AI was supposed to reduce cost. Instead, it created a new, unpredictable cost center
AI didn’t eliminate cost.
It changed how cost behaves.
What Finance expected
The pitch was simple:
- reduce headcount
- automate work
- lower fixed cost
AI was framed like software:
buy seats, deploy tools, realize efficiency.
The assumption:
cost goes down, margin goes up.
What Finance is seeing
The first few months looked fine.
Then the bills showed up.
- inference spend rising month over month
- usage driving cost, not licenses
- teams exceeding “expected” budgets without clear cause
And the question started to surface:
why doesn’t this behave like software?
The break
It behaves like a utility.
Every interaction incurs cost:
- each query
- each retry
- each tool call
- each step in a workflow
And in agentic systems, that work expands:
a task doesn’t execute once
it iterates, branches, calls other systems, and loops until completion
So the cost isn’t tied to:
- seats
- users
- even requests
It’s tied to how much work the system actually performs
The hidden shift
Headcount may go down.
But cost doesn’t disappear.
It reappears as compute:
- dynamic
- usage-driven
- and often invisible until after execution
Two identical “tasks” can produce completely different costs
depending on how the system executes.
Where it breaks for Finance
Finance is accountable for ROI.
But the system commits cost before Finance can intervene.
By the time reporting shows up:
- the work is already done
- the cost is already incurred
- the variance is already baked in
So the process becomes:
observe → explain → adjust
And then it happens again.
Why this matters now
For months, the conversation focused on:
- models
- benchmarks
- infrastructure
- capex
Now it’s shifting to:
- usage
- spend
- ROI
Because the cost is no longer theoretical.
It’s showing up in the P&L.
The real question
This isn’t about whether AI works.
It clearly does.
The question is:
can the cost of that work be predicted and controlled before it runs?
If not:
- ROI becomes uncertain
- scaling becomes risky
- adoption slows
The implication
AI isn’t just a new capability.
And until that cost can be:
- estimated before execution
- constrained before it runs
- and measured in terms that reflect actual work
…it will behave less like software
and more like an unbounded utility.
Finance didn’t get this wrong.
They were given the wrong model.
Now they’re seeing the consequences.
– Published on Monday, March 30, 2026