AI agent ROI is the annual value a workflow returns once the agent is running, minus the annual cost of running it, divided by that cost. The formula is the easy part. The hard part is deciding which of your saved hours are actually money, and which costs the vendor's calculator quietly left out.
Search for a calculator and you will find a page of them, almost all built by companies selling agents. Put in your headcount, get back a number with a lot of zeroes, and give up your email. They are not lying, exactly. They are just counting the generous side of the ledger and skipping the rest. The buyers on the other side of those pages are not convinced, and it shows in what ranks alongside the calculators. Search that query and the discussion threads sitting next to them ask how anyone is measuring ROI on agents "before the API bill eats the profit," and whether agents are delivering real return "or just saving time."
That second question is the whole article. Time saved and money saved are different things, and treating them as the same is the single most common reason an AI business case falls apart in the room where it gets approved.
So this is the version you can defend. It gives you the six inputs, the three cost lines that usually go missing, and a different output than the one vendor calculators produce. Instead of an ROI percentage, it produces your break-even setup budget: the largest one-time fee that still pays for itself inside twelve months. That is the number worth walking into a vendor conversation holding.
What is AI agent ROI?
AI agent ROI measures what one workflow returns after an agent takes over part of it, against what that agent costs to run for a year. It is calculated per workflow, not per company, because an agent handling invoice exceptions and an agent handling inbound lead research have nothing in common on either side of the equation.
The formula, in plain terms:
Annual net = (realized value) minus (annual running cost)
First year ROI = annual net divided by (annual running cost plus setup)
Two things about that. First, "realized value" is doing a lot of work, and the next two sections are about pinning it down honestly. Second, the setup fee sits in the denominator in year one and disappears in year two, which is why a one-year ROI number and a three-year ROI number for the same agent can look like they describe different projects. Always say which one you are quoting.
Calculate this per workflow, and calculate it before you buy. Doing it afterward turns it into a justification exercise, and everyone in the room can tell. If what you are actually evaluating is a whole custom software build rather than a recurring workflow, that is a different unit of analysis and a different calculation.
Why most AI agent ROI calculators give you a number you cannot defend
Vendor ROI calculators inflate in three predictable ways: they convert every saved hour straight into dollars, they count only the subscription as cost, and they assume the agent handles the whole workflow. Fix those three and the number gets smaller, more boring, and far more likely to survive a finance review.
Inflation one: hours saved become money saved. A calculator multiplies hours removed by an hourly rate and calls the product savings. But nothing left your bank account. If nobody is let go, no contractor invoice is cancelled, and no overtime disappears, you did not save money. You recovered capacity. Capacity is genuinely valuable, but it is a different line item, and the person approving your budget knows the difference.
Inflation two: the cost side is one line. The subscription goes in. The model usage, if it is billed separately, does not. Nor do the hours your own team spends reviewing the agent's work, answering its escalations, and improving it. Those hours are real and they recur.
Inflation three: the agent does the whole job. It almost never does. Agents take the routine majority and hand you the exceptions, which are the hardest and slowest part of the work. If the agent removes nine hours a week and creates three hours of review, your net is six, not nine.
There is a fourth thing worth naming, which vendor calculators cannot know: some of the workflows people put into them should not have an agent at all. More on that near the end.
The six inputs you need before you can calculate AI agent ROI
You need six numbers, and five of them come from inside your own business. Gather them before you talk to anyone selling you something, because once you are in a demo the numbers start arriving pre-filled.
| # | Input | Where to get it | Common mistake |
|---|---|---|---|
| 1 | Hours the workflow consumes today, per week | Ask the people doing it, for two weeks, in writing | Asking a manager instead of the person with the inbox open |
| 2 | Fully loaded hourly cost of those people | Payroll, grossed up for taxes and benefits | Using base salary divided by 2,080 |
| 3 | Hours the agent plausibly takes over | Scope conversation, expressed as a share of the routine cases | Assuming 100 percent |
| 4 | New hours the agent creates | Review, approvals, escalations, corrections | Recording zero |
| 5 | Hard costs that actually stop | Contractor invoices, overtime, tools you can cancel, fees you currently pay | Counting salaries nobody is cutting |
| 6 | Annual running cost of the agent | Plan price, model usage if billed separately, your own ownership time | Counting only the subscription |
On input two, use the fully loaded figure. Wages are not the whole cost of an hour. As of June 2026, employer compensation costs for private industry workers averaged $46.89 per hour worked, of which wages and salaries were $32.82, or 70.0 percent, with benefits making up the remaining 30.0 percent, according to the Bureau of Labor Statistics Employer Costs for Employee Compensation release. A workable gross-up is to divide your salary-derived hourly rate by 0.70, then sanity check it against your own payroll, because your benefits load and your paid-versus-worked hours will differ from the national average.
A $70,000 salary grosses up to roughly $100,000 fully loaded, which is about $48 an hour across a 2,080 hour year. That is the number that belongs in the worksheet, not $33.65.
How to value the hours without overstating them
Multiply the net hours the agent removes by the fully loaded hourly cost, then multiply again by a realization rate: the share of those recovered hours that actually turns into money or into work that earns. Setting that rate honestly is the difference between a business case and a brochure.
Net hours first. Take the hours the agent takes over, subtract the new hours it creates, and use the remainder. Nine hours taken over and three hours of review created is a net of six.
Then the realization rate. Ask one question about those six hours: what specifically happens to them?
- A cost line actually falls. You stop renewing a contractor, overtime drops, an open role stays unfilled. Realization is close to 100 percent, and you should be able to name the line.
- The hours go to work that earns or protects money. The person now follows up on quotes that were going cold, or clears a backlog that was costing you renewals. Realization is real but partial. Pick a number you can argue for, and be conservative.
- Nothing in particular happens. The work is less annoying and the week fills back up. Realization is zero. This is not a failure and it is not nothing, but it is not ROI, and calling it ROI is how business cases lose credibility.
Most real situations are a blend. Write down the blend and the reasoning next to it. When someone challenges your number six months later, the reasoning is what saves you.
Count avoided error costs too, but only from a number you already track. If you paid $1,900 in late payment penalties last year and you expect the agent to remove most of them, that is a defensible line. If you "expect fewer mistakes," that is a hope, and hopes do not belong in a spreadsheet.
What to count as cost: the three lines vendor calculators skip
Annual running cost has three parts, and most calculators show you one. Add all three before you compare anything, because the two missing lines can be larger than the subscription itself.
The recurring service cost. This one is easy when it is published. VantaSoft's plans are Launch at $200/mo, Team at $375/mo, Scale at $700/mo, and Enterprise from $1,500/mo. Discovery comes first. A one-time setup fee based on the scope of your agents, workflows, and integrations applies. Setup and monthly billing begin after you approve your Order Form. Enterprise pricing is customer-specific.
Model usage, if your provider bills it separately. This is the line behind that "before the API bill eats the profit" question, and it is the one buyers get surprised by. It is usage based, so it scales with how much work you send the agent, and it is not fixed the way a seat license is. Ask directly whether model usage is inside your quoted monthly figure or on top of it, and ask what happens when volume doubles. Anthropic's engineering team is blunt about the underlying economics: agentic systems "often trade latency and cost for better task performance," and "the autonomous nature of agents means higher costs, and the potential for compounding errors." That is a design tradeoff, not a billing surprise, as long as you priced it in.
Your own ownership time. Somebody internally owns this agent. They answer its questions, approve the actions that need approval, notice when something drifts, and ask for changes. Budget an hour a week at that person's fully loaded rate and you will be closer than budgeting zero. If you have not decided which actions need a person in the loop, that decision changes this line, and we wrote a method for placing it action by action.
Setup sits outside all three. It is a one-time cost, it is scoped to your workflows and integrations, and it is the number the rest of this worksheet is designed to help you evaluate rather than guess at. For what drives the total cost of an agent in the first place, we broke the cost layers down separately.
The AI agent ROI formula, and the better number it produces
Run the annual net first, then invert the question. Instead of asking what ROI a given price delivers, ask what price your workflow can support. That gives you a break-even setup budget, which is a number you can carry into a vendor conversation instead of one you receive from it.
The four steps:
- Realized value = (net hours removed x fully loaded hourly cost x realization rate) + hard costs that actually stop + measured error costs avoided
- Annual running cost = recurring service + model usage if separate + your ownership time
- Annual net = realized value minus annual running cost
- Break-even setup budget = the annual net, if you want the whole thing paid back inside twelve months
Then, once you have a real quote, payback in months is the setup fee divided by one twelfth of the annual net.
Why this direction is more useful: it stops the conversation from being about whether a price feels high. A setup fee is not high or low in the abstract. It is either inside or outside what this specific workflow can pay for, and the difference is arithmetic. It also protects you from the opposite mistake, which is buying something cheap for a workflow that was never going to return anything.
A worked example: invoice exception review
The numbers below are invented to show the mechanics. They are not a VantaSoft customer result, an estimate for your business, or a claim about typical outcomes. Use your own inputs.
A 40 person company handles vendor invoice exceptions by hand. Two people spend a combined 14 hours a week finding mismatches, chasing vendors, and correcting records.
Value side
| Line | Working | Amount |
|---|---|---|
| Hours the workflow takes today | 14/week x 52 | 728 hours/year |
| Hours the agent takes over | 9/week x 52 | 468 hours/year |
| New review hours it creates | 3/week x 52 | 156 hours/year |
| Net hours removed | 468 minus 156 | 312 hours/year |
| Fully loaded hourly cost | $70,000 salary, grossed up, over 2,080 hours | $48/hour |
| Gross value of recovered hours | 312 x $48 | $14,976 |
| Realization rate | Half the hours go to a backlog that protects renewals, half evaporate | 50% |
| Recovered capacity, realized | $14,976 x 0.50 | $7,488 |
| Hard cost that actually stops | An overflow AP contractor, not renewed | $6,000 |
| Measured error cost avoided | $1,900 in late payment fees last year, 60% expected to go | $1,140 |
| Realized value | $14,628 |
Cost side
| Line | Working | Amount |
|---|---|---|
| Recurring service | Launch plan, $200/mo | $2,400 |
| Model usage | Treated as included in the monthly figure for this example; confirm yours | $0 |
| Internal ownership time | 1 hour/week at a $72 fully loaded manager rate | $3,744 |
| Annual running cost | $6,144 |
Result
- Annual net: $14,628 minus $6,144 = $8,484
- Break-even setup budget for a twelve month payback: $8,484
- If the quoted setup came in at $4,000, payback is $4,000 divided by $707 a month, or about 5.7 months
Now the part that matters more than the answer. Change one input, the realization rate, and watch what happens.
| Realization rate | What it assumes | Realized value | Annual net | Break-even setup budget |
|---|---|---|---|---|
| 0% | The hours free up and the week refills. Nothing measurable changes. | $7,140 | $996 | $996 |
| 50% | Half the recovered hours go to work that protects revenue. | $14,628 | $8,484 | $8,484 |
| 100% | Every recovered hour is redeployed into work that earns or a cost that falls. | $22,116 | $15,972 | $15,972 |
One assumption moves the defensible setup budget from about a thousand dollars to about sixteen thousand. That spread is why a vendor calculator that silently picks 100 percent for you is not useful. It is also why the most valuable thing in your business case is not the total at the bottom. It is the sentence explaining what happens to the recovered hours.
How long does it take to see ROI from an AI agent?
Payback is setup divided by monthly net, and nothing starts counting until the agent is actually handling live work. In the example above, a $4,000 setup against $707 of monthly net pays back in about six months. What that arithmetic hides is the ramp, and the ramp is where schedules slip.
Three things stretch it in practice. The agent starts narrow and widens as you trust it, so early months deliver less than steady state. Access to your systems has to be approved and scoped before anything runs, and that is usually an internal calendar problem, not a vendor one. And the first weeks carry more review time than later ones by design, which shows up on the cost side of your own worksheet.
Build your model on steady state, then ask what the first ninety days look like separately. If a provider will not distinguish the two, that is information. The sequencing of the work itself is a separate question, and we laid out what the rollout actually asks of you.
When the honest answer is not to buy an agent
Some workflows return a good number on paper and should still be solved with something simpler. Before you run the math, check that the work genuinely needs judgment, because when the rules are genuinely writable, a rules engine that costs a fraction as much will usually win on ROI.
OpenAI's practical guide to building agents names three conditions where agents earn their keep: workflows "involving nuanced judgment, exceptions, or decision-making," systems that have become "unwieldy due to extensive and intricate rulesets, making updates costly or error-prone," and work with "heavy reliance on unstructured data." Its instruction is worth quoting exactly: "Before committing to building an agent, validate that your use case can meet these criteria clearly. Otherwise, a deterministic solution may suffice."
Anthropic reaches the same place from the engineering side, recommending teams find "the simplest solution possible, and only increasing complexity when needed," and noting that "this might mean not building agentic systems at all."
So run this gate first:
- If the rules are written down and stable, you want an automation, not an agent.
- If the inputs are uniform and structured, you want an automation, not an agent.
- If the work is high volume but low judgment, you want an automation on the majority and possibly an agent on the exception path only.
- If nobody can name who owns the workflow, fix that before you price anything.
If you have not settled on which workflow to put through this in the first place, start from the roles and use cases that tend to qualify and pick the one with the clearest owner.
We would rather tell you this before a scoping call than after one. If you are still deciding between an automation and an agent, the comparison is here, and if you are weighing building it yourself against buying it managed, that changes the cost side substantially.
The worksheet
Copy this into a spreadsheet, fill the left column, and keep the notes. The notes are the part you will need when someone asks where a number came from.
| Line | Your number | Note to your future self |
|---|---|---|
| A. Hours the workflow takes today, per week | Who measured it, and over how long | |
| B. Hours the agent takes over, per week | What share of cases, and which ones it does not touch | |
| C. New hours the agent creates, per week | Review, approvals, escalations, corrections | |
| D. Net weekly hours removed (B minus C) | ||
| E. Fully loaded hourly cost | Salary grossed up for taxes and benefits | |
| F. Gross annual value of recovered hours (D x 52 x E) | ||
| G. Realization rate | Exactly what happens to the recovered hours | |
| H. Realized capacity value (F x G) | ||
| I. Hard costs that actually stop, per year | Name the invoice or the line item | |
| J. Measured error costs avoided, per year | From a number you already track | |
| K. Realized value (H plus I plus J) | ||
| L. Recurring service cost, per year | ||
| M. Model usage, per year | Confirm whether it is inside L or on top of it | |
| N. Your ownership time, per year | Hours per week x their fully loaded rate | |
| O. Annual running cost (L plus M plus N) | ||
| P. Annual net (K minus O) | ||
| Q. Break-even setup budget (P, for a 12 month payback) | The number you bring to the conversation |
If P is negative, the workflow does not support an agent at current assumptions. That is a useful result, and it took you an hour instead of a quarter.
Frequently asked questions
How do you calculate AI ROI in general, not just for agents? The same structure holds: realized value minus running cost, divided by cost. What changes for agents is that the cost side has a usage-based component and an internal ownership component that a fixed software license does not, and the value side depends more heavily on how much of the workflow the system can actually take.
Is there a difference between generative AI ROI and agentic AI ROI? Yes, on the cost side. A tool that drafts something for a person to use is a fixed cost with a fixed output. An agent takes actions across your systems, so its cost scales with volume and its value depends on how many steps it completes without a person. That makes per workflow measurement more important, not less.
Should we count time saved as ROI? Only the portion that becomes money or becomes work that earns. Report recovered hours as a separate line from realized savings. Anyone reviewing your case will respect the distinction, and the ones who do not will find it themselves later.
Do we include the setup fee in year one? Include it in the year one ROI denominator and exclude it from year two onward. Quote both. A three year view is the fairer picture of a recurring service, and a one year view is what most approvals actually need.
What if our workflow is not measured at all today? Measure it for two weeks before you do anything else. A rough two week log from the person actually doing the work beats a precise estimate from someone who is not. Without input A, everything downstream is decoration.
What if the number comes out small but the pain is large? Say so plainly. Some workflows are worth fixing for risk, retention, or customer experience reasons that the ROI line does not capture. Make that argument on its own terms instead of inflating the realization rate until the spreadsheet agrees with you.
Bring your break-even number to the conversation
Fill in the worksheet, land on line Q, and you will know something most buyers walk into a scoping call without: what this specific workflow can actually support. It makes the conversation shorter and the scope more honest, because the discussion becomes what fits inside that number rather than whether the number feels right.
That is the conversation we would rather have. Start Automating. It starts with 30 minutes on your workflows, and bringing your own arithmetic is welcome.
