I've been running a thought experiment lately: what if AI could perfectly replicate an EMR overnight? No bugs, no missing features. Just prompt it and go. Would that change the SaaS model for healthcare?
I'm still leaning toward no. Even in that hypothetical world, the economics don't work for most practices.
Building software has always been expensive. AI is changing that. But running it is expensive too, and that cost doesn't shrink just because a model wrote the code. An EMR needs a production database with backups and failover, secure image storage for documents and clinical files, SMS for appointment reminders, fax (yes, still fax) for communicating with insurance companies and referral sources, and AI inference for the features that are starting to matter, like automated note generation. Each of those is a separate service with its own monthly bill. Add HIPAA-compliant hosting, SSL certificates, monitoring, and someone to wake up when the server goes down at 2 a.m., and the costs accumulate fast.
A SaaS company spreads those costs across hundreds or thousands of practices. The per-practice share of a database cluster, a fax gateway, an SMS provider, an inference endpoint is small because the fixed costs get divided. A single practice absorbs all of it alone. The math is the same math that makes it cheaper to rent an apartment in a building than to build your own house on every dimension except equity. And software doesn't build equity.
Now, there is a version of this where it makes sense to build your own. If your company is a dev shop with engineers on staff, servers already running, and AWS credentials already in hand, maybe you do spin up your own project management tool over a weekend and it serves your needs fine. Your marginal cost of hosting one more app is close to zero because the infrastructure already exists and the expertise is already in the building.
But that's a very specific customer profile. A twenty-person speech therapy practice in suburban Ohio does not have a DevOps engineer, a staging environment, or an on-call rotation. The practice owner is a speech-language pathologist who went to graduate school to help children communicate, and she is already managing therapists, billing, scheduling, compliance, and parent communication. Asking her to also manage a cloud deployment is absurd, no matter how easy AI makes the initial build.
The initial build is the part people fixate on. It's visible and dramatic: look, AI wrote the whole thing in a weekend. But software is a living system.
AI probably does reduce switching costs. Migrating data, rebuilding custom workflows, and adapting interfaces may get dramatically easier.
But lower switching costs don't eliminate the operational burden. The infrastructure still has to run. Integrations need maintenance, compliance keeps evolving, and someone has to monitor the system when something breaks.
AI may change how software gets built and how easily it can be replaced. It doesn't change the reality that most healthcare practices want the responsibility for running that software to live somewhere else.