When most developers think about healthcare AI, they think about models.
Which model should we use?
Should we use a managed API or self-host?
How large does the context window need to be?
What's the cost per token?
After spending the last several months building healthcare AI software, I've come to a different conclusion:
The model is often the easiest part of the system.
The Real Challenge
Building a proof of concept with an LLM has never been easier.
You can sign up for an API, send a prompt, and have something impressive running in an afternoon.
The hard part begins when you try to turn that proof of concept into something people can actually use.
Questions start appearing immediately:
- How does this fit into an existing workflow?
- What happens when the model is wrong?
- How should users understand the limitations of the system?
- How do you authenticate requests?
- How do you monitor usage?
- How do you manage costs?
- How do you maintain trust?
In healthcare, these questions become even more important.
Information vs Decisions
One lesson I've learned is that users rarely need more information.
They need help making decisions.
A patient may already know:
- Their symptoms
- Their insurance provider
- Their local healthcare options
What they often don't know is:
- Whether telehealth is appropriate
- Whether urgent care is a better option
- Whether they should wait for a primary care appointment
- What the likely cost implications are
The challenge isn't generating more text.
The challenge is helping people navigate uncertainty.
Building for Reality
Many AI demos assume a perfect environment.
Real users don't operate in perfect environments.
They are busy.
They are distracted.
They may not understand healthcare terminology.
They may not trust AI recommendations.
That means AI systems have to do more than generate answers.
They need to provide context, communicate uncertainty, and fit naturally into existing decision-making processes.
What I'm Watching
I think the next wave of healthcare AI products will be less focused on model capabilities and more focused on operationalizing AI within real-world workflows.
The winners won't necessarily be the teams with access to the largest models.
They'll be the teams that understand users, workflows, and trust.
The technology is important.
The experience around the technology is what ultimately creates value.
What lessons have you learned while moving AI projects from proof of concept to production?
Top comments (0)