Industry
AI employees in healthcare clinics
Tobiloba Odejinmi · 4 Dec 2025 · 6 min · 1,071 words
Direct answer
An AI employee in a clinic is not a doctor and it is not a credit officer. It chases the incomplete file, answers the same provider questions, and flags the cases a person still has to see. At 10mg Health that work sits under a harder rule: if a pharmacist cannot open the system on Monday morning, nothing else you shipped that week matters.
- Start with the pile that delays treatment, not a chatbot on the website.
- The model can collect and sort. A named person still owns the yes, the no, and the weird case.
- Reliability is a patient problem. Measure whether the path opens, not 'AI usage'.
- If you cannot explain a miss at 2am, you are not ready to put it in front of a clinic.
The clinic does not care about your model
I lead engineering at 10mg Health. We built the platform from zero. It now serves more than 6,000 clinics and pharmacies and has put out $3.4 million in loans. None of those providers asked me what model we use. They asked whether the application opened, whether the decision came back, and whether the disbursement landed.
That is the frame I use for an AI employee in healthcare. The job is the work a person is already doing by hand: chasing papers, answering the same three questions, moving a file from 'started' to 'ready for a human'. If you start with a waiting-room chatbot, you are solving a slide, not a Monday.
What is safe to take off the desk
There is a line I will not cross. The system does not diagnose. It does not quietly decide credit. It does not send a patient a medical instruction that nobody reviewed. Healthcare already has enough ways to be wrong without adding a silent one.
The useful work is boring. A provider starts an application and stops. Someone has to ask for the missing document. Someone has to answer why a field failed. Someone has to queue the file for the person who actually underwrites. That someone can be an AI employee if the output is a shape you can check and the miss has a name next to it.
- Chase the incomplete file and say what is still missing.
- Answer the provider questions the team already answers every day.
- Assemble a packet a credit officer can read in one sitting.
- Hand the weird case to a person before anyone calls it a decision.
Monday morning is still the test
I have sat in reviews where someone showed a slide about uptime. Then a clinic called because a loan would not go through. The slide did not help. The person who knew the payment path did.
An AI employee inherits that test. If the workflow is down, a patient may not get treatment that day. That changes what you call an incident. You watch the paths that move money and the paths that let a provider log in. Everything else can wait.
Pretty dashboards do not replace a person who can explain the database. The same is true of a model that 'usually works'. Usually is not a standard when the user is a pharmacist with a queue.
Keep a person on the judgment
Sixty percent approval is a number we can stand next to because a person still owns the yes and the no. The file can arrive cleaner. The first pass can be automatic. The decline still needs a reason a human can say out loud.
If that sounds slow, try explaining a silent miss to a clinic. They will not care that the model was confident. They will care that a patient waited and nobody can say why.
I do not ship a reviewer-less system for this kind of work. The first pass can be automatic. The miss still needs a name next to it.
Write the runbook before the demo
If the only person who can explain a failed application is the person who wrote the prompt, you have a risk, not a product. The runbook has to be readable at 2am by someone who did not write the code.
That means logs you can search, a clear owner, and a way to turn the thing off without taking the whole platform with it. Clinics will forgive a slow afternoon. They will not forgive a black box that nobody can rewind.
What I would automate first
If you run a clinic network or a health credit desk, do not start with a grand 'AI strategy'. Pick the process that burns the most hours and delays the most treatment. Write down the steps, the tools, and the cases a person still has to see.
Then plug into what the clinic already opens. No new religion. No new login if you can help it. Live in a week, with someone who can explain it. That is the whole job.
I still write the code. I still sit in the reviews. If you have one clinic process that still runs on people copying things, that is usually the first job. Book thirty minutes. If there is nothing worth doing, I will say so.
Questions people ask
Should an AI employee decide who gets care or credit?
No. At 10mg the platform has put out $3.4 million in loans at a 60% approval rate. Those numbers only hold if a person still owns the judgment. The employee can assemble the file. It should not quietly approve or decline.
What is the first job worth automating in a clinic network?
The incomplete application. Missing documents, unanswered provider questions, and follow-ups that sit in a inbox. That pile delays treatment more often than a missing feature does.
Do you replace the clinic's existing tools?
No. Plug into what they already open. A new login dies the first week a locum is on shift and nobody remembers the password.
How do you know it is safe enough to go live?
You can explain every miss. There is a named person for the cases that do not fit. The runbook is readable at 2am by someone who did not write the code. If that is not true, wait.
What does 6,000 providers change about the design?
You stop designing for the demo clinic. You design for the Monday morning when half the network is trying to log in and one payment path is slow. Volume is how you find out what is fragile.
Written by
Tobiloba Odejinmi
Head of Engineering at 10mg Health. I have run engineering at Zeeh Africa and sold Insurpass and Shopl. I still write the code. If you have one process that still runs on people copying things, we can look at it in thirty minutes.
