Value
Token cost is the least interesting number in the business case.
Teams compare model prices and miss the cost of integration, evaluation, supervision, failure, and the old path that never quite turns off. The economic question is what a completed case costs under each design.
18%
ROI in the top decile of organizations — well above the cost of capital. The scaled average is about 7%.
IBM Institute for Business Value, 20261
How I see it
The unit economics of AI work
AI economics starts with a unit: a completed case, a closed ticket, a prepared packet, a reconciled exception. Price the current human path. Price the proposed mixed path. Include the time that remains with people.
Then widen the lens. Build cost, change cost, and three-year operating cost decide whether a cheap-looking pilot is expensive to own. A buy decision that looks cheaper in year one can lose if the workflow is differentiating and the vendor cannot reach into your systems.
Compare designs, not slogans. A copilot that saves drafting time has one cost curve. A redesigned workflow has another. An agent that owns the outcome has a third, with more engineering and more monitoring.
The point is to make the tradeoffs visible before the architecture hardens. Economics should change the design, not decorate a decision that was already made.
How the engagement runs
What you actually buy.
A service is a sequence with an artifact at each step. It is not a transformation program you purchase as a bundle.
Baseline the current cost
Time, volume, error, and capacity as they are today — not a target invented for the deck.
An owner already manages the number.
A benefit story with no starting point.
Build a conservative case
Addressable share, implementation cost, and run cost. If it is not material, stop.
A number that survives a skeptical operating review.
An ROI slide that ignores the cost to operate.
Get the win into the operating numbers
The result has to show up in cost, revenue, or capacity someone already reports.
The win is managed after the project team leaves.
Adoption of a tool as the only metric.
Do not buy this for
These engagements fail for predictable reasons.
Comparing only model invoices
The model is one line. Integration, review, and failure handling are usually larger.
Using company-wide productivity claims
A function-level unit cost is usable. A 10 percent productivity story for the whole firm is not a decision tool.
Forgetting the cost of being wrong
A cheaper case that creates rework, refunds, or compliance exposure is not cheaper.
Before we start
Questions that decide whether to engage.
What is the unit of work, and what does it cost today?
If the unit is unclear, stop. You cannot compare designs.What does the mixed path cost per unit, including human time?
Software cost without remaining labor is incomplete.What is the three-year cost to own the path?
Launch invoices hide the tail.What does a wrong case cost?
If error cost is high, reliability spend is part of the economics.Does the design change if the volume is half, or twice, the forecast?
Fragile economics at the planned volume are not ready to fund.
How we size it
Cost per completed case(software + integration + evaluation + human review + failure cost) ÷ completed cases
Compare that number to the current path. If it only wins at perfect volume, redesign before you build.
This is the wrong engagement if
Do not start this engagement yet.
- There is no unit of work, only a desire to use AI.
- Finance will not engage on loaded cost or error cost.
- The stack is already purchased and economics are expected to confirm it.
- The company wants a thought piece more than a comparison of designs.
A useful next step
Bring the designs you are comparing.
We will price the current path and the proposed paths on the same unit of work, including the people who still have to show up.
Discuss an AI opportunity