Money
CFO questions before you buy
Tobiloba Odejinmi · 12 Jul 2026 · 6 min · 988 words
Direct answer
A CFO should ask which process is being bought, how success will be measured in thirty days, who owns a miss, what tokens and reviewer time will cost, and what would make the company stop. I charge $5,000, $7,500, or $15,000+ for the build. The run costs stay on your books. Do not approve a strategic ROI claim for a week-one workflow. IBM’s Frontier Firm figure is 9% of leaders. You are probably not in that set yet, and that is fine.
- Buy a named process with an owner, not a category.
- Demand a thirty-day scorecard: hours, exceptions, run cost.
- Put inference, reviewer time, monitoring, and integration on the sheet.
- Refuse strategic ROI language on a first workflow.
What should a CFO ask first?
Ask what we are taking off a desk. Not the category. The steps. I will answer that in thirty minutes or I will tell you there is nothing worth doing. That is the same test I use on the sales call. Finance should not have a lower bar than the working session.
Then ask who owns the thing on day eight. If the answer is 'the vendor' forever, you are buying a dependency. If the answer is 'we will figure it out', you are buying a mess. I hand work over. I do not want to be the only person who understands production. I learned that selling a company. Buyers hate magic.
How do we know the process is worth buying?
It costs hours now. Those hours are boring. A miss is survivable or at least visible. The tools already exist. That is worth $5,000 or $7,500 if the mapping is honest. Several workflows at $15,000+ is worth it only if two or more of those processes are already named.
I will not attach a fake case-study dollar amount to convince you. At SmartComply, review time dropped by about half. That is the kind of claim I will make, because we measured the pile. If we have not measured yours, we do not know your payback, and a confident number would be a lie.
Who owns the miss?
A person with a name. Not 'the model'. Not 'AI'. If a lending file is wrong or a clinic path fails, someone has to say what happened. I keep a human in the loop for that reason. Reviewer time is a cost. It is also how you stay allowed to operate.
Ask to see the escalation path. If context dies when a person takes over, you will pay twice: once for the bot, once for the customer repeating themselves. That is not a rounding error. That is the product.
What is the run cost after month one?
Inference on real volume. Reviewer hours on exceptions. Monitoring that a person actually does. Integration when a vendor moves a field. Ask for those four in the approval memo. The build fee is the easy line. These four decide whether the easy line was honest.
If The week is the buy, monitoring and thirty days of support are in my fee. After that, the owner is still yours. Budget like it.
How do we treat this on the books?
Split build and run. The build is a project against a process. The run is operations. If you capitalize a story, you will stop looking at tokens and reviewer time. Those need to stay visible or they grow in the dark.
When I sat through diligence on Insurpass, nobody opened the vision deck first. They opened cost, uptime, and whether one person could explain the database. An AI employee has to survive that conversation. If it cannot, do not buy it yet.
What would make us stop?
Write the stop rule before you start. Exception rate above a number you pick. A silent miss. No owner for two weeks. Inference that makes the process sillier than a person. Stopping is not failure if you learned the process was the wrong first job.
Capability ROI can still be true when you stop: you learned it runs, or it does not. Realized ROI is the only reason to continue. Strategic ROI is not on the table for a first workflow, and a CFO should strike that sentence if it appears.
What does the 9% figure mean for this purchase?
IBM and Microsoft talk about Frontier Firms: humans and agents sharing work on purpose, with the data and governance to match. Only about 9% of leaders say they are there. That number is a cold shower, and it is useful.
It means your first $5,000 or $7,500 is a process buy. It is not an entry ticket to a club. Approve it if the process is real. Reject any memo that needs you to be in the 9% for the math to work. The math should work on one desk.
Questions people ask
What is the first question a CFO should ask?
What process, in steps, are we buying, and who owns it after handover? If those two answers are soft, the rest of the memo is fiction.
How should we treat the fee on the books?
Treat the $5,000, $7,500, or $15,000+ as a build against a process. Treat model usage and reviewer time as operating cost. Do not bury them in 'innovation' so nobody reviews them.
What run costs should be in the model?
Tokens and inference, human review, monitoring attention, and integration slack. If the vendor cannot talk about those, the vendor is selling a screenshot.
When is this cheaper than a hire?
When one process is eating hours and a miss is visible. It is not a headcount replacement plan. It is a first pass. If someone is selling you a hire you can fire, walk away.
What would make a serious CFO say no?
No written process. No owner. A payback slide with invented dollars. A claim that week one is strategic transformation. I will say no in those cases too.
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.
