Industry
AI employees in lending
Tobiloba Odejinmi · 5 Mar 2026 · 6 min · 914 words
Direct answer
An AI employee in lending assembles the file, chases what is missing, and explains why a case is stuck. It does not quietly approve. At 10mg Health we have put out $3.4 million in loans at a 60% approval rate across more than 6,000 providers. Those numbers only mean something if a person can still say why a file moved. In the Zeeh era, books on cleaner rails stayed under 5% NPL. Guesses do not keep that number.
- The pile is incomplete files, not a new scorecard you cannot defend.
- A person owns the approve, the decline, and the reason.
- Clean inputs protect the book. In the Zeeh era, NPL stayed under 5% on those rails.
- If you cannot replay a decision, a regulator or a buyer will do it for you.
The file is the product
Most lending work is not a dramatic yes. It is a file that is almost ready. Missing bank statement. Blurry ID. A provider who started and stopped. Someone has to chase that, every day, or the loan never exists.
That is the employee I will build. Not a mysterious score. A worker that lists the gaps, sends the follow-up, and puts a complete packet in front of a person. At 10mg that person still owns the 60%.
Approval rate is not a model trophy
$3.4 million out the door across more than 6,000 clinics and pharmacies is a operations fact, not a pitch for autonomy. The platform has to open. The decision has to come back. The disbursement has to land. Then a human has to be able to say why.
I have no interest in an employee that lifts approval by guessing. You will pay for that in collections, in arguments, and in the meeting where someone asks you to replay a file and you cannot.
Clean rails, cleaner book
At Zeeh we made identity and account data something a lender could call without a workshop. Adoption grew 890% in nine months. 390+ companies plugged in. The unglamorous win was fewer fictional fields in the file.
In that era, NPL on the books that used those rails stayed under 5%. I connect those facts on purpose. If your employee starts inventing a field because the PDF was ugly, you are spending that number down.
- Pull only fields you can point at in the source.
- Refuse to fill a gap with a guess. List the gap instead.
- Hand the complete file to a named officer.
- Store the reason a file moved, in language a stranger can read.
Stuck is a first-class state
Credit ops lives in 'stuck'. The employee should be good at stuck. Why did this stop. Who has the next action. What document is missing. When did we last ask.
If your system only knows approved and declined, you will hide the real work in Slack. Hidden work is how files rot and how patients or borrowers wait for no reason you can report.
Explain it or do not ship it
Banks, buyers, and your own risk lead will ask the same thing. Can you explain this. Not the architecture diagram. The file. Why this one moved. Why that one did not.
Structured output helps. Logs help. A person on the decision helps. A prompt nobody can replay does not. I will not put that in front of a clinic or a lender and call it done.
One queue, then stop
Start with the incomplete-file queue. Map it. Plug into the tools the credit team already uses. Measure time-to-complete-file and miss rate. Do not measure 'AI usage'.
When that loop is boring, you can talk about the next process. Not before.
What a credit meeting should still sound like
A person should be able to open a file and say: here is what we received, here is what was missing, here is what we asked for, here is why we said yes or no. If the employee cannot help that sentence, it is in the way.
At 10mg the path is simple enough to shout about: can a provider start, does the decision come back, does the disbursement land. The employee should make those questions easier to answer, not add a fourth tab called 'AI said maybe'.
I have no patience for a score nobody can defend. Banks will ask. Buyers will ask. Your own risk lead will ask. Build the chase first. Leave the yes with a name.
Questions people ask
Can the employee approve a loan?
It can say the file is complete. A named credit officer approves. If you skip that, you do not have a faster lender. You have an unexplained book.
What should we automate first?
Document chase and the 'why is this stuck' queue. That is where hours go. The scorecard can wait until you can explain the first pass.
How does 60% approval fit?
It is a rate we can stand next to because the path is visible. Speed without a reason is not a better approval rate. It is a mess you will find later.
What did Zeeh change about how you think of credit data?
Identity and account data other companies could call in a day. Cleaner files. In that era, NPL on the books using those rails stayed under 5%. The lesson is inputs, not magic.
What will a bank or buyer ask?
Cost, uptime, and whether you can explain the system. They will also ask who can turn the employee off. Have that answer before the meeting.
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.
