safety and capacity should be measured together
A fall in manual triage work does not establish that requests were handled appropriately. Equally, a high level of booking automation is not inherently unsafe. Both claims require outcomes beyond a transaction count.
A useful evaluation would track unresolved requests, repeated contacts, clinician overrides, changes in urgency, complaints and relevant adverse events. Where feasible and appropriately governed, it should examine subsequent urgent-care use rather than infer it from the advice initially displayed.
Continuity also deserves a place. A faster booking can still be a poor operational result if it repeatedly separates a person with a complex ongoing problem from the clinician or team responsible for it. The importance and measurement of that trade-off should be defined with the service, not assumed from a generic automation target.
Age, language, disability and deprivation should be considered in the evaluation plan. The question is not simply who completed the digital process, but who encountered difficulty, used an alternative route or did not obtain a completed next step.
What Rapid Health adds to the evidence discussion
Rapid Health has an externally conducted implementation evaluation by Unity Insights, commissioned with Health Innovation Kent Surrey Sussex and Rapid Health. It examined one practice before and after deployment. The published summary gives a more structured account than a testimonial, while still falling short of a randomised comparison. Unity Insights evaluation
The accompanying September 2024 report also records limitations in the baseline data and mixed patient and staff feedback. That is valuable because a service can improve an operational measure without every group experiencing the change positively. Evaluation report
This evidence supports a different kind of question: what happened during this particular implementation? It does not establish that Rapid Health would outperform GP Triage at Langton, or that either supplier's results would transfer unchanged to another practice.
International engine evidence answers another question again
Infermedica's clinical assessment evidence should not be confused with the financial performance of GP Triage. For example, a 2024 Journal of Hospital Administration study compared AI-based virtual triage with rules-based triage using 149 clinical vignettes. Its design concerns triage assessment, not the net workload of an NHS booking implementation. Published comparative study
An evidence review should therefore separate supplier operational reports, implementation evaluations, component validation and prospective patient-outcome studies. A large number of assessments internationally cannot fill a missing local costing assumption. Equally, a favourable cost estimate cannot answer a question about under-triage.
What an informative NHS evaluation would look like
A practical design would establish the baseline before changing the pathway, define the intended population and follow implementation long enough to distinguish settling-in effects from sustained changes. Comparable practices or a phased rollout could strengthen interpretation, provided the analysis accounts for the design and relevant service changes.
The outcome set should be agreed in advance. A practice could define success as releasing clinician capacity while maintaining timely escalation, acceptable patient experience and workable non-digital access. That is more meaningful than maximising the proportion of requests that avoid an inbox.
For a GP-led service with substantial manual review, the Langton report offers a useful hypothesis to test. For a large practice examining booking administration, Swanscombe's activity model offers a starting calculation. For commissioners, neither replaces a net-cost and patient-outcome evaluation across the proposed population.
This comparison is published by iatroX, which provides clinical-reference and educational tools rather than patient booking. Its role here is to examine the evidence and the measurement question, not claim the operational benefits of a different product category.
The most credible time-saving claim names the task, the staff role, the period, the comparator and the work that remained. Everything else is still a hypothesis about capacity.
Frequently asked questions
Has GP Triage independently proved a £163,000 annual saving at Langton?
The case study reports approximately £163,000 in annualised savings from platform and practice reporting for October 2025 to February 2026. It does not provide an independent financial audit or a complete public net-cost model.
Were Swanscombe's 422 hours measured with a time-and-motion study?
The published February to March 2026 estimate applies five minutes to each of 5,068 autonomous bookings. It is a modelled time estimate rather than a directly observed duration for every avoided task.
What should a practice measure before buying autonomous triage?
Measure the existing work by role and request type, then account for exceptions, patient support, oversight and subsequent contacts. Assess that net workload alongside clinical and access outcomes, not automation alone.