
I Can't Touch the $1.6M Medical Records AI; WorkBuddy Handles the Work It Can't Reach
Last night I was browsing the procurement site and saw that Guangzhou Huangpu District Traditional Chinese Medicine Hospital posted a notice. AI TCM outpatient pre-consultation and medical record generation system, AI inpatient medical record quality control system, 800k each, plus a TCM five-tone guided sleep aid at 44k, totaling 1.644 million, bid deadline November 16th.
My first reaction wasn't about whether it's expensive. It was about all that work before discharge records get archived—by the time this system lands, it probably won't be my job to touch it anymore.
The notice says software integration implementation services. Outpatient pre-consultation, inpatient record quality control—both are pushing on the substance of medical records: whether diagnoses are fully written, whether progress notes are compliant. These are the doctors' and QC department's business.
On the nursing side there's still one more layer. Nursing record sheets, temperature charts, shift handover books—all need to be checked before discharge. Missing items, mismatched times, empty signature fields. These don't count as clinical quality, but when archiving gets kicked back for rework, it's on us.
This layer I currently run with WorkBuddy. Been using it for 1 month, let me explain how I set it up.
Draw the boundary first
It handles form, not substance. I only have it check three things: whether required fields are present, whether the time sequence is out of order, whether signature fields are empty. Whether diagnoses are correct, whether medication is reasonable—I don't touch any of that. Touching it means trouble for myself and trouble for others.
My environment is WorkBuddy Enterprise, department shared workspace. The personal version can run too, but there's no collaboration permission, and it gets messy when several people use it together. May not apply to everyone.
How to configure, how to run
Open WorkBuddy Enterprise, select Workspace on the left, enter the department's shared one. There's a Skill entry in the top right, click new, name it "Nursing Record Form Check."
The word Skill is confusing the first time you see it—you can think of it as a saved set of instructions plus rules, configured once and reused, no need to retype the prompt every time.
Go in and paste the rules list first. Just the hospital's archiving requirements, one per line, hardcoded, don't use words like "etc." or "and so on"—it really will take liberties.
Then set the output format. I require a fixed five-column CSV: bed number, name, missing item type, page location, suggested action. CSV is a comma-separated table file, Excel opens it with a double-click. Here's a key parameter: the instruction must explicitly say "output CSV, not markdown table," otherwise it gives you a block of text with vertical bars, and when you paste it into Excel everything crams into one column.
One more line to hardcode: only output the list, don't touch the original files.
Save. From then on, every Monday morning, export last week's discharged patients' nursing records as PDFs, drop them into the dialog box, select this Skill, and wait for it to run. Don't do more than 20 at a time, or it'll miss some. When it's done, out comes a CSV, sorted by bed number, who's missing what at a glance.
Workspace members are split into two tiers. Junior nurses get read-only, can view the list, can't edit the Skill. Only I and one other QC nurse can edit the Skill. After editing, write the version number and date in the description, like "v3 2026-11-10 added signature field check," otherwise two months later nobody can tell who changed what. Original medical record files don't go into the workspace—they only pass through the conversation, and get deleted after running.
At the start of each month, check whether the hospital's archiving requirements have been revised. This isn't hard, but easy to forget—I set a calendar reminder. One more thing: after each run, pull three for manual review. If something goes wrong, I have to cover it myself.
I've stepped in a few pits too. At first I wanted to say everything at once—rules, format, cautions all stuffed into one prompt—and it replied with a big block of prose, reads smoothly, but I still had to copy it line by line into a table. Later I learned better: rules are rules, format is format, write them separately.
File names need to be standardized too. At first the exported files were called "nursing record(3).pdf" and it would get the categorization mixed up. Now it's unified as "bed number_name_discharge date."
The worst pit was when I casually said "fix the ones with wrong format while you're at it." It actually did—and after fixing, it didn't match the original records, and I spent half the night checking it back. Since then the instruction keeps just that one line: only output the list.
Comparing the two setups: on money there's no comparison—theirs is 1.6 million, mine is just usage within the department quota. Speed of changing rules is way different—a dedicated system needs to go through the IT department to adjust flows, a week minimum; I change one line of rules, five minutes. But when things go wrong it's different—system problems go to the vendor, my problems I cover myself, so the spot-check line can't be skipped. Scope of management is different too—theirs handles substance, mine handles form.
When this system actually goes live, whether the form-check part gets merged in, or each handles its own section, I can't guess. Anyway, someone has to do these corner tasks before archiving—mine keeps running for now.
Physix Frontier