Steer Health is the second healthcare company I've built, and by now our AI has handled millions of patient conversations across 63+ hospitals and hundreds of care sites covering 21 million patient lives. Healthcare AI is still mostly uncharted territory, but we have enough scar tissue to know which of our early beliefs survived contact with reality and which didn't. These are the ten that survived.
You can see them in our results: a 2.2 percent no-show rate across our network against a national average of 18 to 20 percent, and 2.6 million patients actively using the platform as of this summer. But I want to be clear about the order of operations. Growth is the output. Patient access is the reason. Every lesson below exists because somewhere, a patient needed care and the system made it hard to get.
1. One channel is never enough
Everyone is building voice AI right now, and I understand why. The phone is still the front door of American healthcare. But after millions of interactions, our data is blunt: patients don't live in one channel. The same person who calls at 7 a.m. will text at lunch, click a web link at 9 p.m., and show up next week through a half-legible faxed referral. Nationally, around 40 percent of referrals are abandoned, and much of that leakage happens right at the seams, where a patient crosses from one channel to another and nobody owns the handoff.
So we built FastTrackCare as an omnichannel platform from the first line of code: web, chat, SMS, fax, and voice through Luna, our embedded voice agent, all feeding one patient journey. Betting everything on a single channel is really a bet that patients will meet you where you're easiest to build. They won't.
2. Your operating playbook is the most powerful prior
The other big debate: can a general-purpose model run healthcare operations out of the box? Try booking one orthopedic appointment and you'll have your answer. Which surgeon takes workers' comp? Who insists on seeing imaging first? Which location actually has Saturday hours this month? Does this payer require a referral on file before you can even offer a slot? None of that lives in a foundation model. All of it lives inside the organization, usually in the heads of your three most tenured schedulers.
We treat each health system's playbook the way a good driver treats a map: as a prior. The rules get codified once in Steer Studio, our AI reasoning layer, and inherited by every channel and every agent. That frees the real-time intelligence to focus on what's genuinely dynamic: the patient in front of it, their intent, their coverage, their urgency. AI without your playbook isn't an assistant. It's a liability with a pleasant voice.
3. One brain is better than ten bots
I've sat in health system conference rooms where the team listed eleven different AI vendors on one whiteboard. A bot for scheduling. Another for reminders. Another for intake, another for claim denial processing. Every one of them demoed beautifully. Together, they were operational spaghetti: separate rule sets, separate versions of the truth, and a patient repeating her story at every seam.
We consolidated in the opposite direction. One reasoning layer, many hands. Define a scheduling rule once and the phone line, the website, and the referral workflow all update together. When the platform learns something, everything gets smarter at the same time.
4. You can't put a black box in front of a patient
AI is powerful, but it isn't magic, and a hospital's main line is a terrible place to find that out. Pure end-to-end generation, where a model improvises its way through a patient interaction, is a trust risk no health system should accept. Our architecture is agentic with guardrails: the AI works like your best employee. It follows the practice's playbook, and it escalates what it shouldn't handle. The hard constraints sit above the model, not inside it, where they can't be talked out of.
The same principle governs the clinical side. SteerNotes, our Agentic Clinical Intelligence, doesn't just hand you a note. It shows its work: suggested codes with the supporting evidence, draft orders queued for physician review and approval. The AI proposes. Clinicians decide. Any system that can't explain itself has no business near a chart.
5. The EHR is the system of record. AI must be the system of action.
Healthcare software spent a decade creating work about the work. Another portal. Another inbox. Another screen for staff to swivel between. So we hold ourselves to a simple rule: if the AI's output is a task a human has to re-key, the AI has failed.
Integration depth isn't a feature for us, it's the product. We integrate with 36 EHRs through FHIR R4 and HL7, including Epic with single sign-on, because the value is in the write-back: the appointment lands on the actual schedule and the order lands in the actual chart. The EHR stays the system of record. AI earns its keep by becoming the system of action on top of it.
6. Every agent needs a critic
An AI that grades its own homework will always pass. So evaluation is built into the platform, not bolted on: automated review of conversations at scale, measuring not just whether the call was answered but whether the intent was resolved, the booking was right, the escalation was graceful, and the patient left better off than they arrived.
That constant critique catches drift before patients feel it, and it lets our human experts spend their judgment on the hardest one-in-a-thousand cases instead of sampling transcripts at random. Your evaluation system is your conscience. Much of the industry hasn't built one.
7. Healthcare AI must think fast and slow
A patient on the phone won't wait three seconds while a model deliberates. Real-time conversation demands instant responses, natural interruption handling, and immediate action against the rules. Coding an encounter or reasoning through a complex referral rewards exactly the opposite: slow, deep analysis across the note, the history, and the payer's requirements.
The mistake is forcing one architecture to do both jobs. We run two speeds in one system. A fast path lives inside the conversation's latency budget. A slow path does the heavyweight reasoning around the encounter, where a few extra seconds buy a lot of rigor.
8. In healthcare, governance isn't overhead. It's the product.
Most teams think the work ends when the agent works. In healthcare, that's where the work starts. For us it means validation before anything reaches production, monitoring for drift after it does, defined escalation paths, clinician review for anything touching the chart, and honest disclosure to patients about what they're talking to. HIPAA is the floor, not the ceiling.
None of that is bureaucracy. It's the reason a hospital can put AI on its main line and sleep at night. The uncomfortable truth about our industry: an AI system is only as good as the governance evaluating it, and much of the market shipped the former without ever building the latter.
9. The flywheel is real, and it compounds
Every conversation makes the next one better. Interactions surface edge cases. Edge cases refine the playbook in Steer Studio. The refined playbook improves every channel at once, and the results come back as data. Run that loop across millions of patient interactions and a no-show rate falls to 2.2 percent and stays there. It's also why platform advantages compound while point-solution advantages plateau: the flywheel needs the whole patient journey to spin.
10. There is no substitute for production
If I could get one sentence tattooed on this industry, it's this one. A great demo is base camp, not the summit. An agent performing in a scripted pilot, with humans quietly cleaning up behind it, hasn't been tested. It matures only when it owns the outcome: the 2:47 a.m. call with nobody else on the line, the fax that arrives half-legible, the patient who changes her mind twice in one sentence.
Production exposes the true weight of every decision and surfaces situations no simulation would think to invent. It's also why we're willing to tie our pricing to performance. When your AI genuinely owns outcomes, you can put your revenue behind your roadmap. Vendors who can't make that offer are telling you something.
The road ahead
Healthcare is facing a labor shortage it can't hire its way out of, while an estimated $150 billion evaporates every year through missed appointments and broken access. Against that backdrop, here are three predictions I'm comfortable being graded on:
First, within three years, AI governance committees will be as standard in health systems as infection control. Vendors should be preparing for that scrutiny now, not lobbying against it.
Second, the point-solution era ends in consolidation. Most of today's healthcare AI companies won't exist as standalone vendors by 2030. Health systems are already exhausted by the eleven-logo whiteboard.
Third, performance-based pricing becomes the filter. When AI owns outcomes, outcomes become the contract, and vendors unwilling to be measured will quietly disappear from RFPs.
These ten lessons are why Steer Health exists as an AI-Native Growth Platform for Health Systems. They're behind the recognition I'm proudest of, including the 2026 Hearst Health Prize and the KLAS Emerging Solutions Top 20. But mostly they're a working philosophy, earned one patient conversation at a time, including one at 2:47 in the morning that ended with “Wait, that's it?”
That's the whole ambition, honestly. A healthcare system or a large medical group where getting care feels that easy, every time. The front door is open.