AI-assisted advice and guidance reduces cardiology consultant response time and variability: a pilot crossover study

Background NHS Specialist Advice and Guidance (A&G) enables general practitioners to obtain cardiology input without formal referral, reducing unnecessary outpatient appointments and elective waiting times. National policy aims to double A&G requests to 4 million annually, yet specialist capacity remains static. Current A&G workflows require consultants to draft responses from scratch for each query, creating a bottleneck. Generative artificial intelligence may support this demand-capacity gap by pre-drafting responses for specialist review. Purpose We hypothesised that AI-generated draft responses would reduce consultant response time and standardise output quality without compromising clinical safety. Methods Five consultant cardiologists from a single tertiary centre completed 200 simulated GP queries in a crossover design: each answered 20 queries manually (Batch A) and 20 different queries with assistance from a novel AI application (Batch B), reviewing and editing AI-generated drafts before submission. Queries were designed by the study team to reflect authentic referral patterns across general cardiology, heart failure, arrhythmia, and valve disease. Time-per-response was self-reported in seconds. AI drafts were rated for edit requirements (none, minor, significant) and clinical safety. Conditions were completed on separate days to minimise fatigue effects. Statistical comparison used independent t-test with significance set at p<0.05. Results Consultants responded 25% faster when using AI assistance (mean 95 seconds versus 127 seconds manually; t-test p=0.008). AI assistance also made response times more consistent across consultants, with the spread of times narrowing considerably (IQR 52–120s versus 52–194s; coefficient of variation reduced from 81% to 64%). Notably, four of five consultants were faster with AI, though one consultant with a particularly efficient baseline workflow was slower, suggesting the tool suits some working styles better than others. Half of the AI-generated drafts (50/100) were sent without any changes. The remaining 50 required minor or significant edits, most commonly to adjust referral thresholds or add local pathway information. No responses were judged clinically unsafe. Conclusions AI-assisted A&G drafting reduces mean consultant response time by 25% and standardises output across consultants with varying baseline practice patterns. The 50% unedited rate and 100% safety signal suggests strong alignment between AI-generated content and specialist reasoning. The narrowing of response time variability may have implications for service planning and capacity modelling. Limitations include simulated rather than live queries and self-reported timing. Prospective validation with real GP referrals and formal regulatory assessment is warranted before clinical implementation.