How to Measure ROI of AI Agents: Baselines and Costs

1. 1Expected benefit and its baseline
2. 2Review, delivery and operating costs

Scenarios expose assumptions. They are not achieved returns or promises.

To measure the ROI of AI agents, start with a baseline, count the full cost including human review and maintenance, and state value as a range. Agentic ROI turns those inputs into a 0–100 ROI confidence score; it is not a promised return. It tests whether an opportunity's expected value is worth building for and how confidently the team can defend that case. It is derived from the Canvas and appears under the Agentic Brain hub.

What the 0–100 score means

The score is ROI confidence. It is not an ROI percentage, a promised financial return, or a probability that the project will succeed.

• 70–100: Strong case — the value and supporting evidence are comparatively strong.
• 45–69: Promising, measure to confirm — the case has potential but still needs measurement.
• 1–44: Baseline-dependent — the claim depends heavily on missing or weak baselines.
• 0: Not yet a case — the Canvas does not yet support a meaningful value case.

The five dimensions

• Value Potential — whether the Canvas identifies credible time, cost, revenue, capacity, quality, risk, or speed improvements.
• Baseline Evidence — whether current volumes, handling time, cost, errors, rework, conversion, or similar measures are known.
• Cost Complexity — whether implementation, integration, data preparation, model usage, human review, monitoring, governance, and maintenance costs are understood.
• Readiness Dependency — whether readiness gaps threaten realization of the expected value.
• Payback Confidence — whether the combined evidence supports a SOW, prototype, pilot, build, pause, or further-measurement decision.

Each dimension is labeled Validated, Estimated, Missing baseline, or Unknown and can include findings, open questions, and a recommended action.

Value drivers and effect order

The assessment can identify time saved, cost reduced, revenue captured, capacity unlocked, quality improved, risk avoided, and speed to market. It also separates effects into:

• Upstream preconditions that must be true before value can occur;
• First-order effects that are direct and more defensible; and
• Downstream effects that may follow but should not yet be treated as committed value.

This prevents a distant strategic benefit from being presented as though it were a directly measured saving.

When numbers appear

Agentic ROI can show confidence, drivers, and missing evidence without displaying a dollar figure. Numeric annual value and payback ranges need the relevant baseline inputs. These may be supplied measurements or explicitly identified estimates; an estimate is not independently verified evidence. Available inputs include monthly case volume, handling minutes per case, loaded hourly cost, error or rework rate, annual run cost, estimated implementation cost, and — for the revenue side of the case — deal value, expected conversion uplift, leads per month, baseline conversion rate, and gross margin.

For a time-saving case, the calculation basis combines volume, handling time, loaded cost, and an estimated recoverable-time range. The How we calculated this disclosure on the ROI page shows the basis used for the current Canvas. Payback also requires an implementation-cost estimate.

Every input is optional. Missing inputs remain visible as baselines to measure next instead of being fabricated by the model.

The revenue upside scenario

Downstream revenue (leads × conversion uplift × deal value, gross-margin-adjusted when a margin is supplied) is never folded into the conservative net benefit, ROI %, or payback. Instead it powers a clearly labelled second scenario — "with revenue upside — expected, not committed" — with its own net, first-year ROI, and payback figures. When the savings-only case is negative but the upside scenario is clearly positive, the verdict reads "Pays for itself if the revenue upside lands", so a deal-driven automation is presented as a case to validate rather than a dead end, without ever claiming uncommitted revenue as cash.

Relationship to the Canvas and Readiness

The authored Business Impact & ROI section captures the team's claims and assumptions. Agentic ROI evaluates them with evidence from goals, target outcomes, use case, constraints, systems, workflow, and AI Readiness.

AI Readiness asks, "Can we build this responsibly?" Agentic ROI asks, "Is the value worth building for, and how strong is the evidence?" A project can be valuable but unready, or ready but not yet supported by a compelling value case.

Working with the assessment

The ROI page can copy or download a SOW-oriented summary, edit baseline inputs, and explain the full breakdown. Its Brain conversation can explain calculations and save explicit answers to the active structured baseline question. It can also present baseline proposals for you to review and apply. It cannot edit authored Canvas sections from the ROI surface.

Canvas-derived numbers remain labeled estimates until confirmed. Confirmation records the user's acceptance; it does not independently measure or audit the underlying business data.

The current product computes the present assessment from the latest Canvas. It does not yet provide persistent ROI snapshot history.

To fill in the baseline inputs, work through the AI use case canvas guide. To turn a value case into a pilot with a success number, use the agentic AI implementation plan. Check readiness separately with the agentic AI readiness checklist. If an earlier pilot stalled on value, read why AI projects fail.

Frequently asked questions

What does the Agentic ROI score mean?

The 0 to 100 score is ROI confidence. It is not an ROI percentage, a promised financial return, or a probability that the project will succeed. A score of 70 to 100 means a strong case, and 1 to 44 means the claim depends heavily on missing or weak baselines.

Why does Agentic ROI sometimes show no dollar figure?

Numeric annual value and payback ranges need baseline inputs such as monthly case volume, handling minutes per case and loaded hourly cost. Without them, Agentic ROI still shows confidence, drivers and the baselines you should measure next, and it does not make numbers up.

Is a revenue upside counted as savings?

No. Downstream revenue is never folded into the conservative net benefit, ROI percentage or payback. It appears as a separate scenario labelled as expected, not committed.

How is Agentic ROI different from AI Readiness?

AI Readiness asks whether you can build this responsibly. Agentic ROI asks whether the value is worth building for and how strong the evidence is. A project can be valuable but unready, or ready but not yet supported by a compelling value case.

What are the hidden costs of an AI agent project?

The costs around the model: integration, data preparation, human review, monitoring, governance and maintenance. Agentic ROI scores whether these are understood under Cost Complexity, and leaves them visible as unknown when they are not.

Next: Outputs.