Measuring the ROI of AI agents: hours, dollars, and risk

The ROI of an AI agent is the value of the work it completes, measured in recovered hours, direct dollars, and reduced risk, minus what it costs to run, divided by that cost. Measured honestly, ROI for scoped HR and finance agents is usually decided by two numbers: the automation rate on a high-volume workflow and the fully loaded cost of the human time it replaces.

This article gives the formulas, works through examples using modeled outcomes from Meridian design-partner deployments, lists what to exclude so the number survives a CFO's scrutiny, and lays out a 90-day measurement plan.

Three kinds of return

• Hours: human time no longer spent on a unit of work. Hours are real only if the time is redeployed, absorbs growth, or reduces overtime and contractor spend.
• Dollars: cash effects that do not pass through labor, such as duplicate payments avoided, penalties not incurred, discounts captured, or audit fees reduced.
• Risk: reduced probability or severity of an adverse event, such as a payroll compliance failure or a material misstatement. Risk value is an expected-value calculation and should be reported separately, not blended into hours.

Keep the three in separate columns. Executives trust a return that shows its parts.

The formulas

Define the terms once:

• V = monthly volume of the unit of work
• A = automation rate, the share of units the agent completes without human handling
• T = human minutes per unit before the agent
• t = human minutes per unit after the agent, for the units it handles (review time)
• C = fully loaded hourly cost of the humans doing the work
• K = monthly agent cost: subscription share plus credits consumed times the effective rate

Hours recovered per month = V × A × (T − t) ÷ 60

Labor value per month = Hours recovered × C

Direct dollar value per month = sum of cash effects attributable to the agent, with evidence

Monthly ROI = (Labor value + Direct dollar value − K) ÷ K

Payback period in months = One-time implementation cost ÷ (Labor value + Direct dollar value − K)

Risk value, reported separately = Reduction in event probability × expected event cost

Four worked examples

The examples below use modeled outcomes from design-partner deployments and Meridian's published credit rates as of September 2026.

Worked example 1: Help Desk Agent

An organization with 6,000 employees, 0.8 HR cases per employee per month, and a tier-1 team of 9 generalists.

• V = 4,800 cases per month
• A = 0.70 in budget, below the modeled 0.75
• T = 18 minutes per case, from the baseline study
• t = 2 minutes, for spot checks and escalation review
• C = $42 per hour fully loaded

Hours recovered = 4,800 × 0.70 × 16 ÷ 60 = 896 hours per month

Labor value = 896 × $42 = $37,632 per month

Credits = 3,360 resolved cases × 2 = 6,720 credits per month, inside a Growth plan at $2,499 per month. Allocating the full subscription to this one agent, K = $2,499.

Monthly ROI = ($37,632 − $2,499) ÷ $2,499 = 14.1, or about 1,400%

The number is large because the workflow is high volume and the human cost per case is high relative to two credits. The number is only true if the 896 hours are redeployed or absorb growth; see the exclusions below.

Worked example 2: Audit Agent

A company that models about 900 hours per year saved on audit evidence collection, at $65 per hour for the senior accountants who do it.

Labor value = 900 × $65 = $58,500 per year

Direct dollars: the external audit firm reduces fees by $30,000 because evidence is packaged to its request list. Evidenced by the engagement letter.

Credits: 300 evidence packages per year × 5 = 1,500 credits, a small share of a Growth pool. Allocate $6,000 per year of the subscription for a conservative K.

Annual ROI = ($58,500 + $30,000 − $6,000) ÷ $6,000 = 13.75

Worked example 3: Controls Agent, mostly dollars and risk

One design partner avoided about $283,000 per year in duplicate payments. Duplicates that are caught before payment are direct dollars. Duplicates recovered after payment are also dollars, net of recovery cost. The risk column carries the reduced probability of an undetected fraud scheme; if internal audit estimates a 2% annual probability of a $2 million event and continuous testing halves it, the risk value is $20,000 per year, reported separately.

Worked example 4: Close Agent, hours plus decision value

A 12-person close team shortens the close by 3 days. Hours are the overtime and contractor time removed: for example, 12 people × 3 days × 2 hours of overtime = 72 hours per month at a $75 loaded overtime rate, or $5,400 per month. Most of the value is elsewhere: earlier results for decisions, and the risk column, which carries the reduced probability of a late filing or a post-close adjustment. Report the hours, describe the decision value, and quantify the risk only where finance will stand behind the probability.

Summary across agents

What to exclude

An ROI number is only useful if it survives challenge. Exclude the following, or report them separately with the label "unrealized."

• Hours that are not recaptured. If the tier-1 team is the same size, doing the same work, with more slack, the hours are capacity, not savings. Count them only when headcount is redeployed, attrition is not backfilled, contractors are released, or volume grows without hiring.
• Avoided future hires. Count these only against an approved headcount plan that was withdrawn.
• Satisfaction and experience effects, unless you measure them with a before-and-after survey and can price the outcome, such as reduced turnover.
• Double counting across agents. The Controls Agent clearing exceptions early and the Close Agent shortening the close overlap; attribute each day or hour once.
• Implementation time of your own team. Include it as cost, not as value.
• Model or vendor claims not measured in your environment. Modeled outcomes set the budget; measured outcomes set the ROI.

A 90-day measurement plan

Two habits make the plan credible. First, have finance own the baseline and the readout, not the team that sponsored the agent. Second, publish the assumptions next to the number. An ROI of 6x with a visible method beats an ROI of 14x that nobody can reproduce.

The number to watch after day 90

After the readout, the metric that predicts whether ROI persists is cost per completed unit, tracked monthly in Registry against the human baseline. If it rises, either volume fell, the automation rate slipped after a model change, or approval tiers widened in a way that increased rework. Each is visible in the audit trail, and each is fixable. ROI is not a one-time calculation; it is a monthly line in the blended workforce report.

Outcome figures in this article are modeled outcomes from design-partner deployments and are not guarantees of results.