Suppose an agent assembles a supplier approval package in minutes. Finance then spends half an hour checking its figures, procurement replaces a missing attachment, and legal confirms which agreement version applies. The package moves forward, but the agent's response time explains little about what the work cost.
AI agent ROI measures the financial return attributable to an agent deployment relative to its incremental cost over a defined period. Establishing that return requires a record of the completed process, including the work people still perform.
The September 2026 IDC InfoBrief sponsored by Leah reports that about 21% of organizations were not yet measuring ROI on AI agent deployments. A pilot should establish that measurement before its scope and spending grow.
TL;DR
- IDC reports that about 21% of organizations were not yet measuring ROI on agent deployments. That finding does not establish whether those deployments were profitable.
- Define a completed business outcome and compare equivalent cases against the existing process. Include review, corrections, failed attempts, and human fallback.
- Report capacity released separately from realized cash savings. Count each benefit once and include implementation, operating, and supervision costs.
- Expand when results hold across representative work, required controls function, and an accountable owner can explain the costs and benefits.
What the IDC research says about measuring value
The IDC InfoBrief, The Three Pillars of the Agentic Enterprise, describes a survey of 410 enterprise decision-makers actively deploying or evaluating AI agents. Its figures identify the underlying source as Leah's Agentic AI Adoption Study, April 2026.
Page 9 reports the measurement gap: about 21% of organizations were not yet measuring ROI on agent deployments. It does not provide a verified financial return for the remaining organizations. Measuring a return and achieving a positive return are different findings.
Page 10 also discusses CFO concerns about cost predictability under token-based applications and AI platform pricing. Together, these observations support a practical starting point: connect the cost of running agents to the business work they complete.
The scorecard and calculations below are an evaluation approach for that task. They are not IDC benchmarks or reported Leah customer results.
Define what counts as completed work
Choose one recurring process with a clear start and finish. For supplier approvals, the start might be receipt of an eligible request. Completion might require the current document, the required decisions, and a confirmed update in the destination system.
Agree on the quality and approval requirements before the pilot. A package that needs material correction is unfinished until someone completes that work. A case that safely stops for an authorized review may be behaving exactly as intended.
In Demystifying evals for AI agents, Anthropic distinguishes an agent's account of its work from the resulting state of the environment. Apply that distinction to commercial operations by checking the agreement, approval, or system record that establishes completion.
Record whether the agent completed its assigned scope, which human steps it required, and whether the overall process finished. Human fallback can produce a successful business outcome while revealing a limitation in the agent's contribution. Keep both facts visible.
Establish a baseline you can compare
Measure the existing process using the same eligibility rules and completion criteria as the pilot. Capture staff handling time, elapsed time, correction work, and the relevant operating costs.
Separate simple cases from exceptions. A pilot that receives standard agreements should be compared with standard agreements, rather than a baseline dominated by negotiated or incomplete requests.
Where practical, compare similar cases running through the existing and agent-assisted processes during the same period. Random assignment can strengthen the comparison when operational conditions allow it. If the comparison is historical, record changes in staffing, workload, policy, and systems that could also explain the result.
Follow each group of incoming cases to a defined cutoff. Report unresolved work and the age of the backlog alongside completed cases. Otherwise, the difficult cases can remain open while the reported average improves.
Use a representative mix and enough observations to understand the variation. A small set of demonstrations can expose defects, but it cannot establish how often rare failures will occur in production.
Use a scorecard that includes the work around the agent
Keep the headline measures readable enough for the process owner to review with finance and operations.
| Measure | What to record | What it helps establish |
|---|---|---|
| Verified completion | Completed cases meeting the agreed criteria, with unresolved cases reported separately | Whether the process delivers the required outcome |
| First-pass acceptance | Cases accepted without material correction, including any planned review, divided by all eligible cases in the group | How much work needs repair or another attempt |
| Staff handling time | Preparation, review, correction, escalation, and fallback time | Whether staff effort actually falls |
| Elapsed cycle time | Time from the agreed start to completion, including queues | Whether the business gets the result sooner |
| Cost per completed case | All in-scope processing costs, including failed attempts and fallback, divided by verified completions | The cost of delivering an accepted result |
| Control failures | Unauthorized actions, missed approvals, and the circumstances of each event | Whether financial gains depend on unacceptable behavior |
| Cost and delay variation | Typical results and expensive or slow cases, separated by case type | Whether averages conceal an operating problem |
| Realized benefit | Documented spending reductions or other finance-approved financial benefits | What belongs in the financial return calculation |
Use consistent cost boundaries for both processes. If staff time is included in cost per case, do not add the same review cost again elsewhere in the comparison. Include costs incurred on unfinished cases during the measurement window and explain how work in progress affects the result.
For staff time and elapsed time, review the distribution as well as the average. A faster typical case can coexist with longer waits for exceptions. The orchestration article in this series explains the dependencies and recovery behaviors behind those exceptions.
Count the full incremental cost
An agentic AI ROI calculation needs a consistent comparison with what the organization would otherwise spend to deliver the same work.
Record implementation costs such as integration, data preparation, testing, security review, and training. For ongoing costs, include platform and model usage, connected services, monitoring, support, evaluation, and the human work that remains. Include recovery and remediation when failures create additional work.
A July 2026 McKinsey interview with Pay-i CEO David Tepper discusses how verification, rework, and different execution paths affect agent economics. Its cost discussion is a useful reason to measure entire cases instead of extrapolating from the price of an individual model call.
Set a reporting horizon before calculating returns. A first-year cash view includes the implementation spending paid in that year. An ongoing operating view answers a different question. Label both clearly when they are useful, and apply the organization's finance rules consistently.
Keep unmeasured costs visible as gaps. Treating them as zero can make a pilot look more attractive without improving its economics.
Separate capacity released from cash saved
When an agent reduces staff handling time, it creates capacity that the organization can use. Record what happens to that capacity: more cases completed, a shorter queue, less overtime, reduced external support, or time redirected to other work.
Multiplying hours released by a loaded hourly rate gives an estimate of labor capacity value. A cash-savings claim needs evidence that spending actually changed. An avoided hire also needs a documented hiring requirement and a credible comparison with what would have happened without the deployment.
Other benefits need their own evidence. Earlier contract completion may support revenue or supplier availability, but the value depends on the business event that followed. Fewer missed obligations may reduce exposure, but a short pilot with no incidents does not establish an annual loss reduction.
Report modeled risk benefits with their assumptions and uncertainty. Keep them separate from realized spending reductions, and check for overlap before combining benefits.
A worked example: faster work with a negative cash return
Consider an illustrative supplier-approval pilot processing 1,000 comparable cases per month. Assume it meets the agreed completion and control requirements.
Average staff handling time falls from 30 minutes to 12 minutes per case, including review and correction. That releases:
1,000 cases × 18 minutes ÷ 60 = 300 staff hours per month.
At an assumed loaded rate of $60 per hour, the capacity value is $18,000 per month. Those hours include work previously covered by overtime and external support.
Suppose the organization documents a $6,000 monthly reduction in that spending. The rest of the released capacity remains within the team. Incremental platform, model, integration-service, and support costs total $6,000 per month, and initial implementation costs $36,000. Assume no other incremental cash costs or financial benefits.
If those monthly results continued for 12 months, the first-year cash scenario would be:
- Realized spending reductions: $6,000 × 12 = $72,000.
- Incremental operating costs: $6,000 × 12 = $72,000.
- Implementation spending: $36,000.
- Total incremental cost: $108,000.
For this simplified cash view:
ROI = (attributable financial benefits − incremental costs) ÷ incremental costs × 100.
($72,000 − $108,000) ÷ $108,000 × 100 = approximately −33%.
The pilot releases capacity, but the scenario has a negative first-year cash return. The $18,000 monthly capacity estimate cannot be added to the $6,000 savings: the spending reduction already reflects part of the same work.
The organization may still value the remaining capacity or shorter queues. It should explain those benefits and decide whether they justify the investment. All figures here are assumptions for an illustration, not survey findings or a forecast for Leah deployments.
Decide what evidence supports expansion
Agree on the expansion decision before the pilot ends. The business owner should define acceptable outcomes, finance should agree on cost and benefit treatment, and the relevant control owners should set the operating requirements.
Expand when the evidence supports the proposed next scope. Check whether quality, review effort, and costs hold across the case types the agent will receive. Test the intended workload against reviewer capacity, connected-system limits, and recovery procedures.
If the result depends on specialists quietly repairing cases, record that contribution and cost. If an update changes the model, tools, or permissions, rerun the relevant checks. Financial performance does not override the action limits established through AI agent governance.
An unresolved result may call for a narrower scope, another measurement period, or a simpler process. Assign an owner and a decision date rather than extending the pilot indefinitely.
Where Leah Maestro fits
Leah Maestro coordinates commercial work across legal, contracting, procurement, and finance. Leah's product page describes execution records, organizational playbooks, and approval before irreversible actions.
In an evaluation, ask to connect those records to your completion criteria and cost data. Follow a representative case through the process, inspect the remaining human work, and establish which benefits finance can verify. The demonstration should make the measurement possible; the return calculation still needs your own evidence.
Read the full IDC study for the measurement gap and the broader findings on enterprise agents. Before expanding a pilot, make sure its owner can explain the completed work, the full cost, and what the organization gained.
Frequently asked questions
How do you calculate AI agent ROI?
Choose a time period and compare attributable financial benefits with incremental deployment costs over that period. Divide the net benefit by incremental cost and multiply by 100. State the baseline, cost boundaries, and whether the calculation uses realized cash benefits or modeled economic value.
Does time saved count as ROI?
Time saved is evidence of released capacity. It becomes a cash benefit when spending changes, such as a documented reduction in overtime or external services. Capacity can also support more work or shorter queues; measure those outcomes separately and avoid counting the same benefit twice.
Which AI agent metrics should a pilot track?
Track verified completion, first-pass acceptance, staff handling time, elapsed cycle time, cost per completed case, control failures, and realized benefits. Include failed attempts, human fallback, and unresolved work. Separate case types so changes in workload do not distort the comparison.
How long should an AI agent pilot run?
Run it long enough to observe the normal workload mix, relevant exceptions, and the costs of ongoing operation. A fixed number of days cannot guarantee representative evidence. Set a review date, identify what remains uncertain, and decide whether more observation would change the expansion decision.
What did IDC find about AI agent ROI measurement?
Page 9 of the September 2026 InfoBrief reports that about 21% of organizations were not yet measuring ROI on AI agent deployments. The study covered 410 enterprise decision-makers. That finding describes measurement practices and does not establish that other respondents achieved positive returns.
