Economics
25%
of AI initiatives have delivered the ROI CEOs expected, over the past three years.
IBM Institute for Business Value, 20261If the agent cannot clearly reduce cost, increase revenue, or create capacity, it should not be funded.
Teams present model quality and adoption. Operators need annual value, investment, and payback on a workflow they can name.
How I see it
AI agent ROI
Agent ROI starts with the current cost of the work. Employees, loaded cost, time on the workflow, volume, and the share that is actually addressable. That is the surface area.
Then apply a conservative improvement. Not a transformation story. A number a controller will not laugh at. Twenty to thirty percent on the addressable share is often more honest than a 10x slide.
Subtract implementation and the ongoing cost to run and operate the agent. The output should be annual benefit, payback, and a first-year view. Sensitivity matters more than a single heroic number.
A $540K surface area with a 25% improvement is $135K. That can justify a $20K diagnostic and a contained build. It does not justify a multi-year program.
Common mistakes
What teams usually get wrong.
Revenue fantasies with no capacity math
Start with time and cost you can see. Add conversion later if you can measure it.
Ignoring run cost
Model calls, review, and operations can erase a pretty first-year ROI.
Counting 100% of the workflow as addressable
Judgment, relationships, and rare exceptions stay human.
A useful diagnostic
Five questions before you fund the work.
Can you name the workflow and its annual cost?
If not, you are not ready for an ROI conversation.What improvement would a skeptical CFO accept?
Use that number, not the vendor’s.What will the agent cost to run each month at expected volume?
If unknown, the ROI is incomplete.When does the investment pay back?
If payback is a story instead of a date range, keep working.What would make this a no-fund?
A useful ROI model includes a kill number.
Economic model
First-year ROI(annual benefit − ongoing cost − first-year implementation) ÷ first-year implementation
Show a conservative, base, and aggressive case. The conservative case is the one to fund against.
Three credible paths
How far should you go?
Do not force one solution. Choose the path the economics, the risk, and the organization can support.
Capacity and cost only
Fund the project on hours and cost you can baseline now.
The workflow is internal and the volume is stable.
May understate a real revenue effect.
Capacity plus a measured commercial lift
Add conversion or retention only after you can measure it.
There is a clean funnel or renewal process.
Easy to double-count.
Do not fund
If conservative value does not cover implementation and run cost, stop.
The surface area is small or the data is not there.
This is the correct outcome more often than teams admit.
When this is the wrong next step
Do not fund an agent here.
- The only available number is a vendor case study from another industry.
- Leadership wants a transformation ROI on the whole company.
- Nobody will give you headcount, volume, or loaded cost.
A useful next step
Bring one workflow. Get guided into production.
We guide the implementation, go deep on the technical path, and stay hands-on through operations — or tell you when a simpler answer is better.
Discuss an AI opportunity