Money
Hidden costs of AI employees
Tobiloba Odejinmi · 24 May 2026 · 6 min · 903 words
Direct answer
The hidden costs of an AI employee are token and inference spend, the hours a human still spends on exceptions, monitoring that someone actually reads, and integration work when tools change. They are not hidden because they are exotic. They are hidden because the sales deck stops at the build fee. I put them on the table before you buy $5,000, $7,500, or $15,000+ of process work.
- Inference is a volume bill. Price it on real cases, not the demo.
- Human-in-the-loop time is a line item, not a rounding error.
- Monitoring without an owner is how silent misses pile up.
- Integration is never finished if the vendor can change an API.
Why does the invoice look smaller than the project?
Because the invoice is for the build. I charge $5,000, $7,500, or $15,000+ to make a process live. That is real work. It is also the part that photographs well. The rest looks like 'just software' until month two, when the model bill, the reviewer, and a broken webhook show up in the same week.
I would rather you be slightly annoyed by a complete list now than surprised later. Surprise is how AI employees get pulled. Pulled systems do not get realized ROI. They get a postmortem.
What do tokens and inference actually cost?
They cost whatever your volume, prompt size, retries, and model choice add up to. A support reply that pulls three fields is one thing. A compliance document that needs structured output and a second look is another. I used structured output at SmartComply so the model had to return a shape we could check. That is more tokens than a vibe. It is also how you stop invented fields.
Do not take a vendor’s per-seat 'unlimited usage' at face value if the usage is your data leaving through a meter you cannot see. Ask for a way to read the meter. If they cannot show you one, the hidden cost is opacity.
Why reviewer time belongs on the budget?
I do not ship reviewer-less systems for work that can embarrass you. The first pass can be automatic. The miss still needs a name. That name has a salary. Even a few hours a week is a cost you should write down so nobody calls the project 'free to run'.
If realized ROI depends on pretending the reviewer is idle, the ROI is fake. Capability can exist with a heavy reviewer. Realized ROI cannot. Strategic ROI, the 9% Frontier Firm version, assumes humans are conducting the work on purpose, not hiding in the spreadsheet.
What does monitoring cost if you take it seriously?
Time and attention. The week includes monitoring after go-live because I have watched teams skip it. A clipboard in a warehouse does not update itself. Neither does a workflow log. Someone has to look. Someone has to get paged when the connector dies.
The cash cost might be a small tooling line. The real cost is the owner. If you cannot name that person, you are buying a system that will fail politely until a customer names it for you.
What integration work is not in the demo?
Auth that expires. Sandbox fields that do not exist in production. Rate limits. The one custom object your CRM uses for 'real' status. The note that has to land in the same thread a human already started. Demos skip those. Production does not.
At Zeeh the product won when another team could plug in before lunch. That took boring API work, not a keynote. An AI employee attached to a messy join will spend its life in that join. Budget a little slack. The process will use it.
How do you put these on a sheet before you sign?
Four lines under the build: inference, reviewer hours, monitoring owner, integration slack. Give each a number or a name. If a line is 'unknown', say unknown. Unknown is allowed. Invisible is not.
Then decide if $5,000 or $7,500 still makes sense against the process you picked. If the hidden lines dwarf the build, you may still buy. You just buy with your eyes open, which is the only version of this I will help you do.
I would rather a CFO see four small, ugly numbers than one clean fee and a surprise in month two. Write the lines. Name the owner. Then sign, or do not.
Questions people ask
What are the four costs people forget?
Tokens and inference. Reviewer time. Monitoring. Integration. If your quote has none of those, it is a build quote, not an ownership quote.
Are token costs large?
They can be small on a tight first pass and ugly on a long-context review that retries. I will not guess your invoice. I will tell you to measure a week of production before you lock a budget story.
Why is reviewer time 'hidden'?
Because the pitch is autonomy. The work is a first pass. Someone still owns the miss. If that someone is a senior hire, the hidden cost is not hidden for long.
What does monitoring actually mean?
A log you can open, an alert that reaches a person, and a weekly look at exceptions. A dashboard nobody reads is decoration.
Who pays when a tool changes?
You do, unless we agreed I am still on support. APIs move. Fields get renamed. The AI employee does not get a moral exception. Budget slack for the join.
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.
