Industry
What banks will ask about your AI employee
Tobiloba Odejinmi · 13 May 2026 · 6 min · 919 words
Direct answer
A bank will not ask which model you like. They will ask what it costs to run, how often it breaks, who can change it, whether they can plug in before lunch, and whether you can explain a miss. At Zeeh Africa we got 390+ companies on boring APIs and adoption grew 890% in nine months. The AI employee only survives that meeting if it looks like another dull, logged system, not a new religion.
- Time-to-first-call still decides the deal. Clever setup loses to a documented endpoint.
- Expect questions on cost, uptime, data access, and a kill switch.
- If you cannot replay a miss, the bank will treat the whole thing as risk.
- One engineer who 'just knows' the workflow is a finding, not a compliment.
They want a pipe, not a story
I ran engineering at Zeeh Africa. Banks and lenders did not adopt us because we had a better story about the future of money. They adopted us because identity and account data could be called like any other API, and because the errors were boring enough to handle.
An AI employee attached to that world gets the same interview. If you cannot draw the request, the response, and the failure, you are not in a product meeting. You are in a stall.
The list they actually walk through
I have been in enough of these rooms to stop being surprised. The questions are unglamorous. That is the point. They are trying to see whether you will become an incident on their license.
- How do we connect, and can we do it without a three-week project.
- What does it cost at our volume, and what happens when volume doubles.
- What is the uptime, and what is the fallback when it is down.
- Who can see the data, where it lives, and how we delete it.
- Who can change the workflow, and will we hear about it.
- Can we turn it off without taking our own systems down.
- Can you explain the last miss in language our risk team will accept.
Explainable is not a slide titled Explainable
They will pick a file. They will ask why it moved. If the answer is 'the model thought so', the meeting is over even if nobody says so.
Give them a log, a schema, and a person. That is what explainable means in a bank. Not a heatmap. Not a blog post about alignment.
Cost has to stay dull
Diligence looks at cost. If your employee is a pile of unmetered calls and a surprise invoice, they will treat you like a vendor who cannot count. Count first. Put a number on a successful job and a failed job.
I price a process, not a seat, when I build these. Banks understand a process. They do not understand why a 'digital worker' costs more every time you add a teammate who never uses it.
Single points of failure have names
If only you can explain production, that is a finding. I learned the same thing selling companies: buyers open cost, uptime, and whether one person is the database.
Write the runbook. Name a second person. Keep the workflow in systems other people can open. The bank is not impressed that you are indispensable. They are worried you will get sick.
Make it look like the rest of their stack
The employee should arrive as an API, a queue, and a dashboard they did not have to invent. No new religion. No mandatory mobile app for the compliance officer.
When I say make the API boring, this is why. 890% adoption in nine months is what boring looks like when it works. Your AI employee should be trying to inherit that, not replace it with charm.
Walk in with the file, not the deck
If I were sitting on their side of the table, I would pick one failed job from last month and ask you to narrate it. Who saw it. What the customer felt. How you turned it off. How you shipped the fix. How you know it did not happen again.
That story is the product. The model name is a footnote. Bring the log. Bring the owner. Bring the cost. If you cannot, do not book the meeting yet. Fix the system until the story is dull. Dull is what they buy.
I still write the code. I still sit in the reviews. Banks can smell a founder who only has slides. They have been in that meeting before.
Questions people ask
What is the first question, really?
How do we connect, and how long does that take. If the answer is a workshop and a new portal, you are already behind. Banks want to plug in before lunch.
Will they ask about the model?
Sometimes. They will care more about logs, training data you should not have used, and whether a change can ship without them knowing. Boring versioning beats a brand name.
What about uptime?
They will ask how often it breaks and what happens to their customers when it does. Have numbers. Have a fallback that is a person, not a shrug.
Do they care what it costs?
Yes. Diligence looks at cost. If your employee gets more expensive every time volume is healthy, they will model that as a tax on success.
What did Zeeh teach you here?
Adoption grew 890% in nine months because companies could connect in a day. 390+ companies later, the lesson has not changed. Dull and documented beats clever and closed.
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.
