AI Agent ROI: How to Calculate It and Prove It

AI agent ROI is the measured change in cost from one agent workflow, minus everything the agent costs, divided by what it costs, over a stated period. To calculate it, sample the current process for a baseline, measure the agent against a control group, and count review time, exceptions, and rework.

Most ROI numbers skip at least one step. Below are the method, a worked IT service desk example, and four ways the number gets inflated.

What does AI agent ROI actually measure?

ROI compares the world where the agent runs with the counterfactual world where it does not. You never see the counterfactual, so you estimate it. Most bad ROI numbers estimate it badly.

The gap between felt and booked value is wide. In McKinsey’s State of AI 2026 (1,719 respondents, fielded May 4 to June 8, 2026), 80% said AI has improved their individual productivity. Only 37% attribute at least some EBIT impact to AI, about the same share as a year earlier.

A Danish study linking AI chatbot adoption surveys to administrative labor records found “precise null effects on earnings and recorded hours” at both worker and workplace levels, ruling out effects larger than 2% two years after ChatGPT launched. That study measured chatbots and labor outcomes, not agents or company profits. My inference is that the same mechanism applies to agents: saved time becomes realized labor-cost savings only when someone changes a budget, a contract, or a staffing plan. Extra revenue is a separate path.

Scope every ROI calculation to one workflow, one period (usually 12 months), one written counterfactual, and one owner who signs the baseline. The funding decision around the number is in the CFO guide to AI agent investment.

How to measure the baseline before an AI agent launches

The baseline is what the workflow costs today, per case, with a known error bar. There are two ways to cost a case, and you need both. Touch-time cost is the minutes a person actually spends on the case, times a loaded hourly rate. Absorbed cost is the team’s total monthly cost divided by monthly volume. MetricNet’s Jeff Rumburg defines cost per ticket this way: total monthly operating expense divided by monthly tickets. Absorbed cost runs higher because it includes meetings, training, idle time, and tools. Use touch time for what the agent changes, and absorbed cost only for outside benchmarks. Mixing them is how 10 minutes saved becomes a $40 saving in a deck.

Do not trust the “time worked” field. People fill it in from memory and round. Flag a random sample of cases and have the people working them log start and stop times for two to four weeks. Stratify by category and by person, and avoid weeks with known spikes such as quarter-end.

How many cases? NIST’s engineering statistics handbook gives the bound for estimating a mean at 95% confidence: N ≥ (1.96 / δ)² × σ², where δ is the margin you accept and σ is the standard deviation. With a standard deviation of 11 minutes, you need about 465 cases to know the mean within plus or minus 1 minute, or about 207 for plus or minus 1.5. Take a small first sample to estimate σ.

For the hourly rate, use payroll data, or label a public fallback. BLS reports that private-industry employer costs averaged $46.89 per hour worked in June 2026, with wages at 70.0% of that. BLS puts the May 2025 median wage for computer user support specialists at $61,860 a year. Divided by 2,080 hours and then by 0.70, that is a rough $42.50 an hour loaded, which I use below.

Also record what most baselines skip: rework rate, escalation share, and escalated handle time. Timing matters too. HM Treasury’s guidance on evaluating AI interventions warns that AI systems are often built iteratively and change fast, which makes it hard to set a baseline “quickly or before the roll-out of the intervention begins.” Measure before the pilot, not during it.

The AI agent ROI formula, with every term defined

ROI = (Benefit - Cost) / Cost, over a stated period

Benefit = Labor value + Cash removed + Error cost avoided
  Labor value       = (Baseline human minutes - After human minutes) / 60 x L
  After human minutes = assisted cases x minutes
                      + escalated cases x minutes
                      + agent-completed cases x review share x review minutes
                      + reopened cases x rework minutes

Cost = One-time cost + Run cost x months
  Run cost = model + monitoring and maintenance + tooling + eval runs

Two counting rules. Put human review time in one place only; I put it in after human minutes, where the measurement captures it. And if freed hours are what let you cancel a contract, count them once, as cash removed.

Then report two ROI numbers. Capacity ROI uses labor value. Cash ROI uses only cash removed and error cost avoided. When they differ a lot, the gap is a management decision about the hours. More analysis will not close it.

How do you measure AI agent benefit after launch?

Use the strongest design you can.

Randomized split. Route in-scope cases to the agent or the current process at random, such as by even or odd case ID. Both arms see the same weeks, spikes, and staff, so the difference in human minutes per case is the effect. The HM Treasury guidance notes that the digital nature of AI tools opens the door to randomized trials.

Staggered rollout. Turn the agent on one team at a time and compare each group with those still waiting. Brynjolfsson, Li, and Raymond used it in Generative AI at Work (Quarterly Journal of Economics, 2025), covering 5,172 customer-support agents. Issues resolved per hour rose 15% on average. The average hid the useful part: less experienced workers improved in speed and quality, while the most experienced saw small gains in speed and small declines in quality. Break your results out by case type and team.

Before and after, with a comparison group. If you cannot split, compare your workflow’s change with the change in a similar workflow the agent does not touch (difference-in-differences). Hypothetical: in-scope minutes per case fall from 14.0 to 8.4 after launch. Over the same weeks, a comparison category falls from 16.0 to 14.5 because of a seasonal lull. The estimated effect is (8.4 - 14.0) - (14.5 - 16.0) = -4.1 minutes, not -5.6. About 27% of the naive saving was the season. A plain before-and-after with no comparison is weakest; if it is all you have, compare the same calendar weeks year over year.

In every design, freeze the model version, prompt, and tools during the window, or log each change with a date. Run at least one full business cycle. For weekly numbers after the measurement ends, see three AI agent KPIs a COO should review.

Worked example: AI agent ROI for an IT service desk

Everything here is hypothetical except the BLS wage data and the model price. For triage design, see AI agents for IT service desk triage.

A 4,000-person company’s IT desk gets 3,000 tickets a month in one scope: how-to questions and requests from a pre-approved software catalog. A 465-ticket sample put touch time at 14 minutes. The loaded rate is $42.50. Baseline: 3,000 × 14 = 42,000 minutes, or 700 hours a month.

The agent runs on Claude Sonnet 5.5, which Anthropic prices at $2 per million input tokens, $10 per million output, and $0.10 per million for cache hits. At 6 calls per ticket, 12,000 input and 700 output tokens each, and 60% of input cached, that is $0.104 per ticket, or $0.125 with 20% added for retries and cache writes.

A six-week randomized split measured these outcomes:

Look at the escalated row: 19 minutes, not the old 14. The agent keeps the easy tickets, so what it hands back is harder. Measure this number; never assume it.

One-time cost is $72,000 (build and integration $60,000, baseline study and pilot $8,000, training $4,000). Months 1 and 2 run at half benefit, which removes one month of labor value. Error cost avoided is zero here, because reopen time already carries the rework.

12-month capacity ROI = (11 × $11,857.50 - ($72,000 + 12 × $2,574.11)) / ($72,000 + 12 × $2,574.11) = ($130,433 - $102,889) / $102,889 = 26.8%. Over 24 months, the same math gives 103.9%.

Now the cash view. The 279 hours equal about 1.6 analysts of paid time (173 hours a month each), and cash arrives only in whole-person steps. Say the desk decides not to backfill one analyst who leaves in month 3. Loaded annual cost is $61,860 / 0.70 = $88,371, or $7,364 a month. Nine months of that is $66,279, so year-one cash ROI is -35.6%. Over 24 months, 21 months of savings ($154,650) against $133,779 of cost gives +15.6%.

Sensitivity, holding everything else fixed:

Exception handle time moves ROI nearly as much as resolution rate. Model price barely moves it. For payback timing, see AI agent payback period.

Four ways AI agent ROI gets inflated

Here is the same example, inflated one cumulative step at a time.

Counting freed hours as cash. Freed hours become cash only when a contract, a hire, or overtime goes away. Detection: ask which budget line changes, and in which month. If nobody can say, label it capacity.

Ignoring review time. A reopen often costs more than the original ticket, because someone has to rebuild context. In the example, review and rework are only 26 hours a month, yet dropping them adds 13 points of ROI. Detection: track time to verify per case. If verifying takes as long as doing the task by hand, the agent saves nothing on that case type.

Using vendor accuracy instead of measured accuracy. Vendor numbers come from someone else’s tasks, usually on one try. The τ-bench paper introduced pass^k, the share of tasks an agent completes correctly on every one of k runs. For its May 2026 benchmark submission, Automation Anywhere reports that its own agent scored 74.5% on single runs across 375 tasks in four service domains, and 60.3% when it had to succeed four times out of four. That is a vendor’s self-reported score, not your workflow. Detection: no rate enters the model unless it was measured on your cases. Set thresholds ahead of time with scale-or-stop criteria.

Ignoring exception handling. Do not leave this out. When the agent takes the easy cases, the human queue gets harder per case. Detection: compare escalated handle time with the baseline for the same categories, and track the human queue’s category mix weekly.

How to report AI agent ROI

One page, every field filled. A blank field is a finding. Here is the template, filled with the example.

Mark each number as measured or assumed. If you are still estimating the one-time cost, AI agent development cost breaks down what drives the range.

Bottom Line

AI agent ROI is a measurement of one workflow, not a forecast you defend.

Sample the baseline, measure against a control group, and count review, rework, and exceptions as human time. Report capacity ROI and cash ROI side by side.

Next step: log touch time on a random sample of 200 to 500 cases before anyone builds anything.

FAQ

What is a good ROI for an AI agent?

No reliable public benchmark exists for a single agent workflow. In McKinsey’s 2026 survey, about 6% of respondents were high performers: they attribute at least 5% of EBIT to AI and call its impact significant. Judge your number against your own hurdle rate.

Does switching models change the ROI?

Yes. Anthropic’s pricing page says Claude 4.7 and later models use a tokenizer that produces about 30% more tokens for the same text, and US-only inference costs 1.1x. In the example, 30% more tokens cuts 12-month ROI from 26.8% to 25.1%. An accuracy drop would cost far more, so rerun your test set before any switch.

How long should I measure before reporting AI agent ROI?

My rule is at least four to six weeks, covering a month-end, and longer when volume is low. The HM Treasury guidance treats a randomized trial finished within a few months as fast. I’d expect shorter windows to overstate results, because early weeks get extra attention.

Can I calculate AI agent ROI before building anything?

You can forecast it: a real baseline plus outcome rates from a shadow test on past cases, where the agent runs but takes no action. Replace the forecast with measured rates within 90 days of launch.

Three Ways to Work With TMA

Need an agent built? We deploy production AI agents in your infrastructure. Working pilot. Real data. Measurable ROI. → Schedule Demo

Want to co-build a product? We’re not a dev agency. We’re co-builders. Shared cost. Shared upside. → Partner with Us

Want to join the Guild? Ship pilots, earn bounties, share profit. Community + equity + path to exit. → Become an AI Architect