Hana Kobayashi
“'Human in the loop' is easy to write on a roadmap slide. It's hard to design when the human is exhausted at 2 a.m.”
- Martin Knapic, Product MoatHana Kobayashi leads product for clinical tools at Meridian Health in Osaka, where the cost of an overconfident feature is measured in patient outcomes, not churn. Ten years in, she's become one of the most careful, precise voices in this series on what 'human in the loop' actually costs to design properly, versus what it costs to write on a slide.
Identity & Background
- Name
- Hana Kobayashi
- Current role
- Product Lead, Clinical Tools
- Current company
- Meridian Health
- Location
- Osaka, Japan
- Years in product
- 10 years
Your story: how did you end up in product?
I trained as a nurse before I moved into product - three years on a hospital floor before a health-tech startup hired me to help design a tool I'd actually complained about using. I brought the floor experience into every spec after that.
Core PM Philosophy
How do you define product success?
Minutes returned to patient care per clinician per shift, and whether clinicians would notice if the tool disappeared. Being missed is a higher bar than being used.
What's your decision-making philosophy?
I ask a clinician on the team to make the final call on anything that touches the workflow at the bedside. Product decides what to build; clinicians decide whether it's safe to use.
What's your biggest PM lesson or mistake?
We once auto-summarized patient histories in a way that was accurate on average and dangerously wrong on outliers. Average accuracy is close to meaningless in this domain; we now report confidence per-field, not per-document.
Product & Decision-Making
How do you do customer discovery and validation?
Shadowing shifts, not scheduling interviews. Twenty minutes watching a nurse chart between interruptions taught me more than any research session could.
AI & the Future
How is AI changing product management?
The cost of overconfidence is measured in patient outcomes, not churn. Every AI-assisted feature we ship has to degrade to a safe, clearly labeled manual fallback, and we test that fallback path as hard as the happy path.
What's one thing about AI everyone gets wrong?
That 'human in the loop' is a design constraint with real cost, not a checkbox. If reviewing the model's output takes as long as doing the task manually, you haven't actually saved the clinician anything.
Beliefs & Opinions
What's your unpopular product opinion?
Clinician trust decay, specifically - how quickly a clinician stops trusting a tool after just one bad recommendation, even if the tool is right the other ninety-nine times. Most teams measure model accuracy obsessively and barely measure trust recovery, which I think is the harder and more important problem.
Which author, thinker, or influence shapes you?
I let the far more aggressive consumer AI space be the early testing ground and specifically study its failure cases, not just its wins. By the time something's proven safe enough to seriously consider for a clinical setting, someone outside healthcare has usually already found its sharp edges for us.
What makes a great product?
One that's calm. In healthcare, a great product lowers the cognitive load of a tired person making a high-stakes decision.
Company & Product
What does your product solve?
Decision-support software for hospital clinicians - tools that surface relevant patient history, flag potential drug interactions, and now increasingly summarize and highlight what a model thinks is most relevant, always with a clear, labeled fallback to the raw record.
What's the market opportunity?
Our ICP is a mid-sized hospital system too small to get real attention from the giant EHR vendors' product roadmaps, but too complex to run on something generic. Against the incumbents, we win on actually shipping fast in response to a specific hospital's workflow complaints, not a five-year enterprise roadmap.
More from the conversation
What do you do outside of work?
I practice kyudo, Japanese archery, several times a week. It's entirely about form and repeatable precision under calm conditions, which is the exact opposite of a hospital floor, and I think that contrast is why it clears my head so effectively.
Was there a point you nearly left healthcare product work?
After a feature I'd shipped contributed to a near-miss incident - nobody harmed, but close enough to shake me. I considered moving to a lower-stakes industry entirely. I stayed because that incident is exactly why I think careful people need to be in this specific room, not fewer of them.
How big is your team, and the company?
Product for clinical tools is 14 people, inside a company of about 340. I lead the physician-facing surface specifically, as distinct from the separate team building patient-facing tools.
How do you design for an exhausted user making a high-stakes decision?
Minimize what the interface asks the clinician to hold in their head. Surface the model's reasoning inline, in the same glance as the recommendation, not behind a link they're too tired to click at 2 a.m.
Enjoyed this conversation?
A new interview with a product person lands every week. Get it in your inbox.