How to calculate the ROI of AI automation

A vendor presentation says an automation will save 2,000 hours a year. Your finance lead asks the obvious follow-ups: whose hours, at what cost, and what changes on the P&L? If nobody can answer, the project either stalls or gets approved on faith, and neither outcome is good. A credible ROI case is built from your own process data, a complete cost list and assumptions that still hold when someone pushes on them.

What should you measure before calculating AI automation ROI?

Every ROI calculation rests on a baseline: how the process performs today, measured rather than estimated. Without one, the savings figure is a guess presented as a forecast, and you will have nothing to compare against after launch.

Measure these for the specific workflow you plan to automate:

1. Volume. Items per week or month (orders, invoices, requests, applications), including seasonal peaks.
2. Touch time. The minutes someone actively spends on one item, which is different from how long the item waits.
3. Cycle time. Elapsed time from arrival to completion, including waits for approvals or missing information.
4. Error and rework rate. The share of items corrected later, and what each correction costs.
5. Exception rate. The share of items that fall outside the normal path and need judgment.
6. Who does the work. Roles, pay grades and how much of each person's week the process takes.

Pull volume and cycle time from timestamps in your ERP, CRM or ticketing system wherever you can. For touch time, a short time study over two to four weeks, with the people doing the work logging a sample of items, beats a manager's estimate. If the process is seasonal, include a busy period, or the baseline will understate the peak.

How do you calculate the ROI of AI automation?

The core formula is simple. The discipline is in what you allow into each line.

• Monthly net benefit = value of hours saved + value of errors avoided + value of faster cycle time, minus recurring running costs.
• Payback period in months = one-time costs ÷ monthly net benefit.
• Return over a period = total benefit minus total cost over the period, divided by total cost over the period.

Hours saved at loaded cost

Hours saved = monthly volume × (minutes per item today minus minutes per item after automation) ÷ 60.

The after-automation figure is never zero. Someone still reviews exceptions, spot-checks routine items and handles the cases the system routes to them. Estimate it from a pilot or shadow run, not from a vendor's best case.

Value those hours at loaded cost: wages plus benefits, payroll taxes and overhead. The hourly wage alone understates it. The US Bureau of Labor Statistics reports that private industry employer compensation costs averaged $46.89 per hour worked in June 2026, with wages and salaries at $32.82 (70.0 percent) and benefits making up the remaining 30.0 percent (BLS). Use your own payroll figures for the roles involved, but the national split shows why wage-only math comes out too low.

Errors and rework avoided

Price each error by what it costs to fix: the staff time to correct it, plus direct costs such as credit notes, re-deliveries, late payment fees or missed early payment discounts. Count only the error types the automation prevents, and subtract any new ones it introduces, such as a misread field that passes through unnoticed. Do not count rework minutes that are already inside your touch time figure.

Faster cycle time

Cycle time has a dollar value only where speed changes money. Invoicing customers two days sooner brings cash in sooner. Capturing early payment discounts is real money. Answering quote requests faster may lift win rates, but only count it once you can measure the effect. If faster processing does not change cash, revenue or penalties, report it as an operational improvement and keep it out of the dollar figure.

Capacity redeployed versus headcount

This is where most ROI cases break. Saved hours turn into money in only three ways:

• Avoided hiring. Volume grows and the team absorbs it without a hire you had planned.
• Reduced spend. Less overtime, fewer temporary staff or a smaller outsourcing contract.
• Redeployed capacity. People move to work with measurable output, such as collections calls or supplier negotiations.

If none of these will happen, the hours are real but the cash is not. Name the path you expect and agree it with the managers of the team. Gains are also rarely uniform. A study of 5,179 customer support agents found a generative AI assistant raised issues resolved per hour by 14 percent on average, with a 34 percent improvement for novice and low-skilled workers and minimal impact on experienced, highly skilled staff (NBER). Applying one percentage across a whole team can overstate the result.

What belongs on the cost side of the calculation?

Build cost is the line everyone remembers. The recurring lines and your own team's time are the ones that surprise people in year two.

Model usage is usually billed per token, a unit of text that averages roughly four characters of English. Anthropic, for example, publishes separate prices per million input and output tokens, and a 50 percent discount for requests sent through its Batch API, which suits work that does not need an immediate answer (Anthropic). Price usage at your real volume and document length, then double it in your sensitivity test.

Maintenance is not optional. Model providers retire older models on a schedule, and Anthropic commits to at least 60 days' notice before retiring publicly released models (Anthropic). Each change means re-testing, so budget for it. For a fuller breakdown of running costs, see what AI workflow automation costs to run and what it saves.

What does a worked ROI calculation look like?

Illustrative example: Consider a regional distributor whose order desk keys emailed purchase orders into its ERP. Automation would read each order, match the customer and products, and prepare the ERP entry. Clean orders get a quick check, while exceptions such as unknown product codes or price mismatches go to a person. Every figure below is a placeholder chosen to make the arithmetic easy to follow. None of them is a benchmark or a quote for anyone's work.

Cycle time is deliberately missing. Orders do move faster, but unless the distributor invoices sooner or wins more business because of it, the speed has no dollar value here. The simple payback also assumes full benefit from the first month. In practice, the first month or two after launch deliver less while people adjust, so add that ramp to the answer.

How do you test whether the ROI holds up?

A sensitivity analysis changes one or two assumptions at a time to see how far the answer moves. It tells finance which inputs matter and shows you looked for the weak points before they did.

In this example, two inputs drive the result: the exception rate and the share of saved time that turns into money. That tells you where to spend effort before committing. A shadow run on real orders would pin down the exception rate, and a conversation with the order desk manager would settle how saved hours will be used.

Habits that keep assumptions conservative:

• Use the median handling time from your time study, not the fastest person's.
• Take the exception rate from testing on your own documents or requests, then add a margin.
• Count only the share of saved hours you can tie to a specific hiring, overtime or contractor decision.
• Add a ramp period of reduced benefit after launch.
• Double the model usage estimate.
• Agree the decision rule in advance, for example proceeding only if the downside case pays back within a period finance accepts.

What should you leave out of an AI automation ROI calculation?

Some numbers make a business case look stronger while making it less credible. Leave these out, or report them separately:

• Hours with no change in spending or output. Freed time without a plan is slack, not savings.
• Vendor benchmark percentages. A reduction measured on someone else's process is not evidence about yours.
• Process fixes that need no AI. If removing a redundant approval saves an hour, credit the process change, not the software.
• Double counting. Rework minutes already inside touch time, or the same speed gain valued as both labor and cash.
• Revenue gains without a tested link. Put a number on them only once you can measure the effect.
• Soft benefits priced in dollars. Better morale and fewer late nights matter, so describe them in words.

Leaving these out rarely sinks a good project. It does stop a weak one from being approved on numbers nobody can defend later.

How do you present AI automation ROI to finance?

Finance teams do not need the technology explained. They need to see where each number came from and what happens if it is wrong. A one-page summary usually covers it:

1. The process and its baseline. Volume, touch time, error rate and cycle time, with the data source and the dates measured.
2. The assumptions table. Each input, its value, where it came from and who owns it.
3. Costs split by type. One-time and recurring, including internal time and support.
4. Three scenarios. Expected, conservative and downside, with net monthly benefit and payback for each.
5. How cash is realized. Avoided hires, reduced overtime or contractor spend, or redeployed capacity, named specifically.
6. The measurement plan. The same baseline metrics tracked after launch, with reviews at 30, 60 and 90 days.
7. Stop conditions. The results that would pause or end the project.

Your finance team may also apply its own discount rate to calculate net present value, so give them monthly cash flows rather than only a payback figure. To run your own numbers before that conversation, use the AI automation ROI calculator. If you have not chosen the workflow yet, start with which business processes to automate first.

How Kastling approaches AI automation ROI

We start with the work, not the model. A discovery call covers the process, who owns it and whether a business case is likely to exist at all. For AI and operations work, a separately scoped paid audit is common: we interview the people who do the work, review the systems and data involved, and establish the baseline the ROI depends on. Not every project needs one.

Before development begins, the proposal sets out the measures of success and how they will be tracked, so the business case and the acceptance criteria use the same numbers. We test against real business scenarios, keep a named person in control of consequential decisions, and involve the people who will use the solution, because adoption is where projected savings are won or lost. When the numbers do not support automating a process, we say so. Read more about our AI Integration & Automation service.