Money
Hours saved are not the whole story
Tobiloba Odejinmi · 14 Apr 2026 · 6 min · 925 words
Direct answer
Hours saved matter, and they are not enough. An AI employee can cut the first pass and still create waits, rework, or a reviewer who is now the bottleneck. Count errors, time-to-correct, customer wait, and the hours you still pay a person to watch the exceptions. If those get worse, you did not save money. You moved it.
- Hours off a desk can hide a slower customer and a busier reviewer.
- A miss that needs cleanup can wipe a week of 'saved' time.
- Measure wait, rework, and exception load next to the hour count.
- If quality fell, the hour number is a vanity metric.
Why do hours saved mislead?
Hours are easy to put on a slide. They are also easy to steal from the wrong place. I have watched teams celebrate a shorter queue while a senior person quietly rebuilt the work at the end of the day. The junior hours went down. The expensive hours went up.
An AI employee is good at the first pass. That is the point. If you only measure the first pass, you will declare victory on day eight. The customer, the clinic, or the lender lives at the end of the path. Measure that end.
What else moves when you automate a process?
Wait time moves. A bot that answers fast and then parks the weird case in a dead queue is not faster. It is louder. Rework moves. A wrong field in a lending file is not a saved hour. It is a file someone has to reopen. Trust moves. If a pharmacist stops believing the screen, they will go around it.
At 10mg the test is still Monday morning. If the path that moves money or lets a provider in is fragile, I do not care how many 'assistant hours' you logged. The same test applies to an AI employee you bolt onto that path.
- Time from trigger to a finished case, not time to first token
- Rework rate on cases the model already touched
- How long a customer or clinic waits for a human after a handoff
- Whether the reviewer is drowning in the new exception pile
How do errors change the money?
I will not invent a cost-per-error for a company I have not sat with. I will say this: one silent miss can cost more than a month of inference. In health credit or insurance, a wrong field is not a typo. It is a person who does not get treated or a policy that should not have bound.
Price the miss as a risk, even if you cannot put a clean dollar on it. If you cannot stand the miss, do not automate that step yet. Automate the step next to it. Hours saved on a dangerous step are not a bargain.
What happens to reviewer time?
Reviewer time is the line most decks skip. The pitch is that the human goes away. The honest version is that the human sees fewer easy cases and more strange ones. Strange cases take longer. If you cut junior hours and raise senior hours, say that out loud.
I keep a person in the loop on purpose. That time belongs on the ROI sheet next to tokens. Hide it and the AI employee looks like a miracle for a quarter, then the team hires the hours back and nobody can say why.
When is a slower hour more valuable?
When the hour is spent on a case that needed judgment. A hiring screen that sends a shortlist and a reason is more valuable than a faster pile of maybe. A support handoff that keeps the thread is more valuable than an instant reply that makes the customer repeat themselves.
I would rather your scorecard show fewer hours and a cleaner handoff than a huge hour number and a mess. The second one is how AI projects get a reputation for being sloppy.
How should you report this to finance?
Give them hours, then give them the quality lines in the same breath. Do not lead with a multiple. Lead with the process name, the owner, and whether customers waited less. If realized ROI is small and quality held, that is still a good first buy. If hours dropped and quality slipped, you do not have a win to report.
Capability is 'it runs'. Realized is 'the desk changed and the work got no worse'. Strategic is a later chapter. Do not use hours saved to jump the queue.
Questions people ask
Why are hours saved a weak headline?
Because the hour can move to a reviewer, a cleanup queue, or a customer who now waits for a person. The timesheet looks better. The process did not.
What should I measure instead?
Measure the hour count and the quality of the hour. Completion without rework. Time from trigger to a finished case. Exception rate. Reviewer hours. Inference. Then decide.
Can saved hours still be real?
Yes. At SmartComply the pile got shorter. That was real. We still kept a person on the weird cases. The win was the loop, not a story that humans left the building.
How do I stop a team gaming the number?
Name the process. Name the owner. Sample the misses. If nobody can explain a wrong output, the hour metric is not safe to show finance.
Does this change how you price?
I still price the process. I will not sell you a 'hours saved guarantee' with a fake dollar amount attached. We pick a process where a miss is survivable and visible.
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.
