Money
How much does an AI employee cost?
Tobiloba Odejinmi · 8 Mar 2026 · 6 min · 1,113 words
Direct answer
An AI employee costs a build fee and a run fee. I charge $5,000 for one workflow, $7,500 for The week, and $15,000+ for several workflows. After handover you still pay tokens and inference, the hours a person spends reviewing the weird cases, monitoring, and any integration work the process still needs. If someone quotes you only a seat price, they are hiding the bill.
- The invoice for the build is not the cost of ownership.
- One workflow is $5,000. The week is $7,500. Several workflows start at $15,000.
- Tokens, reviewer time, monitoring, and integration keep showing up after day seven.
- Budget the first ninety days as a package, not a one-line fee.
What does an AI employee cost to start?
People ask for a monthly number because that is how they buy software. An AI employee is closer to a short build with a residual bill. I sit down, we pick the process that eats the most hours and is boring enough to automate, and I make it run in the tools you already have.
One workflow is $5,000. You get a live process, a write-up your team can follow, and a handover call. The week is $7,500. That is the same job with a plan for what to automate next, proper API connections, monitoring after it goes live, and thirty days of support. Several workflows start at $15,000 because the second process is never just a copy of the first. The architecture has to hold both.
I do not pretend those figures cover tokens for the next year. Anyone who does is selling you a slide.
What is included in the build fee?
The build is the mapping, the connections, the tests, and the moment someone on your team can explain it without me in the room. At Zeeh we learned that adoption dies when setup needs a workshop. The same rule applies here. If the AI employee needs a new religion, the price is already wrong.
What you are not buying is an unlimited license to point it at every pile in the company. $5,000 is one process. If the first call reveals three processes that share no tools and no owner, we are not going to squash them into one invoice and call it a deal.
- A written path: trigger, steps, tools, and the cases a person still sees
- A live run in production, not a recorded demo
- Logs a reviewer can open without guessing
- A named owner on your side after handover
What does it cost to keep one running?
After day seven the bill changes shape. The model charges you for tokens and inference. A person still reads the exceptions. Someone has to notice when a connector fails on a Monday. If a vendor changes an API, someone has to fix the join.
None of that is exotic. I have sat on systems that moved real money. At 10mg a broken path is a clinic that cannot serve a patient. You do not get to treat monitoring as a nice-to-have after the screenshot looks good.
Put a monthly line on the sheet even if the number is small. A line that exists gets reviewed. A line that does not exist becomes a surprise in month three.
How do token and inference costs show up?
They show up as a function of volume and how much context you stuff into each run. A first-pass screen of a CV is cheap if the prompt is tight. A document review that hauls in a whole PDF and a knowledge store is not. I will not invent a per-ticket dollar figure for your stack. I have not seen your volume.
What I will say: measure tokens on the real path, not the happy-path sample. The weird case is longer. The retry is longer. If you only price the demo, you will be shocked the first week a queue actually fills.
Who pays for the human in the loop?
You do. I do not ship a reviewer-less system for work that can hurt a customer or a clinic. At SmartComply we took the first pass off people. We did not take the name off the miss. Review time dropped by about half. The remaining half is still a cost.
Budget the reviewer as a fraction of a role, not as zero. If the process is inbound support, someone still owns the tickets that do not fit. If you hide that time, the AI employee looks cheaper than it is, and then the team quietly puts the pile back on a person.
What should you budget for the first ninety days?
Budget the build. Budget a conservative inference number from a week of real volume. Budget reviewer hours as if the exception rate is higher than the demo. Budget a few days of integration slack for the tool that will change. Then look at the total, not the $5,000 line.
If that package is still worth it against the hours the process currently burns, buy it. If you need a fantasy payback slide to make the number work, wait. I would rather tell you that on a thirty-minute call than watch you fund a seat you cannot explain to a CFO.
Questions people ask
How much does an AI employee cost to start?
For the work I do, one live process is $5,000. The full seven-day build is $7,500. More than one process across the company starts at $15,000. That is the build, not the rest of the year.
Is that a monthly subscription?
No. I price the process, then I hand it over. Model usage is separate. If you want me to keep tightening it, that is a different conversation, not a silent seat charge.
What do I still pay after it is live?
Inference. The time a named reviewer spends on exceptions. Whatever you use to watch logs and failures. Any extra connector work when a tool changes. Those four lines are the ones people forget to put in the deck.
Can a cheap tool replace the $5,000 build?
A chatbot in a vacuum is cheap. A process that reads your ticket queue, writes back into the same system, and leaves a trail a person can audit is not the same product. You are paying for the mapping and the handover, not a login.
How should finance compare this to a hire?
Do not compare it to a full-time salary. Compare it to the hours on one named process, plus the cost of a miss on that process. Then add the run costs. If that still looks messy, you do not have a pricing problem. You have a process you have not written down.
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.
