Objectives To evaluate the cost-effectiveness of the implementation of the Identification and Referral to Improve Safety (IRIS) programme using up-to-date real-world information on costs and effectiveness from routine clinical practice. A Markov model was constructed to estimate mean costs and quality-adjusted life-years (QALYs) of IRIS versus usual care per woman registered at a general practice from a societal and health service perspective with a 10-year time horizon. Design and setting Cost-utility analysis in UK general practices, including data from six sites which have been running IRIS for at least 2 years across England. Participants Based on the Markov model, which uses health states to represent possible outcomes of the intervention, we stipulated a hypothetical cohort of 10 000 women aged 16 years or older. Interventions The IRIS trial was a randomised controlled trial that tested the effectiveness of a primary care training and support intervention to improve the response to women experiencing domestic violence and abuse, and found it to be cost-effective. As a result, the IRIS programme has been implemented across the UK, generating data on costs and effectiveness outside a trial context. Results The IRIS programme saved £14 per woman aged 16 years or older registered in general practice (95% uncertainty interval -£151 to £37) and produced QALY gains of 0.001 per woman (95% uncertainty interval -0.005 to 0.006). The incremental net monetary benefit was positive both from a societal and National Health Service perspective (£42 and £22, respectively) and the IRIS programme was cost-effective in 61% of simulations using real-life data when the cost-effectiveness threshold was £20 000 per QALY gained as advised by National Institute for Health and Care Excellence. Conclusion The IRIS programme is likely to be cost-effective and cost-saving from a societal perspective in the UK and cost-effective from a health service perspective, although there is considerable uncertainty surrounding these results, reflected in the large uncertainty intervals.