AI Behind the Curtain

Healthcare is one of those industries where the gap between what technology can do and what workers actually experience day-to-day is almost embarrassingly wide. I work on a healthcare product, and I spend a lot of time watching people perform tasks that feel like they were designed to exhaust them — calling and re-calling patients, juggling calendars, doing data entry at the end of a long shift. It’s not that the tools are broken. It’s that they weren’t designed with the whole human in mind.

So I’ve been thinking: what if AI didn’t show up as a chatbot you have to babysit? What if it just… handled things?

Over the last couple of years, LLMs have gotten genuinely capable. Vision models can read and interpret images with surprising accuracy. Voice models are good enough to be useful in real workflows. And agents can work together to knock out multi-step tasks without constant hand-holding. That changes what’s possible — not in a theoretical, TED talk kind of way, but in a “we could actually ship this” kind of way.

The question I keep coming back to isn’t “can AI do this?” It’s “where is the friction, and can AI absorb it?”

Thinking out loud

Here are a few places where AI working quietly in the background could genuinely make healthcare workers’ lives better:

• Generating patient forms and questionnaires that get printed and distributed — no more copy-pasting templates
• Drafting communications between sites and patients based on visit context
• Coordinating across multiple calendars to find available appointment windows without the back-and-forth
• Surfacing health characteristics for a given patient population so coordinators can be proactive instead of reactive
• Using vision models to ingest paper forms — a clinician snaps a photo of a completed form, and the system extracts and persists the relevant information automatically

That last one is my favorite. Paper forms aren’t going anywhere in clinical settings. But if snapping a photo is the only thing a coordinator has to do, you’ve just saved them 10 minutes per patient — multiplied across a site’s entire census. That adds up fast.

None of this requires a robot or a total system overhaul. It requires good design and AI doing the boring parts so humans can focus on the parts that actually matter.