The first time I saw a company brag about "AI automation," it turned out to be a spreadsheet with a nicer haircut. A few weeks into the job, I learned the "model" was mostly column mapping, data formatting, and a support team quietly doing the real work off to the side. I got the classic line: "Come to our office and watch us work." That's not artificial intelligence. That's Excel with a witness protection program.
Healthcare is full of that same temptation. Big promises. Small reality. But the interesting part is: real AI in medicine isn't mainly about replacing doctors. It's about replacing the junk work around doctors. And that's where the upgrades actually land.
Hospitals and clinics already have plenty of smart people. What they don't have is time. Or clean data. Or a workflow that doesn't feel like it was designed by a committee that hates joy.
The boring parts are the expensive parts
If you've ever been in a clinic waiting room, you've already seen the problem. Intake forms. Repeated questions. Insurance details. Medication lists that get typed three different times by three different people.
Doctors aren't slow. The system is slow.
Here's a normal day in a normal clinic:
- The patient fills out an online form at home.
- The same patient gets asked the same questions again at check-in because the form didn't sync.
- A nurse retypes the medication list because the patient wrote "blood pressure med, little white pill."
- The doctor opens the chart and spends the first 6 minutes figuring out what happened last time.
- After the visit, someone codes the visit, submits prior auth, and sends follow-up instructions.
- Then the patient messages: "Wait, which dose do I take?" and the staff repeats the same answers again.
None of that improves care. It's just plumbing. And plumbing is expensive when it's done manually.
AI is a good fit for:
- Patient intake (turning messy symptoms into structured notes)
- Chart summarization (what happened last time, what meds changed, what tests mattered)
- Back-office workflows (coding, billing, prior auth, scheduling)
- Follow-up messages (the "here's what to do next" part that staff retypes all day)
This is the stuff that burns out teams. It's also the stuff that makes care worse because it steals time from the actual patient.
A useful mental test: if the task is mostly reading, sorting, reformatting, and copying, AI can probably help. If the task is mostly judgment under uncertainty, you still want a clinician in the loop.
"Diagnostics AI" is the flashy demo. Intake AI is the paycheck.
Diagnostics gets headlines because it's dramatic. "AI detects cancer better than humans" is a great pitch deck slide.
But if you want real value fast, look at the front desk and the nurse station.
Intake is where clinics lose time and accuracy. Patients don't speak ICD-10. They speak human. They say stuff like:
- "My stomach is off."
- "I feel weird after I eat."
- "My heart is doing that thing again."
- "I can't sleep and I'm tired all day."
That's not useless. It's just unstructured. A good intake system can turn that into a clinician-friendly starting point.
Imagine a patient types: "I get dizzy after standing, heart races, worse after hot showers, started after COVID." A decent system can ask follow-ups, normalize the timeline, and hand the clinician something like:
- onset: 6 months ago
- triggers: standing, heat
- associated: tachycardia, fatigue
- relevant history: post-viral
- suggested next questions: hydration, salt intake, BP/HR changes, meds
Not a diagnosis. A clean starting point.
That's an upgrade.
And it can be very practical. You can build intake AI as a guided chat that does three simple things:
- Collect: Ask targeted follow-ups based on what the patient says.
- Normalize: Convert "a while ago" into "~6 months," and "heart races" into "palpitations/tachycardia (patient-reported)."
- Summarize: Produce a structured note and a short narrative the clinician can scan.
The tradeoff is obvious: if you ask too many questions, patients abandon the form. If you ask too few, clinicians don't trust it. The sweet spot is usually "short first pass, then adaptive follow-ups."
A good pattern is:
- Start with 5–7 core questions (chief complaint, duration, severity, meds, allergies, red flags).
- Then ask up to 5 adaptive follow-ups based on the complaint.
- Stop early if the patient is struggling or on mobile with low patience.
Also: don't pretend it's a doctor. Label it as an intake assistant. People are fine with that. They just want the visit to be smoother.
Chart summarization: the hidden time sink
If you want to see clinician pain, watch someone scroll through a chart with 10 years of notes. Half of it is copied forward. A quarter of it is irrelevant. The rest is buried in PDFs.
Summarization sounds simple until you try it. A good summary has to answer:
- What changed since last visit?
- What are the active problems and current meds?
- What tests were done, and what mattered?
- What's the plan, and what's pending?
A bad summary is worse than no summary because it creates false confidence.
So the implementation detail that matters is grounding. The model should cite sources in the chart. Not in an academic way, but in a "this came from the cardiology note dated X" way.
A workable approach looks like:
- Pull the last N relevant notes (not everything).
- Extract meds, problems, labs, imaging, and plan sections using rules + model.
- Generate a summary with links back to the source text.
- Show "uncertain" flags when the model is guessing (ex: conflicting medication doses in different notes).
Clinicians don't need poetry. They need a reliable index.
Back-office workflows: where money leaks out
Everyone talks about "saving doctors time," which is true. But the other big lever is back-office work that delays care and burns cash.
Prior auth is the classic villain. It's not medically hard. It's administratively hard.
A prior auth packet often needs:
- the diagnosis code
- prior treatments tried and failed
- relevant imaging or lab results
- a rationale statement
- the right form for the right payer
- follow-up when the payer asks for more info (which they will)
Most of that is document assembly. AI can help by:
- pulling the right evidence from the chart
- drafting the rationale in plain language
- filling payer-specific forms
- flagging missing pieces before submission
The tradeoff: you must keep humans in the loop, because a wrong code or missing contraindication can cause denial or, worse, a compliance problem. Think "draft + checklist," not "auto-submit and pray."
Billing and coding is similar. AI can suggest codes based on the note, but it should also show why it suggested them, and what documentation is missing to support them. Otherwise you get upcoding risk, audits, and a mess.
Scheduling is another quiet win. Clinics lose appointments because of simple friction: phone tag, unclear instructions, patients not knowing where to go. AI can handle:
- confirming appointment details
- sending prep instructions
- collecting pre-visit forms
- rescheduling with guardrails (don't book a 15-minute slot for a complex follow-up)
Again, not magic. Just fewer dropped balls.
Follow-up messages: the endless loop
Patient portals are great until you realize they create a second job for staff: answering the same questions forever.
A safe and useful AI system can draft replies like:
- "Here's how to take the medication we discussed."
- "Here are the red flags that mean you should go to urgent care."
- "Here's how to schedule the lab test and when results usually come back."
But the key word is draft. You want a clinician or nurse to approve, especially for anything that could change a plan.
Implementation detail: build a library of clinic-approved templates, and let the model fill in the blanks from the chart. That keeps tone consistent and reduces hallucinations. It also makes the legal team less twitchy, which is its own kind of medical necessity.
Small language models are the quiet win here
I'm still obsessed with small language models. I want these things shrunk down enough to live on a phone, or at least run close to where the data lives. Healthcare data is sensitive, and shipping everything to a giant cloud model is not always the smart move.
A practical setup looks more like:
- a small on-device model for dictation cleanup, basic summarization, and redaction
- a clinic-side box (think Jetson GPU plus small servers) for heavier lifting
- optional cloud calls for the rare "big brain" tasks
This isn't just a privacy thing. It's also a reliability thing.
If your workflow depends on a cloud call, you inherit:
- network outages
- latency spikes
- vendor rate limits
- surprise model changes
- unpredictable costs
Local models give you stable behavior and predictable costs. They also make it easier to prove where data went, which matters in healthcare.
The tradeoff is capability. Small models can be great at narrow tasks (cleanup, extraction, routing). They can struggle with long-context reasoning across a huge chart. That's why hybrid is realistic: local first, cloud only if needed, and only with redaction and consent.
A simple architecture that works:
- Redact locally: remove names, addresses, phone numbers, and anything not needed for the task.
- Route the task: is this extraction, summarization, drafting, or classification?
- Run locally if possible: small model + rules + templates.
- Escalate if necessary: send minimal context to a larger model, log it, and store outputs with citations.
This is the same pattern I keep thinking about for manufacturing too. Smaller AI chips. Smaller robots. More local autonomy. Less "send it to the mothership and wait."
Healthcare has different constraints, but the architecture rhyme is there.
The real risk: fake AI and hidden humans
That old SaaS job taught me something: if a company won't show you how the sausage is made, it's probably because the sausage is made of people.
Medicine can't afford that. If your "AI scribe" is actually a human in another country cleaning up notes at 2 a.m., you have a privacy and compliance nightmare. Also, the latency will be "mysteriously variable," which is a polite way of saying "someone went on lunch break."
There's also a subtler risk: even if humans are "quality checking," you need to know who they are, what they can see, and what gets stored. A clinic can't just shrug and say "the vendor handles it."
So I look for boring proof:
- audit logs (who accessed what, when, and why)
- clear model boundaries (what it can and can't do)
- measurable error rates (not vibes, numbers)
- a real fallback path that isn't "panic and hire contractors"
- data retention rules (what is stored, for how long, and where)
If a vendor can't answer those questions plainly, that's the answer. The warning signs that distinguish real AI products from manual labor disguised as automation apply just as much in healthcare as anywhere else.
Doctors stay. The admin sludge goes.
Replacing doctors is a bad idea and a worse product plan. Upgrading them is realistic.
Give clinicians cleaner notes, better summaries, fewer clicks, and fewer repetitive questions. Let staff focus on patients instead of chasing forms. Do that, and you get higher accuracy because humans have time to think again.
The best version of "AI in healthcare" is boring in the right way:
- The visit starts on time.
- The clinician walks in already oriented.
- The note is mostly done without turning the doctor into a typist.
- The patient leaves with clear instructions.
- The clinic gets paid without 14 rounds of paperwork.
No robot doctor. A doctor with fewer stupid obstacles.
-Sethers