engineering notes
How to compute the payback of AI agents
A model for computing AI-agent payback: what to count as benefit, how to account for token and operation cost, which assumptions are dangerous.
In brief for executives. An AI agent’s payback is computed by a specific process, not “by AI adoption”. An honest calculation accounts not only for saved hours but for the cost of tokens and operation — otherwise the figure comes out pretty and false. The expected ROI in surveys runs far above the realized one precisely because benefit is counted and ownership cost is forgotten. Below is a model that counts both sides.
“When will it pay off?” is the right question, usually answered wrong: saved salaries are taken and divided by development cost. That formula lacks half the costs, so it is always optimistic.
Payback is counted by a process — and by full cost of ownership.
Hypothesis: payback is computed by a process, not “by adoption”
An AI agent doesn’t pay off “in general”. What pays off is automating a specific operation with a measurable output. So the unit of calculation is the process: what it cost before and after, accounting for all costs, not just development.
Expectations run far above realized impact. ROI is computed honestly — by a specific process and full cost of ownership, not by a 171% expectation.
The gap between expected (171%) and actually realized (significant ROI — for under 1%) is exactly the price of an incomplete calculation.
Problem: benefit is counted, ownership costs forgotten
The typical model: (saved hours × rate) ÷ development cost. It lacks: token cost on every request, operation and monitoring, support on process change, the share of cases still requiring a human. Each of these is significant in production, and without them payback is overstated by multiples.
Why the usual approaches don’t work
“Count by saved salaries” doesn’t work: saving is only one side; an AI agent, unlike ordinary software, costs money on every request.
“Take expected ROI from surveys” doesn’t work: expectations are systematically far above the realized — this is visible in the data.
“Compute payback after the pilot” doesn’t work as a budget rationale: without a cost model the pilot shows the effect but not the cost of ownership at scale.
Engineering model: how to count honestly
Benefit — by process. Removed repetitive work, shortened cycle time, fewer approvals — in money and in a specific operation, not “across the company”.
Costs — full. Development + operation + tokens at your volume + support on process change + residual human share. Token cost is taken not from the prototype (with dozens of requests) but projected onto the real volume.
Sensitivity to volume. Token cost grows with the number of requests; so does benefit. You compute not a point but a curve: where the process goes positive and at what volume.
Term and risk. Payback is a term (in how long the process goes positive) and the assumptions it holds under. A calculation without explicit assumptions is useless.
What not to count as benefit. “Adopted AI”, “became more modern”, “employees are happy” are not money. Only the measurable goes into the calculation.
Practical takeaway for business
Require the calculation as two curves: benefit by volume and cost of ownership by volume. Their intersection is the payback; one figure without the second curve is marketing.
Per IDC, of 33 launched pilots only about 4 reach production. The cause of failure is not technology — it is the underestimated complexity of taking it to a process.
Most pilots don’t reach production — so payback is computed on a process that will actually go into operation, not on a demo.
Don’t take expected ROI for a plan. Plan by the conservative cost-of- ownership curve; a 100%+ expectation is the market’s appetite, not a forecast.
Apply this to your processes — .
Open questions
How to value decision quality in money (not just its cost) is a task without a common standard. Fixed price or time-and-materials for AI systems is an open question: the process is refined along the way. How to count benefit when it is removed coordination is a cycle-time model, not an industry norm.
Name a process and your request volume — and payback can be computed as two curves before the start. — we’ll assemble benefit and full cost of ownership for your process.