Most AI agents ROI content stops at a vendor claim and a big percentage. A CFO needs something else: a model with named inputs, stated assumptions and a result that survives a stress test. This article gives you that model, built for teams of 20 to 200 people rather than the enterprise.
The model uses four levers: hours saved, headcount avoided, error rate and speed to output. It subtracts the full cost of running agents, including the human review time that most business cases leave out. It then runs the numbers three ways: base case, stress case and upside case.
Key Takeaways
- Measure AI agents ROI on four levers: hours saved, headcount avoided, error rate and speed to output.
- Never count the same hour twice. A task pool is valued as hours saved or as headcount avoided, not both.
- Net the cost side honestly: platform and model usage, set-up, human review and owner time.
- Keep speed to output out of the base case and show it as upside until you have your own conversion data.
- Record a baseline before launch. Without one, no ROI figure holds up in a board meeting.
Why most AI ROI numbers do not survive a finance review
Three problems show up again and again.
The numbers come from someone else's business. A vendor case study reports hours saved at a company with ten times your volume and a different process. Your CFO cannot trace it to your payroll or your ledger.
Freed time is treated as cash. An hour saved is not an hour that produces revenue or removes cost. It turns into value only when someone redeploys it, or when a hire does not happen.
The cost side is incomplete. Licenses and model usage are easy to list. Review time, exception handling, set-up and the owner's attention are not, and they can be a large share of the real cost.
The market has noticed. Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls. Two of those three causes are measurement problems. A model that makes cost and value explicit addresses both.
The AI agents ROI framework: four levers, one cost line
The framework has one rule: value only what you can trace to a record you already keep. Payroll, the CRM, the ticket system and the ledger all count. A feeling that the team is faster does not.
Lever 1: hours saved
Hours saved is the most common lever and the one most often overstated. Calculate it in three steps.
- Gross hours handled. Count the hours your team spent on the task before agents, taken from time logs or a short time study.
- Subtract review and exception time. Agents still need an approver. Under a Requires approval setting, a person checks each output. Under Notify only or Autonomous, someone still handles exceptions. Subtract those hours.
- Apply a redeployment rate. Decide what share of the net hours becomes productive work. Start at 50%. Raise it only when you can name what the team does with the time.
The formula:
Hours value = (gross hours − review hours) × redeployment rate × loaded hourly cost
Lever 2: headcount avoided
Headcount avoided applies when agents absorb work that would otherwise have justified a hire. It is a larger number than hours saved, so it needs a stricter test.
- The hire must have been planned and funded, not merely wished for.
- You value it for the months the hire is actually deferred, not for the whole year.
- You apply a confidence factor, because many planned hires get canceled or delayed anyway.
This lever does not mean replacing people. It means repeated production and admin work goes to agents, so the next hire can be a role that needs judgment. If you count a pool of tasks here, do not count it again under Lever 1.
Headcount value = loaded annual cost of the role × (months deferred ÷ 12) × confidence
Lever 3: error rate
Errors cost money twice: once to fix and once downstream, in a late invoice, a wrong order or a lost customer. Measure it from a sample.
- Pull a sample of the task output from before agents, for example 200 records.
- Count the errors and set a rate.
- Do the same on agent output after review.
- Cost one error. Start with fix time only. Add downstream cost later, when you can document it.
Error value = (baseline error rate − new error rate) × volume × cost per error
Do not assume agents reduce errors. For data entry, matching and follow-up work with clear rules, they often do. For judgment-heavy work, the error rate may rise unless a person approves the output. Measure it.
Lever 4: speed to output
Speed is the lever with the biggest upside and the weakest evidence at first. Faster lead response, faster invoice follow-up and faster report delivery all have a value. That value depends on your conversion rates, win rates and deal sizes.
Speed value = affected volume × lift in conversion × value per conversion × gross margin
Treat this as an upside case. Show it to the board, but keep it out of the base case until you have two or three months of your own data.
The cost line
Every ROI figure needs a full cost line:
- Platform and model usage. On Harnyss, model usage runs on your own provider keys, so your provider bills it to you directly. Self-serve plans use credits, at one credit per agent action.
- Set-up. Mapping the process, connecting systems and recording the baseline.
- Owner time. The monthly time of the person who reviews results and sets direction.
- Review time. Already netted in Lever 1. Do not subtract it twice.
A worked example: a 60-person B2B software company
The numbers below are illustrative assumptions, not Harnyss results. Replace each one with your own before you present anything.
The situation. The company has revenue operations, finance and marketing admin work that nobody owns full time. It had planned to hire a revenue operations coordinator at a $75,000 salary.
Step 1: loaded cost
Salary is not cost. According to the Bureau of Labor Statistics, benefits were 30.1% of employer compensation costs in private industry in March 2026, so wages were 69.9%. A rough loaded cost for the hire is:
$75,000 ÷ 0.699 ≈ $107,000 per year, or about $8,940 per month.
Divided by 2,080 working hours, that is about $52 per hour. Use your own payroll figures where you have them. The BLS figure is a national average and your benefits mix will differ.
Step 2: annual value, base case
The headcount and hours rows cover different task pools. The coordinator's duties are counted once, under headcount. The scattered admin work is counted once, under hours.
The error row works like this: 4% of 1,500 records is 60 errors per month. At 1.5% it is 22.5. The difference is 37.5 errors, at $40 each, which is $1,500 per month or $18,000 per year.
Step 3: annual cost
The platform and set-up figures are placeholders. Harnyss does not publish managed-services prices, so book an assessment to get a proposal and replace them. For software-only plans, see pricing.
Step 4: the result
- Net benefit: $111,762 − $48,120 = $63,642.
- ROI: $63,642 ÷ $48,120 ≈ 132%.
- Payback: the one-time $15,000 is recovered in about 2.3 months, using straight-line monthly benefit ($9,314) minus monthly running cost ($2,760).
One caution. The headcount benefit runs for nine months because that is how long the hire is deferred. After month nine, you either make the hire or re-base the model. Do not carry that line forward in year two.
Stress-test it before the CFO does
A single number invites a challenge. Show a range instead.
Stress case. Cut the redeployment rate to 25%. Cut hire confidence to 40%. Assume agents deliver only half the error improvement.
- Headcount: $80,460 × 40% = $32,184
- Hours: 120 × 12 × 25% × $52 = $18,720
- Errors: $9,000
- Total value: $59,904. Net benefit: $11,784. ROI: about 24%.
The project still pays back, with little room to spare. That is a useful finding. It tells you which assumptions to protect: the redeployment rate and the review time.
Upside case. Add speed to output. Suppose 200 inbound leads per month, a one-point lift in lead-to-opportunity conversion, a 25% win rate, $12,000 in annual contract value and a 75% gross margin. That is 2 extra opportunities per month, 0.5 extra deals, and $4,500 of first-year margin per month, or $54,000 per year. Haircut it by 50% to $27,000.
- Total value: $111,762 + $27,000 = $138,762. Net benefit: $90,642. ROI: about 188%.
Productivity gains are uneven. In a study of 5,179 customer support agents, the NBER found a 14% average productivity gain from an AI assistant, with a 34% gain for novice and low-skilled workers and minimal impact on experienced workers. Expect the same spread inside your team, and measure by workflow, not by company.
Where to get the baseline
An ROI model is only as good as its starting point. Collect these before launch:
- A time study of the target tasks over two to four weeks.
- Volume per month for each task.
- An error sample, drawn from real records.
- Cycle time from request to output.
- Loaded cost per role, taken from finance, not estimated.
On a managed engagement, this work is part of the process. Harnyss operators map processes, hours, cost and systems in the 30-day assessment, and record a performance baseline during launch. Your team brings one time-consuming process and time with the people who do the work. If you run Harnyss yourself, the savings calculator gives you a first pass at the hours and cost inputs.
Where the model breaks
- Overlapping levers. Counting a pool of tasks under both hours and headcount inflates the result.
- No owner. Unreviewed output creates errors, and errors erase the savings. The platform scores every output and holds weak work, and a person still needs to approve what matters.
- Moving baselines. If volume doubles after launch, compare per-unit cost, not totals.
- Free-floating hours. If nobody can name how freed hours are used, set the redeployment rate near zero.
- Money and customer actions. Where agents touch invoices, payments or customer replies, keep approval on. The Finance and admin department does not move money without approval, and the ROI model should assume the review time that implies.
FAQs
What is a good ROI for AI agents?
There is no universal benchmark worth trusting. A better test is the stress case: if the project still returns more than it costs when your assumptions are cut in half, the business case is sound. Treat any vendor ROI figure without stated assumptions as marketing.
How long until an AI agent project pays back?
It depends on set-up effort and how soon agents reach steady volume. In the example above, payback is about 2.3 months on a straight-line basis. Your timeline will differ. Check at 30, 60 and 90 days against the baseline, not against a forecast.
Should I count headcount avoided as savings?
Only for a funded, planned hire, and only for the months it is deferred. Apply a confidence factor. Do not count the same tasks under hours saved as well.
How do I measure speed to output without good data?
Track cycle time from request to delivery, such as time to first response on a lead. Show the conversion value as upside until you have a few months of your own conversion data.
Does this model work for tools I build myself?
Yes, and it matters more there. Self-built agents add engineering and maintenance time to the cost line. Add those hours at the loaded rate of whoever does the work.
Next step
Pick one repeated process that takes at least one person's time each week, and run the four levers on it using your own payroll and volume numbers. If you want help mapping the hours, cost and systems, book an assessment. It is a 30-minute first call, and you can see how Harnyss works with growth-stage teams on the work with us page.
Sources
- Gartner — Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 (2025)
- National Bureau of Economic Research — Generative AI at Work, Working Paper 31161 (2023)
- U.S. Bureau of Labor Statistics — Employer Costs for Employee Compensation, March 2026 (2026)
