Healthcare technology founder Shirin Krall argues that AI is already more capable in healthcare than we admit. The real challenge is deploying it safely.
Healthcare has become one of the most discussed frontiers for artificial intelligence. From diagnostic models to medical imaging systems, the conversation often centres on whether machines might eventually replace clinicians.
But this framing misses the more immediate reality.
The real question in healthcare AI is not whether machines can perform certain tasks. In many cases, they already can. The harder challenge is deploying these systems safely in environments where the cost of error is high.
Modern AI systems are capable of interpreting unstructured information, summarising complex inputs, identifying patterns across large datasets and generating structured outputs at speed. These capabilities make them particularly useful in areas where healthcare systems struggle most: navigating fragmented information and coordinating operational workflows.
AI can assist with tasks such as patient intake, information triage, documentation, and interpreting unstructured patient queries. By structuring large volumes of information quickly, these systems can reduce much of the operational friction that surrounds care delivery.
The technical capability is advancing rapidly. The harder question is not whether these systems can work, but how they should be integrated into healthcare environments where reliability, accountability, and safety matter more than speed alone.
In other words, the bottleneck in healthcare AI is no longer capability. It is safe system design.
Error tolerance is fundamentally different in healthcare
Healthcare operates under fundamentally different constraints from most industries. In many digital products, occasional errors are inconvenient. In healthcare, they can affect patient outcomes, which means the acceptable margin for error is far smaller.
Modern AI models remain probabilistic systems. Even when they perform well overall, they can still produce inconsistent outputs or confidently generate incorrect information, a phenomenon commonly referred to as hallucination.
The challenge in healthcare AI is therefore not simply improving models. It is designing systems that allow those models to operate safely within environments built around verification, accountability, and professional responsibility.
Any AI deployed in this context must include mechanisms for review, escalation, and validation.
Human-in-the-loop systems work best today
One of the most effective ways to deploy AI safely in healthcare today is through human-in-the-loop system design.
In these systems, AI performs tasks where it has clear advantages: interpreting unstructured information, summarising inputs, structuring data, and surfacing relevant insights. Clinicians then review those outputs and make the final judgement.
Human oversight should not be treated as a fallback when systems fail. It should be designed into the architecture of the system itself.
Both clinicians and AI systems have error rates. Human decision-making can be affected by fatigue, workload, or incomplete information. AI systems, while increasingly capable, can still produce unreliable outputs.
Combining the two can therefore produce more robust outcomes than relying entirely on either. AI processes information at scale, while clinicians apply contextual judgement where decisions carry real consequences.
Where AI can deliver immediate impact
Much of the public discussion around healthcare AI focuses on clinical decision-making. In practice, some of the most immediate improvements may come from addressing operational inefficiencies.
Healthcare systems remain highly fragmented. Patient enquiries arrive as unstructured text, clinical services are described inconsistently across providers, and administrative workflows often rely on manual coordination.
Much of this complexity comes from organising and navigating fragmented information across patients, providers, and systems – exactly the type of problem modern AI systems are well suited to address.
In aesthetic medicine, these challenges appear early in the patient journey. Patients rarely arrive requesting a specific treatment. Instead, they describe a concern, such as improving skin quality or addressing lines, leaving clinics to interpret intent and determine the appropriate next step.
In systems we built at MARBL for aesthetic clinics, where we structured and normalised more than 100,000 individual services across providers, models were generally capable of interpreting these unstructured patient requests and narrowing down appropriate treatment options.
Rather than replacing the consultation, the system acted as a form of prequalification. It could interpret the patient’s concern, capture relevant context, explain appropriate treatment options, and narrow the range of suitable paths before the patient reached the clinic.
Clinicians still made the final decision on treatment, but the intake, fact gathering and patient education that often consume much of the consultation process could already be completed.
Interestingly, when workflows failed, the cause was rarely AI misunderstanding. In many cases, the system reached the correct decision point, such as initiating a booking, but the process broke down during execution.
Completing the action required interacting with the clinic’s practice management software. While the AI could determine what should happen next, implementing that action reliably through existing clinic software proved far less predictable.
In other words, the system often knew what should happen next. The difficulty was making it happen within the constraints of existing healthcare infrastructure.
The balance between humans and AI may evolve
As models improve, the balance between automation and human oversight will likely evolve.
Today, the safest way to deploy AI in healthcare is through systems that combine machine capability with human review. AI can interpret information, structure complex inputs, and surface insights at speed, while clinicians apply judgement and accountability where decisions affect patient outcomes.
This reflects the realities of healthcare environments, where the cost of error is high, and current AI systems remain probabilistic.
As models become more reliable and evaluation methods improve, AI systems may take on greater levels of autonomy in certain parts of care.
For now, the most effective healthcare AI systems are those designed as structured collaborations between machine capability and human expertise.
The real constraint in healthcare AI is no longer model capability. It is the systems that those models must operate within.
Shirin Krall is a healthcare technology founder working at the intersection of artificial intelligence and healthcare infrastructure. She previously held senior leadership roles at Revolut during a period of hypergrowth and now focuses on how AI systems can improve safety, transparency and operational efficiency in healthcare delivery.
