Industry
AI employees for African ops teams
Tobiloba Odejinmi · 18 Feb 2026 · 6 min · 912 words
Direct answer
An AI employee for an African ops team has to survive WhatsApp, patchy power, and a ledger that cannot be 'usually right'. I have sat on systems at Coral Stone that moved more than ₦2 billion in a month. If a row is wrong, someone does not get paid. Importing a Silicon Valley stack and hoping Lagos will adapt is how you get a demo that dies the first time NEPA takes the light.
- Build for WhatsApp, retries, and a human fallback. Do not assume Slack, Notion, and a perfect CRM.
- Scale is useless if the data is sloppy. ₦2B months teach that faster than a workshop.
- The first job is the copy-paste between tools, not a new dashboard the team will not open.
- If it cannot fail loud and hand off, it does not belong on a live book.
Lagos is not a constraint. It is the spec.
I build from Lagos. The useful systems here already assume the lights will go, the network will flake, and the real conversation is on WhatsApp. An AI employee that needs a perfect laptop, a perfect CRM, and a two-hour onboarding call is a campus project.
The spec is: can a person on a phone finish the step. Can the job retry. Can someone see that it failed. If you cannot say yes, do not attach it to money or customers.
A wrong row is not an engineering metric
I have sat on systems that moved ₦2 billion in a month at Coral Stone. Scale is useless if the data is sloppy. The ops team does not need another chart. They need the posting to be right when someone is waiting on the other side.
That is why I do not let an employee 'fix' a ledger. It can draft the match. It can point at the break. A person posts. If that sounds old, keep it. Old is how you still have a book.
Volume arrives before headcount
Kwaba got to about 164,000 users. You do not staff that with a Silicon Valley ratio. You staff it with a small team and processes that do not require a meeting for every exception.
An AI employee helps when it takes the first pass off that team: the repeated question, the incomplete signup, the ticket that is the same photo again. It does not help when it adds a tool they have to babysit during a blackout.
WhatsApp is the interface
Most of the work I see in African ops already lives in chat. The customer sent a receipt. The agent forwarded it. Someone typed it into a sheet. Tomorrow they will do it again.
Put the employee there. Read the message. Pull the fields. Write them into the system you already run. Reply with what is still missing. If you make them open a portal, you have not taken work off the plate. You have added a tab.
- Read the inbound chat and extract a shape you can check.
- Write into the ledger or CRM you already trust.
- Queue and retry when the network dies.
- Escalate with the original message still attached.
Do not import a religion
I have watched teams try to install a full US stack and then wonder why nobody uses it. The stack was not wrong in California. It was wrong for a floor that runs on cash, naira, and a generator.
At Zeeh, the thing that scaled was a boring API other companies could call in a day. Same idea here. Meet the team in the tools they already open. Then take the copy-paste away.
Fail loud
When power drops mid-job, the employee should not pretend it finished. It should say it did not. A person picks it up. That is the whole reliability model.
If you cannot explain a miss the next morning, you are not ready for a ₦2B book or a 164k-user inbox. Write the runbook. Name the owner. Then go live on one process, not twelve.
I started on Fiverr. Two hundred and seventy jobs. The useful ones were finished and used the next morning. African ops is the same test at a larger scale. If the employee cannot finish a step a person on a phone can finish, it is not ready. If it cannot say it failed when the tower died, it is not ready either.
Questions people ask
Why not just use the same tools US teams use?
Because the work happens on WhatsApp, the power drops, and the customer will not open a fifth tab. A tool that needs a stable office and a SSO workshop will sit unused.
What did Coral Stone teach you?
I sat on systems that moved more than ₦2 billion in a month. If a row is wrong, someone does not get paid. That is the standard I use for any employee you attach to operations.
How does Kwaba fit?
Kwaba reached about 164,000 users. At that size the ops pile is not theoretical. Support, onboarding, and follow-up have to work when the team is small and the inbox is not.
What should an African ops team automate first?
The fields someone copies from a chat or a PDF into a sheet, every day. Then the follow-up that slips when the network dies. Leave judgment with a person.
Can this work offline?
The employee cannot pretend the tower is up. Queue the work. Retry. Tell a person when it did not land. Silent failure is worse than no automation.
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.
