A clinician using a tool badly affects one clinic. A board approving the wrong platform commits the institution for years.
Healthcare AI training is typically aimed at clinicians, who the AMA's 2026 survey shows are already requesting it, with 92% wanting more education. The group less often served is the executive and board layer approving procurement and accepting risk. Under Singapore's AIHGle 2.0 the deployer organisation carries responsibilities that cannot be transferred to a vendor, and those responsibilities are discharged by people at that level.
Where AI training budget goes in healthcare, it usually goes to the frontline. That is understandable. It is visible, it is popular, and the demand is genuine: the AMA's 2026 survey of 1,692 physicians found 92% wanting more education and training in AI.
But consider the blast radius of a bad decision at each level. A clinician who over-trusts an AI scribe produces notes that need correcting. A board that approves a five-year platform commitment on a weak evaluation shapes the institution's capability, data position and clinical workflow for the length of that contract, and usually beyond it.
Under AIHGle 2.0, the deployer organisation carries accountability that cannot be handed to the developer. In the AI-SaMD sandbox arrangement, HSA went further and required endorsement from the medical board chair or CEO before deployment, alongside consultant-level oversight of design and validation.
That is a regulator stating plainly that senior leaders are expected to have a view. Endorsement is not a signature on someone else's assessment.
The objection we hear is that boards cannot be expected to develop technical depth, and that this is what management and advisers are for. We would push back on the framing. Boards are routinely expected to exercise judgment on financial instruments, clinical quality and legal risk without being accountants, clinicians or lawyers. The standard is not expertise, it is the ability to interrogate the expert in front of them.
Nobody argues a board needs to model a swap to ask whether the hedging strategy is sound. The same standard applies here and is not being applied.
Our practical advice: run one session where the board evaluates a real vendor validation study from a live procurement, with no technical preamble. It exposes the gap immediately, it is more useful than an overview of AI, and it changes what the next paper submitted to them looks like, because management knows the questions are coming.