AI Future Fridays: Medicine Gets a Copilot, and So Do You

Welcome to AI Future Fridays. Every week this column takes one force shaping the near future, tells you the truth about it, and hands you something to build before Monday. This week's force: medicine finally got a copilot, and for once the copilot landed on your side of the desk first.

That's the part nobody put on a conference slide. The most capable diagnostic reasoning ever assembled arrived in a chat window before it arrived in an exam room, and the raw material it needs is a file you already own. Your labs, your imaging reports, your visit notes. They're sitting in a portal you have the password to, written in a language nobody taught you.

The gatekeeper stopped being the technology. It's the reading.

The Patient Advocacy Signal Hiding in Your Own Chart

The signal isn't a hospital press release. It's what people are quietly doing with the portal login they already have. Picture Ravi, a fleet mechanic in Akron, sitting in his truck at 6:40 in the morning with an after visit summary open on his phone. Two lines in it mean nothing to him: "mild hepatic steatosis noted incidentally" and a ferritin value with an asterisk. His follow up is in nine days. Last time he had four minutes of face time and thought of his real question in the parking lot afterward.

So he does what a growing number of people do now. He drops the report into an assistant, asks it to turn the dense document into plain language, and asks a second thing that matters more: what should I be asking about this? Nine days later he walks in with six questions on one page. The visit is still fifteen minutes. It's a completely different fifteen minutes.

Nobody diagnosed anything. A mechanic read his own file.

The Forecast for AI Diagnosis Over the Next 18 Months

Here's the honest read: the machine reasoning is further along than the machine deployment, and that gap is where you live for the next year and a half.

Start with capability. Microsoft's research team built a system called MAI-DxO that works through cases the way a real clinician does, ordering tests and revising as evidence arrives. Run against 304 of the hardest published case records, Microsoft AI reported 85.5 percent accuracy where a panel of 21 experienced physicians working without references or colleagues averaged 20 percent. Those cases are deliberately brutal and the comparison is a benchmark, not a clinic. Still, the direction is not subtle.

Now look at what's approved. The FDA's list of authorized AI-enabled devices has climbed past 1,500 entries, and The Imaging Wire's breakdown puts 1,104 of them in radiology, roughly 76 percent of the total. Nearly all of that cleared, regulated, hospital-grade AI is doing one job: looking at pictures faster than a human can. Useful work. It just never reaches you. It reaches your radiologist.

Drug discovery is moving on its own clock. Insilico Medicine's rentosertib, a compound whose target and molecule both came out of generative models, posted phase 2a lung function results in Nature Medicine showing a mean gain of 98.4 mL against a 20.3 mL decline on placebo, and the company began a phase III trial in July. That's a decade-long pipeline compressed, not skipped. It's also a preview of the shift from models that generate to systems that do the job.

Capability arrives on a two year curve. Institutional adoption arrives on a ten year one. You get to live in the gap, and the gap is where a patient with a good agent has more leverage than they've had in a century.

The Backstory of Open Notes and Patient Record Access

None of this would be possible if a stubborn group of doctors hadn't spent fifteen years arguing that patients should be allowed to read what was written about them.

For most of modern medicine, the chart belonged to the institution. You could request records, pay a per-page fee, and wait weeks for a photocopy. The OpenNotes movement started in 2010 as a study asking a heretical question: what happens if patients can just read their doctors' notes? The predicted disasters, panic and confusion and flooded phone lines, mostly didn't happen. Then came the 21st Century Cures Act, whose information blocking rule made open notes federal policy in the United States in April 2021.

• 2010: the first OpenNotes study puts clinical notes in front of patients and measures what breaks. Very little does.
• 2016: the 21st Century Cures Act passes, setting up rules against blocking a patient's access to their own data.
• 2021: the information blocking rule takes effect, and notes, labs, and imaging reports start landing in portals without delay.
• 2024: federal survey data shows 77 percent of adults were offered online access to their records, and app-based access climbed from 38 percent in 2020 to 57 percent, per the ASTP health IT data brief.

The law handed you a locked room full of documents written for somebody else. The reading tool showed up five years later, and it runs on the same unglamorous machinery behind ordinary document workflows: point it at a dense file, get back something a person can use.

The Access Gap Between Flagship Hospitals and Everyone Else

Now the honest part. The best AI-assisted medicine in the country is concentrated in a handful of academic centers with research budgets, and that is not where most people get care.

The cost of the gap is measurable. A Johns Hopkins team estimated in BMJ Quality & Safety that 795,000 Americans die or are permanently disabled by diagnostic error every year, and that just five conditions account for 38.7 percent of the serious harm. Stroke. Sepsis. Pneumonia. Blood clots. Lung cancer. These aren't exotic. They're the things a rushed visit misses when the patient can't describe the pattern well enough to trigger the right thought.

Three constraints are worth naming plainly, because your build has to work around all three:

• The models are confidently wrong sometimes. A language model will produce a fluent, reasonable, incorrect paragraph without any change in tone. This is exactly why the output you want is questions, not conclusions.
• Your data is worth money to people who aren't you. Health records are the most sensitive file you own. Where the reading happens matters as much as how good the reading is.
• The clinician is not the obstacle. Most doctors want a prepared patient. What they don't have is time to reconstruct your history from scratch at every visit.

So the target was never replacing anyone. Preparation is the whole play, and it happens to be the highest leverage move a patient can make. That's the same principle behind every agent worth building in the workflow directory: narrow the job until it's unambiguous, then run it every single time.

The Countermove: Build a Patient Advocacy Agent

The street finds its own uses for things. The same models that hospitals are cautiously piloting will read a PDF you paste into them at eleven at night, and that asymmetry is yours to use.

Get the boring foundation right first, because the fancy part fails without it.

• Export everything. Pull labs, notes, imaging reports, and medication lists from every portal you have. One folder, named by date.
• Strip your identifiers. Remove name, address, member ID, and date of birth before anything leaves your machine. The clinical content is what matters.
• Turn off training. Whatever tool you use, find the setting that stops your inputs from becoming training data, and switch it off.
• Keep a running timeline. One file, one line per visit. This is the thing no clinic has and every specialist wishes you did.

Then build the agent. It takes your latest export and produces exactly one page: a plain language translation of what the results say, a diff against your previous visit, every value outside its reference range with the range printed beside it, and five to seven ranked questions. Ranked by how much the answer would change what you do next, not by how interesting they are. Instruct it explicitly to refuse diagnosis and to phrase uncertainty as a question for your clinician. That constraint is a feature. It keeps the output in the one register where a language model is reliably excellent, which is turning jargon into questions. If you've written a workflow spec before, this is that same shape with higher stakes and a much shorter output.

Set it to run whenever a new file lands in the folder. Then forget about it until the day before an appointment, which is exactly when you'll be glad it's there.

The Horizon for Everyday Medicine

Here's the future worth building, and it's smaller and better than the one in the keynote videos.

A woman in a rural county with one clinic and a three month wait understands her own bloodwork before the appointment, so the appointment is about treatment instead of translation. A man managing two chronic conditions across four specialists hands each of them a timeline none of them had. A nurse practitioner gets a patient who can describe the pattern, not just the worst day. Multiply that by a few hundred million fifteen minute visits and you get something that no hospital procurement cycle could have delivered: care that fits the person, in a building that didn't have to change first.

Two things this future is not. It's not a world where your chart follows you into a lender's risk model or an insurer's scoring engine, because a copilot that reports to anyone but the patient isn't a copilot. And it doesn't outrank the work that comes first. A planet still burning, still at war, still watching genocides unfold, doesn't get fixed by a better MRI reader, and anyone selling you medical wonder as the top of the list is selling you a smaller world than the one worth wanting. Fix the sky. End the wars. Then, yes, let everybody read their own chart.

The remarkable thing is that this particular piece needs no permission. Not a bill, not a budget, not a vendor. Just a weekend, per last week's argument about passkeys: the infrastructure shifts when enough regular people quietly adopt the better thing.

Before Monday: Your Medical Records Checklist

You've got two days and a portal password. Here's what to do with them.

1. Log into every patient portal you have and export the last two years of labs, notes, and imaging reports into one folder.
2. Open the training and data controls on whichever AI tool you use, and turn off anything that retains or trains on your inputs.
3. Redact your name, address, member number, and date of birth from the exported files before you upload a single one.
4. Run your most recent lab panel through the agent and ask for a plain language summary plus every out-of-range value with its reference range.
5. Ask it for seven questions ranked by how much the answer would change your next step, then cut the list to five.
6. Build the timeline file: one line per visit, date, who you saw, what changed. Twenty minutes now, permanently useful.
7. Print the page. Put it in the bag you're already bringing.

Frequently Asked Questions

• Can an AI agent diagnose me?+
  No, and a well built one will refuse to try. A patient advocacy agent translates the language in your records, tracks what changed between visits, and drafts the questions you bring to your clinician. Diagnosis stays with the person who can examine you, order tests, and take responsibility for the answer. The agent's job is to make sure your fifteen minutes with that person get spent on the right things. If you want the general pattern for scoping an agent this tightly, the task-based workflow library is a good place to see how narrow instructions produce better output.
• How do I get a copy of my own medical records?+
  Start with your patient portal. Under the federal information blocking rule, US providers have to give you electronic access to the health information in your record without unreasonable delay, including clinical notes, lab results, and imaging reports. Most portals have an export or download option that produces a PDF or a structured file. If your care is spread across several systems, export from each one and keep them in a single folder so your agent can read the whole picture.
• Is it safe to paste my medical records into an AI tool?+
  Treat your chart the way you'd treat a tax return. Check whether the tool trains on your inputs and turn that off if it does, strip your name, address, and member numbers before you upload anything, and keep the files in a folder you control rather than a shared drive. A local model or a workspace with training disabled is the safer setup. The clinical content is what the agent needs. Your identifiers are not.
• What should a patient advocacy agent produce?+
  One page. A plain language summary of what the latest results say, a short list of what changed since the previous visit, any value sitting outside its reference range with the range printed next to it, and five to seven questions ranked by how much the answer would change your next step. Anything longer stays in your bag unread. The same discipline that makes a research agent useful applies here: the value is in the compression, not the volume.