A lot of AI ROI conversations start with one question:
How much human work can we automate?
If an AI system saves 1,000 hours, the calculation can look simple.
But labor saved isn't automatically economic value.
Because automation introduces other costs.
What happens when the system is wrong?
Who catches the mistake?
How much does human oversight cost?
How much does recovery cost?
And what happens when the system is making thousands of decisions instead of ten?
I've been thinking about AI value more like this:
AI value = upside from automation − cost of errors − cost of oversight − cost of recovery
Not as a precise accounting formula.
More as a way to think about the economics of AI systems.
Consider an AI agent that saves a company 500 hours a month.
That's the upside.
But suppose it occasionally makes an expensive mistake.
Someone has to monitor it.
Someone has to review edge cases.
Someone has to correct failures.
Some failures may be reversible.
Others may affect customers, revenue, or trust.
The spreadsheet that only counts the 500 hours saved misses that side of the equation.
And there's another variable that makes AI different:
Scale.
A human making one mistake is one mistake.
An automated system making the same mistake across thousands of workflows is an entirely different economic problem.
This is why I think AI startups should evaluate more than:
How much work can we automate?
They should also ask:
What does one failure cost?
How many times can that failure repeat?
How quickly can we detect it?
How much human oversight is actually required?
What does recovery look like?
The interesting part is that the economics can change as the system scales.
More automation can mean more value.
But it can also mean a larger surface area for failure.
So perhaps the real question isn't:
“How much labor does AI replace?”
It's:
“How much net economic value does the system create after accounting for what it costs to be wrong?”
I'm still exploring this idea.
But I think it's a much more useful conversation than simply putting “hours saved” into a spreadsheet.
Top comments (0)