Reframing AI Support in Oncology (Hint: It's Not Decision-Making)
· Gena Health Team
The major AI labs have made significant moves into healthcare over the past few months. OpenAI launched ChatGPT Health and ChatGPT for Healthcare, giving patients the ability to connect medical records and wellness apps and offering healthcare enterprises access to GPT-5 models built for clinical workflows. Amazon, Microsoft, and Perplexity followed with their own health-focused LLMs shortly after, and Anthropic announced Claude for Healthcare, with HIPAA-ready infrastructure and connectors to databases including CMS, ICD-10, and PubMed.
The clinical adoption tells a similar story: an American Medical Association survey published in February 2025 found that 66% of physicians reported using some form of AI in practice during the year prior, up from 38% in 2023. OpenAI reports that over 230 million people globally ask health and wellness questions on ChatGPT every week. The adoption curve from both patients and physicians is steep, and the investment has been substantial.
For oncology, AI has already started transforming the way care is delivered, but it also raises real questions. Cancer care involves high-stakes decisions, complex patient histories, and treatments that require careful weighing of survival against quality of life. While the benefit of AI in oncology is undeniable, the stakes of getting AI wrong here are different from almost any other clinical context, so skepticism and concerns are legitimate. But the more important question isn't whether AI belongs in oncology - it's "where?" and "when?"
Reframing the Question
Much of the public conversation about AI in healthcare centers on whether AI can make clinical decisions, or whether it might eventually replace physician judgement. This captures both the enthusiasm and skepticism of technological developments in healthcare. While the opportunity to positively transform healthcare is clear, the “black box” nature of many deep learning models often raises skepticism and slows down clinician and patient trust. When AI tools provide a diagnosis or recommendation, the reasoning behind a recommendation isn’t always transparent. Surveys show that most oncologists feel unprepared to evaluate AI tools, and in oncology, a field where treatment rationale must be communicated to patients and accountability rests with the physician, these concerns deserve serious consideration.
But this framing often also obscures where AI can actually already help. When we speak with oncologists about their daily work, the pain points they describe are rarely about the clinical decision itself. Experienced clinicians already know how to evaluate treatment options, weigh risks against benefits, and navigate difficult conversations with patients. That's their training and what they do best. Where they could use real support is with everything that happens before the decision.
The Bottleneck Isn't Reasoning
For patients with complex or refractory cases, particularly those who have exhausted standard guideline options, identifying what's available is incredibly time and labor-intensive: it involves confirming that guideline options are truly exhausted, then searching for clinical trials, first within the institution, then externally. It also means evaluating eligibility criteria against a patient's specific clinical history, consulting colleagues or professional networks (sometimes through WhatsApp groups of fellow oncologists as a last resort), and tracking down information about specialist centres or emerging treatments that may not be well-documented.
To compound this already massive lift, this information is scattered across multiple systems. Trial registries update inconsistently and off-label evidence lives in literature that takes time to search and synthesise. Institutional knowledge often exists only in informal networks. The information is out there, but pulling it together for a single patient can take hours. Oncologists we've spoken with describe spending upwards of two hours per complex case on this research phase alone. That's two hours of gathering and organising information before the clinical conversation can even begin.
What Clinicians Are Asking For
Survey data reinforces this picture. Oncologists overwhelmingly reported that AI tools need to be explainable by clinicians before they can be used in practice (over 80% for both prognostic and clinical decision models). The message is clear: clinicians want to remain in control of the decision-making process.
In practice, what oncologists describe wanting is more specific than better AI. The research burden is the most immediate problem, not just the time it takes but also the cognitive weight of fearing you haven't exhausted every option because you haven't found it. Beyond just getting more sources, clinicians want sources they don't have to second-guess. And when they've identified relevant options, they need a way to bring the patient into this discussion properly.
The Infrastructure Opportunity
This reframes the AI opportunity in oncology. The question isn't whether we can build models intelligent enough to match a physician's judgment. It's whether we can build systems that integrate into how physicians actually work. This goes beyond having the most powerful LLM model because it requires a smart layer that pull together trials, off-label evidence, specialist centres, and emerging therapies into something a clinician can actually query against, with the context of a real patient's profile. Diagnosis, biomarkers, treatment history, location - this all has to be grounded in clinical context, not keywords. From there, the coordination and onward actionability follows naturally: reaching trial sites, tracking responses, knowing what still needs attention.
Although these tools won't make the clinical decisions, trust and information transparency remains a critical component. The results should be easily auditable: where they came from and why it was surfaced. This transparency is table stakes and it's the condition under which a clinician can actually use the tool.
Where Gena is Focused
At Gena, this is the layer we're building. We're not building AI that makes clinical decisions, instead we are building workflow infrastructure that supports the clinicians who make them.
Our platform is designed to support oncologists in the research and coordination phase of treatment exploration: surfacing relevant options faster, reducing time spent on fragmented searches, making the results actionable by streamlining tedious but necessary workflows, and enabling structured collaboration with patients.
We're currently in beta with oncologists across the UK and US, developing the product based on direct input from the clinicians who will use it. The goal is to build something that fits into how they actually work, and that earns trust through utility rather than claiming to replace expertise.
The AI-in-healthcare conversation will continue to evolve with new models, weekly announcements, and constantly-evolving claims about what AI can do for clinical decision-making. Our view is that the most consequential applications won't come from that direction - at least not yet. The biggest value add now is making decisions less burdensome, more reliable, and better connected to the patient sitting in the room.
If it resonates with how you work or you're interested in trialling the platform, we'd like to hear from you.