Use cases
AI employee for internal operations
Tobiloba Odejinmi · 20 Feb 2026 · 6 min · 1,185 words
Direct answer
An internal ops employee does the copy-paste your team still does between two tools. Good looks like a finished record in the system of record, with a log. Escalate money, access changes, and anything that touches a customer. Do not start with a company-wide “agent”. Start with one path. Measure cycle time and error rate on that path.
- The first job is the path people already copy between two tools.
- Good is a finished record and a log, not a chat summary.
- Money, access, and customer-facing changes stay gated.
- Measure cycle time and misses on one path, not a platform score.
Which internal jobs are worth an AI employee?
The ones your team still copies between two tools. A form lands. Someone retypes it into the CRM. Someone else files the PDF. Someone pings Slack to say it is done. That path has a start, a finish, and a miss you can name. That is a job.
Anthropic’s 2026 numbers put 48% of agent use on internal automation and 60% on reporting. That matches what I see. People do not need a mascot. They need the weekly report and the intake that does not wait for Tuesday.
Pick one path with volume. Vendor intake. Clinic onboarding paperwork. A weekly ops report that already has a template. If the first project is “an agent for the company”, you will spend a quarter on a demo and still have the spreadsheet.
What does good internal automation look like?
The record is complete in the system of record. The log says what was written and why. A person who did not build it can replay the path. The old spreadsheet is no longer the source of truth. That last one is the real test. If people still update the sheet, you have two truths. Two truths is how you bounce a payment.
I have sat on systems that moved serious money. A sloppy row is not an ops anecdote. It is someone who does not get paid. Scale is useless if the data is sloppy. An internal employee that writes junk faster is not a win.
Good also looks boring. No new login if we can help it. No new status language. Use the statuses the team already understands. If you need a workshop to explain the workflow, you overbuilt it.
- One system of record for the finish line
- A log a stranger can read
- A named owner for misses
- The old sheet is archived or read-only
What should still hit a person?
Money. Access. Anything that changes what a customer sees. Anything that looks like a legal commitment. The employee can assemble the packet: fields, files, a recommended action. A person approves. That is a hard gate, not a suggestion.
Escalation is a state handoff. The approver gets the record, the diff, and why it needs them. “Looks off” is allowed if you say which field. A Slack ping with a screenshot is how this falls over when the approver is out.
If 10–15% of items need a person, you are in a healthy range for a path with real judgment. If nothing escalates, you are either on a trivial path or you are hiding judgment inside the model. Find out which one before you add volume.
What should you not automate first?
Do not start with payroll, production access, or anything that can email every customer. Do not start by letting a model open tickets in five systems “to be helpful”. You will spend a month cleaning ghosts.
Do not start with the exception path. Start with the happy path that already has a checklist. Exceptions are why you still have a person. If you automate the exception first, you will encode someone’s worst week.
Skip the company-wide knowledge bot as the first internal job. People will ask it for policy, it will miss, and you will have a trust problem before you have a win. File the policy. Then automate the form that already uses it.
How do you measure internal work that nobody sees?
Cycle time on the path you named. Misses on a weekly sample. Share of items that still go through the old sheet. Those are enough. Do not invent an “automation adoption” score to put on a slide. BCG’s 30% versus 13% is a market number. It is not your KPI.
If cycle time falls and miss rate rises, you shipped speed. That is not the goal. If both fall, you can talk about the next path. If neither falls, the employee is sitting beside the work, not in it.
Write the runbook as if the builder is on a plane. I still write code. I also still leave notes a teammate can use at 2am. Internal automation that only you can restart is a single point of failure with a nicer name.
How do you avoid a new login nobody uses?
Build where the work already happens. The inbox, the admin, the warehouse tool. At Zeeh, companies connected in a day. That is why adoption moved. Internal ops is the same product problem with worse patience. Nobody in ops wants a new religion.
If you need an API, make it boring. If you need a person to click approve, put the button in the tool they already open. If you need a report, put it in the same place last week’s report lived.
When we sold Insurpass, diligence was a tour of the boring systems. Cost, uptime, who can change it. Your internal employee will get the same tour one day, even if the buyer is just a new ops lead. Leave logs. Leave an owner. Leave a way to turn it off.
Questions people ask
What internal work is worth an AI employee?
A path that happens every day, has a clear finish line, and currently lives in a spreadsheet or a Slack thread. Intake to ticket. Invoice to folder. Form to CRM. If you cannot name the finish line, it is not a job yet.
Do we need a new ops platform?
No. Connect the tools you already have. If the employee cannot write to the system of record, it is a notepad. I have the same rule I used at Zeeh: another team should be able to use this before lunch.
What should a person still approve?
Payments, refunds, access grants, contract changes, and anything that emails a customer. The employee can prepare the packet. A person clicks the gate. If you skip that, you will find out on a Friday afternoon.
How do we keep this from becoming shadow IT?
Name an owner. Put the logs where the team already looks. Write a runbook a person who did not build it can follow at 2am. If only you understand it, you built a risk. Buyers treat that as a risk. So should you.
What do we measure?
Time from intake to a finished record. Error rate on a weekly sample. Number of times someone still did it by hand. If hand work does not fall, you automated a side path and the real path is still the old one.
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.
