INTRODUCTION Cataract remains one of the most important and most easily reversible causes of blindness worldwide.1 In addition, Cataract surgery continues to be framed by global health agencies as among the most cost-effective interventions in health care.2,3 But if we are honest about what challenges the quality of service offered to patients today, the ‘cataract conversation’ is no longer only about surgical technique.4 It is increasingly about clinical systems; how efficiently patients move through preoperative assessment, get their biometry done, movement in and out of the operating theatre, the most efficient use of postoperative surveillance with compliant follow-up, and how safely that entire pathway performs at scale. In mature systems, the reason this conversation is easier is that national audits show what ‘good’ looks like at the population level. The UK National Ophthalmology Database (NOD) cataract audit, which spans very large numbers, shows that most patients achieve good functional outcomes. It also benchmarks key complications in a way that allows services to learn and improve rather than blindly guess. In the NOD report summary table, 91.7% of cases achieved postoperative presenting visual acuity of 0.30 LogMAR (6/12) or better. Posterior capsular rupture (PCR) was tracked as a core safety metric in this report.5 Against that backdrop, artificial intelligence (AI) is best understood not as a replacement for cataract surgery, but as a set of tools that can tighten the operational chain. It is best seen as a tool which can aid in reducing avoidable variation and improving follow-up reliability. Thus, saving the surgeon’s valuable time. The real question is not ‘can AI be used?’ or ‘should it be used’. The question is: ‘where does it add measurable value compared to current best practice?’ and ‘how do we prevent digital enthusiasm from outrunning clinical evidence?’ METHODS This article is written as an editorial/perspective article, not a systematic review. Therefore, no primary data were collected, nor was a statistical analysis of a local patient cohort presented. The approach is a focused, operations-oriented narrative which is anchored to (i) benchmark expectations from national audit reporting,5 (ii) guideline standards shaping postoperative follow-up models,6 and (iii) looking into peer-reviewed studies examining AI performance in cataract-related tasks that have clear operational consequences. With special focus on refractive prediction and postoperative triage.7-12 The intent is to translate high-signal evidence into practical implications for cataract services. This is particularly important in settings where geographical isolation and workforce constraints make consistency and patient compliance hard. RESULTS First, it is important to name the baseline. Cataract surgery is already highly effective when performed within stable systems, and this is repeatedly demonstrated in large datasets. The NOD audit illustrates strong functional visual outcomes at scale, alongside benchmarked complication monitoring.5 Similarly, very large registry analyses demonstrate that PCR, even if uncommon, remains measurable and trackable across years and centres. In a European registry analysis of over 2.85 million cataract surgeries, 1.1% were complicated by PCR (with variation by year).7 These references are important because they frame the goal for AI in cataract surgery correctly. The aim is not to ‘make cataract surgery work’, but to reduce variation around a high baseline and to help services reproduce good outcomes more reliably. Second, refractive planning remains one of the most measurement-sensitive, high-volume steps in cataract operations. It is therefore one of the most plausible places for AI to add operational value. Conventional intraocular lens (IOL) power formulas have been compared rigorously. In a large study evaluating 18,501 cases, modern formulas (e.g. Barrett Universal II formula) demonstrated small median absolute prediction errors, but still with meaningful differences between formulas and patient subgroups.8 Other comparative work across multiple formulas similarly demonstrates that ‘good outcome’ is achievable, but this is not uniform. There is considerable room to reduce refractive surprises, particularly in eyes that challenge standard assumptions (long/short axial length, prior corneal surgery and unusual keratometry).9 Within that context, AI-based methods become operationally interesting when they show improvements on metrics that matter to both patients and surgeons. Namely, outcomes with tighter prediction error distributions and higher proportions of eyes which lands closer to the target refraction. In one study describing an AI-based IOL calculation method, the AI approach reported a smaller spread of prediction error (standard deviation 0.30) and a higher proportion within ±0.50 D compared with several established formulas (with the AI group reporting 90.38% within ±0.50 D).10 Even when differences seem numerically modest, the operational impact is straightforward. Fewer refractive surprises means fewer dissatisfied and ‘uncomplicated’ post-ops returning for time-consuming explanation visits. This leads to fewer downstream spectacle dependence complaints in patients expecting emmetropia as well as a clearer separation of ‘true complications’ from perfect quality-of-vision concerns. Third, imaging-based AI for cataract detection and cataract grading can support standardisation in referral and preoperative documentation workflows. In a validation study using slit-lamp and retroillumination photographs, deep learning models demonstrated high discrimination for cataract detection and grading, which matters operationally because it can reduce inter-observer variability and improve the consistency of baseline documentation. This is an important application in an operational system in mixed referral pathways and settings where ophthalmologists are rare and their time is scarce.11 This would not remove the need for clinician assessment, but it can make triage and surgical decision-making more consistent. It will also improve how reliably patients enter the right part of the pathway at the right time. Finally, postoperative follow-up is where the ‘clinical operations lens’ becomes unavoidable. Modern guidelines already push services away from the universal early review after uncomplicated surgery, which is considered as the normal practice among many surgeons. NICE guidance advises against routine first-day postoperative review in uncomplicated cataract surgery and emphasises on follow-up that is tailored to risk rather than habit.6 The real question is simple: ‘How do we keep patients safe without filling our clinics with visits that don’t actually change management?’ That is where AI triage makes sense. It is basically built for the ‘most people are fine, but a few need urgent attention’ reality of post-op care. In a large evaluation of an autonomous AI clinical assistant (Dr Dora) used for postoperative cataract follow-up, the system showed strong performance in detecting complications and was able to handle most follow-up interactions on its own. The AI clinical assistant was also able to escalate the smaller, higher-risk group to clinicians for review.12 Operationally, that is the blueprint many strained systems are trying to achieve manually, which is, fewer routine visits for low-risk patients and faster escalation for symptoms that matter. DISCUSSION A useful way to judge the role of AI in cataract surgery is to ask a deliberately comparative question. ‘Relative to current best practice, what problem does this tool solve and what new failure modes does it introduce?’ The reason this particular framing matters is that cataract surgery outcomes (when services are well organised) are already strong and audits like NOD demonstrate that clearly.5 Therefore, the argument for AI should be uniformly reduce variation in all measurements and improve reliability in cataract surgery. In refractive outcomes, conventional formulas remain evidence-based and perform well at scale.8,9 Hence, AI only earns its place when it demonstrates measurable improvement on patient-relevant targets (e.g., proportion within ±0.50 D), or when it improves performance in subgroups where traditional formulas struggle. The operational consequence of better refractive prediction is not academic. In the system, it shows up as fewer unplanned postoperative consultations, fewer unhappy patients and a smoother separation of refractive counselling from complication management. For imaging-based detection and grading, AI’s comparative advantage is standardisation. When deep learning models show strong discrimination on slit-lamp/retroillumination photographs,11 they can support more consistent referral quality. The risk, however, is deployment drift. A model that performs well in one dataset may degrade across different camera systems, lighting conditions, cataract morphologies, internet speeds and across different populations. That means operational governance is not optional. If a service adopts an AI triage or grading tool, it needs periodic performance monitoring and a clear clinician override pathway. In small systems, the governance piece can actually be simpler if it is designed intentionally, since fewer sites and fewer workflows would mean better accountability. Moreover, since cataract outcomes are not only about the lens, AI-enabled retinal image analysis can strengthen pre-op decision-making by flagging co-pathology that changes prognosis (e.g., referable diabetic retinopathy). This way patients are not overpromised perfect vision, and follow-up will be planned accordingly as per risk stratification method.13 Postoperatively, the comparative discussion should be grounded in what modern guidance already encourages. Focus on risk-stratified follow-up and strong safety-netting, rather than automatic next-day review for everyone.6 AI triage tools such as Dr Dora are persuasive because they formalise that principle into a scalable workflow. It also has high reported performance for detecting complications and the ability to complete most interactions autonomously while escalating the remainder.12 But the operational question that matters is the edge case, which is ‘what happens when a patient does not engage digitally, misunderstands questions or provides incomplete information?’ The answer is not to reject AI. The answer is to design hybrid models intentionally with clear escalation thresholds. Symptom questions that actually fit the patient’s language and culture, clinician dashboards that are genuinely practical (and not another logistical hurdle), and a proper plan for non-responders, so the system doesn’t just go quiet and log the patient as mute. For small-island and geographically dispersed settings, the conclusion is pragmatic. AI will not substitute for surgical training, theatre capacity, sterilisation standards or a functional referral network. But it can empower and help improve the efficiency of surgeons by helping reduce friction at predictable bottlenecks. The strategic priority should be to implement AI where it can be locally validated, where it produces measurable operational gains (wait times, follow-up completion, and escalation latency), and where clinician accountability remains clear. CONCLUSION Cataract surgery outcomes in high-performing systems are already strong, as evidenced by national audits and modern guideline approaches that support streamlined, risk-stratified follow-up.5,6 AI’s credible role is therefore not a rhetorical transformation, but it is one of operational refinement. Improving refractive prediction, standardising triage, standardising documentation and extending safety-netting through scalable symptom screening are extremely important roles AI can play in cataract surgery. Where the evidence supports high performance in real workflows, particularly in postoperative triage models that escalate risk appropriately, AI can function as capacity protection rather than clinical replacement.12 For health systems modernising cataract pathways under resource constraints, the central task is disciplined implementation. This is by local validation, governance, increasing accountability and outcomes measurement, so that AI becomes a lever for quality rather than a new source of variability.
Saraa Yoosuf (Thu,) studied this question.