AI automation ROI: how to calculate it, and how long payback takes

To calculate the ROI of AI automation, start with the weekly value: the hours it saves each week, times the number of people who benefit, times their loaded hourly cost. Payback in weeks is the build cost divided by that weekly value. First-year ROI is the first-year value minus the build cost, divided by the build cost. As a rule of thumb, a workflow that saves ten or more hours a week pays back within a few months, and one that saves under two hours a week rarely justifies a custom build.

This guide covers the formula, the inputs and how to estimate them honestly, a worked example, how long payback takes, what to count in each department, and how to check your estimate after launch. If you would rather skip the arithmetic, the free AI automation ROI calculator does it for you.

#The formula

weekly value   = hours saved per week × people × loaded hourly cost
annual value   = weekly value × 50
payback weeks  = build cost ÷ weekly value
12-month ROI   = (first-year value − build cost) ÷ build cost

Fifty weeks, not fifty-two, to allow for holidays. First-year value is a little lower than annual value because of the ramp-up period, which is covered below.

#The five inputs

#1. Hours saved per week

This is the input people most often get wrong, and nearly always in the same direction.

Measure the task as it is done today. For a week or two, count how many times it happens and how long each one takes, including the checking and the fixing. Then take about a third off. An automation rarely removes a task completely. Someone still reviews the output, and the unusual cases still need a person.

If the task takes six hours a week today, plan on saving about four.

#2. Loaded hourly cost

Use the full cost of an hour of the person's time, not their hourly wage. That means salary plus benefits and overhead, divided by the hours they actually work in a year.

A rough shortcut is annual salary divided by 1,500. Someone on $60,000 a year costs about $40 an hour.

Use the cost of the people who do the task. Four hours a week saved for a founder is worth far more than four hours saved for a junior coordinator, and the estimate should reflect who it really is.

#3. People who benefit

Count the people who do this task every week. It is usually fewer than the team's headcount. A "marketing team automation" may touch two of eight people.

#4. Build cost

For a custom build, a single workflow typically costs $3,000 to $15,000 and a program covering several workflows $10,000 to $50,000. Those are approximate market ranges, covered in detail in what AI automation consulting costs. If an existing product does the job, the "build cost" may be a few hours of setup.

#5. Ramp time

A new automation does not deliver its full value on day one. Instructions get tuned, edge cases surface, and people take a few weeks to trust it and stop doing the task by hand as well. A month or two is common for a single workflow.

A simple way to model it: assume half the full value during the ramp. Over a fifty-week year that gives:

first-year value = annual value × (1 − ramp weeks ÷ 100)

A six-week ramp costs about 6 percent of the first year's value. An eight-week ramp costs about 8 percent.

#A worked example

A three-person sales team spends 12 hours a week each researching new leads and entering the details into the CRM.

• Hours saved: 12 hours each today, less a third for review and exceptions, gives 8 hours each.
• People: 3.
• Loaded hourly cost: $65.
• Build cost: $9,000, as an illustration.
• Ramp: 6 weeks.

The arithmetic:

• Weekly value: 8 × 3 × $65 = $1,560
• Annual value: $1,560 × 50 = $78,000
• First-year value: $78,000 × 0.94 = $73,320
• Payback: $9,000 ÷ $1,560 = about 6 weeks at full use, or closer to 9 weeks once the ramp is allowed for
• 12-month ROI: ($73,320 − $9,000) ÷ $9,000 = about 7 times the build cost

Now subtract the running costs. If AI usage and hosting come to $300 a month, that is $3,600 a year, and the first-year return falls to about 6.7 times. Still a clear yes.

Run the same sum for one person saving two hours a week at $40 an hour and the weekly value is $80. A $4,800 build would take more than a year to pay back. That is a job for an existing tool, or for leaving alone.

#How long payback takes

Payback depends on two things: what the saved time is worth each week, and what the build costs. The table shows weeks to break even at full use, before ramp-up and running costs.

To turn that into a date, add the build and the ramp. A single workflow takes about two weeks to build, and the ramp adds a month or two. So a workflow worth $800 a week, built for $4,800, breaks even roughly three months after the project starts: two weeks of building, then about ten weeks in which the saving first ramps up and then pays the build back.

Two things follow from the table:

• Volume matters more than cleverness. The projects that pay back fast are the ones with many hours behind them, not the most impressive ones.
• Below about five hours a week, custom work is hard to justify. Look for a feature in a tool you already have.

#What to count, by department

The formula is the same everywhere. What changes is the unit of work you count.

#Customer service and help desk

Count requests per month, minutes per request, and the share the automation would fully resolve. Sort three months of tickets by type first. Only the repetitive categories, such as order status or password resets, should go into the estimate. Requests that are handed to a person still save a little time if the handover arrives with a summary, but do not count them as resolved. More in AI agents for customer service.

#Sales

Count leads per month and the minutes of research, data entry and follow-up each one takes. This is the measurable part. Faster replies to new leads may also win more business, but leave that out of the estimate unless you can measure it. More in AI agents for sales.

#Marketing

Count the recurring deliverables: the weekly report, the monthly newsletter, the posts cut from each webinar. Hours per deliverable times how often it happens. Remember that editing time stays. More in AI agents for marketing and how to automate reporting.

#Finance and accounting

Count documents per month and minutes per document: invoices, expense claims, payment reminders. Add the hours spent preparing reconciliations and month-end commentary. Approvals stay with people, so do not count that time as saved. More in AI agents for finance and how to automate data entry.

#HR and recruiting

Count repeated questions answered, onboarding checklists run, and interviews scheduled. Be careful with anything that screens candidates or informs decisions about people. Those need human review, and in some places the law requires it, so the time saved is smaller than it first looks.

#IT support

Count tickets per month by type and the minutes each takes. Password resets, access requests and "how do I" questions are the usual candidates. As with customer service, estimate only on the categories that follow a rule.

#Documents and operations

Count documents per month, minutes per document, and what an error costs when it slips through. For orders, claims and applications, fewer errors can be worth as much as the time saved, but only include that if you already track the error rate.

#The running costs to subtract

A build quote is not the whole cost. Take these off the annual value before you decide.

• AI model usage. Paid to the model provider by volume. For most small and mid-sized workflows it is tens to a few hundred dollars a month.
• Hosting and tools. A workflow platform, a small server or database, monitoring.
• Maintenance. A few hours a month of someone's attention, in-house or through a support plan.

We cover these in what it actually costs to run an AI automation in production.

#Five common mistakes

1. Counting the whole task as saved. Take a third off for review and exceptions.
2. Counting the whole team. Count only the people who do the task every week.
3. Quoting the full-year figure for year one. Allow for the ramp.
4. Forgetting running costs. They are small, and they are not zero.
5. Leaning on soft benefits. If the hours alone do not justify it, the project is not ready.

#Checking the estimate after launch

An estimate is only useful if you find out whether it was right.

1. Record the baseline before launch. Volume, minutes per item and error rate.
2. Track the same numbers monthly for the first 90 days, plus the share of items handled with no manual work.
3. Ask the people involved how many hours a week it is saving them. Self-reported time is good enough for this.
4. Compare with the estimate. Within about 30 percent, your method is sound. Further off, look first at the hours saved and the number of people, which are the two inputs most often overstated.
5. Watch the running cost each month against what was projected.

The second estimate you make will be better than the first, and the third better again.

#Getting a scoped estimate

You can do all of this yourself with the calculator and the steps above.

If you would like it done for you, our AI Opportunity Audit is a one-week, $1,500 review. We speak with the people doing the work, review your systems and data, and give you a written report: opportunities ranked by value, with the hours saved, cost and payback estimated for each, and a fixed-price quote for the top recommendation. Where the numbers do not support building something, the report says so and suggests what to do instead.

For a single workflow you already have in mind, book a free 20-minute call and we will work through the estimate with you.

#Bottom line

Three numbers decide it: hours saved a week, the loaded cost of those hours, and the build cost. Measure the first one before you estimate, take a third off, allow for the ramp, subtract the running costs, and check the result after ninety days. If a workflow takes your team ten or more hours a week and follows rules you could write down, the numbers usually work.