Economics
The first invoice is not the cost of ownership. The third year is.
7%
average ROI once AI projects scale — below a typical 10% cost of capital.
IBM Institute for Business Value, 20261Implementation gets the attention. The durable cost is inference, internal labor, monitoring, maintenance, evaluation, and the next change request.
How I see it
AI agent TCO
TCO for an agent should cover three to five years. Initial build. Licenses or vendors. Model inference. Internal operators. Infrastructure. Monitoring. Evaluation. Security. Change. The people who keep the exception queue moving.
This is how you compare build and buy honestly. A cheaper first project that needs a hidden team is not cheaper. A product with a higher license and a lower operating load may win.
TCO also keeps the portfolio honest. The fifth agent is not free because the first one exists. It adds evaluation, overlap, and another owner.
If you cannot show TCO, you cannot show ROI. You can only show a launch.
Common mistakes
What teams usually get wrong.
Calling implementation the investment
The investment continues every month the agent runs.
Leaving internal labor out
Owners, reviewers, and engineers are part of the system.
Assuming year-two cost falls automatically
It falls only if you design for it and measure it.
A useful diagnostic
Five questions before you fund the work.
Have you listed run cost, people, and change, not only the build?
If not, the TCO is a project budget.Who is in the operating cost in year two?
If the answer is “the vendor,” look again.What happens to cost if volume doubles?
Linear surprise is how TCO blows up.What is the exit cost if you turn the agent off?
Ownership includes the fallback process.Does TCO still beat the current workflow?
If not, do not scale.
Economic model
3-year TCOimplementation + 3 × (run cost + internal labor + monitoring + maintenance + governance)
Split by category. A single lump number hides the line you can actually manage.
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.
Pilot TCO
Price a 90-day supervised path, including review time.
You are deciding whether to build at all.
Will understate production operations.
Production TCO
Include year-two operations, evaluation, and change.
You are asking for implementation funding.
Needs honest volume and review assumptions.
Portfolio TCO
Add shared platform and overlapping agent cost.
A second or third agent is on the table.
Easy to use this as a reason to build a platform too early.
When this is the wrong next step
Do not fund an agent here.
- The company will only look at this year’s project budget.
- Vendor pricing is unknown and nobody will ask.
- Internal time is treated as free.
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