How to measure AI agent ROI
Updated 1 October 2026 · Alem product methodology
Measure one run against one defined outcome. A model response, an approved spending request and a successful payment are different events. Mixing them makes ROI look better than the evidence supports.
Calculate both return and ROI
Return multiple = attributed value / cost. ROI percentage = (attributed value − cost) / cost × 100. With zero cost, report ROI as undefined rather than infinity. Compare runs in the same currency and reporting period.
| Illustrative run | Amount | Interpretation |
|---|---|---|
| Recorded cost | $500 | Model, tools and services included in this example |
| Attributed value | $1,000 | Operator estimate, not necessarily cash received |
| Net attributed value | $500 | Value less recorded cost |
| Return / ROI | 2× / 100% | Two different ways of expressing the same calculation |
This is a worked example, not a verified customer result. Current aggregate values are available on Data.
Define attribution before running the agent
- Choose a measurable outcome: collected revenue, avoided expense or time saved.
- Set a baseline. What would happen without the agent?
- Include human review and correction time in costs, where relevant.
- Record the source of each amount and the measurement period.
- Separate cash revenue from estimated savings; do not add overlapping benefits twice.
For time savings, write down minutes saved and the hourly rate used. Two hours at $30/hour implies $60 of estimated value, not $60 deposited into a bank account. Revenue should account for refunds and costs of delivery when those are material.
What Alem records
Alem stores cost events and attributed outcome events separately. A Challenge card is a snapshot of the recorded totals when the owner creates it. Updating later outcomes does not make an earlier snapshot a new measurement. The optional screenshot can help explain a claim, but attaching evidence does not independently verify it.
Frequently asked questions
Does a screenshot prove revenue?
No. It can document a claim, but it does not establish payment settlement, ownership or causation. Public cards are labelled self-reported.
Is sandbox settlement actual money movement?
No. It is a record supplied through the sandbox. Real payment confirmation requires an execution integration and reconciliation.
Can I compare two agents?
Only with the same task, baseline, cost coverage, currency, period and attribution method. A high percentage on a tiny or selectively reported run is not a portfolio benchmark.