Ajai Sehgal joined IKS Health in Aug 2025 as its Chief AI Officer.gettyHealthcare doesn’t have an AI adoption problem; it has a trust problem. Millions of patients and clinicians now rely on AI tools daily, but in healthcare, success isn’t measured by helpfulness. It’s measured by accuracy under uncertainty. A confident wrong answer isn’t just inconvenient; it can directly impact patient outcomes and financial integrity.Recent data underscores how quickly behavior is shifting. According to a January 2026 OpenAI report, tens of millions of people now ask ChatGPT health-related questions every day, with roughly seven in 10 of those conversations happening outside normal clinic hours and notably heavy use in rural and underserved “hospital deserts.” AI is already woven into the healthcare journey, whether the system is ready for it or not. The question isn’t whether it gets used—it’s how responsibly and reliably we deploy it.That is the gap. General-purpose AI is built to be useful, and it’s very good at it. But in a clinical or operational setting, usefulness isn’t enough. An answer has to be verifiably correct, and the system has to know when it isn’t sure. Without domain-specific grounding, these models struggle with the things that actually matter in healthcare: the nuances of an individual patient, guidelines that keep changing and reimbursement rules that are complex and change frequently.What A 'Glass-Box' Approach Really MeansThis is where a glass-box approach earns its place. A black-box system hands you an output with no way to see how it got there. A glass-box system is built to be transparent, traceable and auditable. It doesn’t just give you an answer; it shows you how it got there, what data it leaned on and where it’s still uncertain. In healthcare, that visibility isn’t a “nice-to-have.” It’s a requirement.We’re already seeing what happens when you get this wrong. Low-quality, unverified output, the stuff that’s come to be called “AI slop,” leads to coding errors, denied claims, compliance exposure and ultimately compromised patient care. In an industry where margins are thin and regulators are watching closely, none of this is theoretical. The consequences are operational and financial, and they show up fast.Why A Small Language Model Makes Sense In HealthcareThe answer is to move toward domain-specific AI architectures built for control, context and compliance. Instead of leaning on a single large, general-purpose model, these systems bring in smaller, specialized models trained on curated clinical and operational data. The payoff is huge with better auditability, lower latency, tighter data governance and far more predictable performance where it’s needed.It starts with data. In most hospital systems, the information you need is scattered across heavily siloed legacy systems, locked inside an EHR that treats your own data as something you need permission to touch, with much of it still sitting on-prem or even on someone’s thumb drive. No model, however capable, can reason well on top of that. The boring but essential first move is to liberate and normalize that data: Pull it out of the silos, give it full clinical context and land it somewhere governed and accessible at scale, mapped to standards like FHIR R4 and HL7. Get this right, and everything downstream gets much easier. Get it wrong, and no amount of model sophistication will save you.From there, a handful of things matter. Ground the models in real clinical knowledge. It's important to match the model to the job. A large general-purpose model is fine for open-ended questions, but for high-volume, high-risk work, a smaller specialized model is easier to audit, cheaper to run and more predictable. Keep clinicians in the loop wherever the risk is high, and let the system route its own low-confidence cases to a human automatically. Build in traceability from day one: Determine what data informed an answer and where the model wasn’t sure, so the audit trail is simply there. Pick one well-bounded, high-value workflow; prove it, and expand from there. That’s how trust gets built, one defensible result at a time.Inside A Layered, Transparent AI StackIn practice, this is a layered stack with deep integration into the EHR, with access to structured and unstructured clinical data that still needs to be normalized and curated alongside everything else sitting in silos. Above that sits an orchestration layer that manages workflows and how the models interact, then a trust and compliance layer that enforces the regulatory and governance guardrails. At the top are the task-specific applications people actually use.What really sets this architecture apart is a clinical knowledge layer with biomedical ontologies and interconnected knowledge graphs drawn from peer-reviewed research. This is the grounding mechanism. It lets the AI map what it’s given to validated medical concepts and the established relationships between them. A central reasoning engine can then work over that structured knowledge to produce outputs that are explainable and aligned with clinical and operational standards.​Glass-Box AI In PracticeTake medical coding. In a conventional AI setup, the model infers codes straight from the clinical notes but sometimes misses the subtle detail that changes everything. A glass-box system works differently. It maps the documentation to standardized ontologies, validates those relationships against known medical knowledge and, when its confidence drops, flags the ambiguity and routes the case to a human reviewer. The result is higher accuracy, fewer denials and a clean audit trail.The same approach carries straight into prior authorization, denial prevention and clinical decision support—places where transparency and traceability aren’t optional (for regulators and providers alike). Ground the AI in domain-specific knowledge and make its reasoning visible, and you move from experimental pilots to something you can actually scale and trust.The Real Measure Of AI SuccessAs healthcare keeps evolving, the winners won’t be the organizations that adopt AI the fastest. They’ll be the ones that adopt it most responsibly. Glass-box AI is how you get there—a path where innovation and accountability aren’t in tension but pulling in the same direction.​Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?