The ROI of AI automation: how to calculate it, step by step

The ROI of AI automation is the money a workflow costs you today, minus what it costs after automation (including the human review that remains and the cost of running the system), compared with what you spent to build it. To get a number you can defend, measure the current workflow first, estimate conservatively how much of it can actually be automated, and include every running cost. The method below takes you from a blank page to a payback period, with a worked example.

#What you are measuring

An automation business case has three parts:

• Benefit: staff time no longer spent on the work, errors and rework avoided, and sometimes money captured, such as early-payment discounts or fewer late fees.
• One-time cost: discovery, build, integration, testing, rollout, and your own team's time during the project.
• Running cost: model usage, hosting, monitoring and maintenance, plus the people who still review, approve and handle exceptions.

Most inflated estimates come from counting the benefit in full and the running cost not at all.

#The method, step by step

1. Pick one workflow. Choose steady volume, repetitive steps, digital inputs and clear rules for common cases: entering supplier invoices into the ERP, triaging support tickets, processing onboarding requests. Give each workflow its own business case.
2. Measure the baseline. Record monthly volume, handling time per item, fully loaded hourly cost of the people doing it, error or rework rate, the cost of a typical error, and cycle time. Time a real sample of 30 to 50 items.
3. Split the volume into three groups. Straight-through items that need no human touch. Assisted items where the system does the work and a person checks it. Manual items, the exceptions that stay with people. Base the split on real items.
4. Estimate the human time that remains. Review time per assisted item, full handling time for manual items, and a few hours a month for someone to oversee the system, answer its escalations and approve sensitive actions.
5. Estimate the remaining errors. Automation reduces some errors and can introduce new ones. Use a reduced error rate only if review gates and validation rules justify it.
6. Count the one-time build cost. Vendor or contractor fees, plus internal time from the workflow owner, subject-matter experts and IT for access and security review.
7. Count the monthly running cost. Model usage (priced per token or per document), hosting, software licenses, and monitoring and maintenance, whether managed by a partner or by your own engineers.
8. Model the ramp-up. Rollout is staged, so assume partial savings for the first few months.
9. Calculate payback and ROI. Use the formulas below, then run a low, base and high case.

#The formulas

Baseline monthly cost   = (Volume × Minutes per item ÷ 60 × Hourly cost)
                          + (Errors per month × Cost per error)

Remaining monthly cost  = (Assisted items × Review minutes ÷ 60 × Hourly cost)
                          + (Manual items × Minutes per item ÷ 60 × Hourly cost)
                          + (Oversight hours × Hourly cost)
                          + (Remaining errors × Cost per error)

Gross monthly saving    = Baseline monthly cost − Remaining monthly cost
Net monthly saving      = Gross monthly saving − Monthly running cost

Payback (months)        = Months until cumulative net saving ≥ Build cost
First-year ROI          = (First-year gross saving − Build cost − 12 × Monthly running cost)
                          ÷ (Build cost + 12 × Monthly running cost)

#A worked example (illustrative)

The numbers below are made up to show the arithmetic. They are not benchmarks, and your own will differ.

An accounts payable team processes 2,000 supplier invoices a month. Each takes 6 minutes to key into the ERP, match to a purchase order and route for approval. The loaded cost of the team is $45 an hour. About 2% of invoices need rework, at an average cost of $25 each.

Baseline: 2,000 × 6 ÷ 60 = 200 hours, × $45 = $9,000. Plus 40 errors × $25 = $1,000. Baseline monthly cost: $10,000.

After automation, based on a sample of real invoices, the split looks like this:

Remaining labor: 45 hours × $45 = $2,025. Errors fall to 1%: 20 × $25 = $500. Remaining monthly cost: $2,525. Gross monthly saving: $7,475.

Running cost: $200 model usage, $300 hosting, $2,000 monitoring and maintenance. Total: $2,500 a month. Net monthly saving at full rollout: $4,975.

Build cost: $40,000, including internal time.

Ramp-up: 40% of the gross saving in month one, 70% in month two, full saving from month three. Running cost applies from month one.

• Month 1 net: $2,990 − $2,500 = $490
• Month 2 net: $5,233 − $2,500 = $2,733
• Months 3 to 12: $4,975 each

Cumulative net saving passes $40,000 during month 10. That is the payback period.

First-year ROI: gross saving over 12 months is $82,973. Total cost is $40,000 + $30,000 = $70,000. ROI = ($82,973 − $70,000) ÷ $70,000 = about 19%.

Second year, with no build cost: $89,700 gross saving against $30,000 running cost, a net of $59,700.

Two honest notes on this example. First, a modest first-year ROI is normal. Most of the return arrives in year two and after. Second, the 155 hours freed each month are only a cash saving if the work is redeployed, overtime or temp spend stops, or a planned hire is avoided as volume grows. If nothing changes in how the team is staffed or used, the saving is capacity, which is still valuable but should be labeled as such.

Key point: An AI automation business case is only as good as its baseline. Measured handling times, a realistic straight-through rate and a full running cost will give you a smaller number than a vendor slide, and one you can actually defend to your CFO.

#Benefits to track separately

Some benefits are real but hard to price. Keep them out of the headline ROI unless you can measure them, and list them alongside it:

• Shorter cycle time, such as invoices approved in one day instead of five
• Early-payment discounts captured or late fees avoided
• Faster first response on support tickets
• A complete audit trail of every step, approval and change
• Capacity to absorb growth without hiring at the same rate

#How to validate the estimate before you commit

A spreadsheet estimate is a hypothesis. Two steps turn it into evidence.

A workflow audit. Over one to two weeks, map every step and exception in the workflow, time a real sample, and test a prototype on a few hundred historical items to measure the actual straight-through and review rates. The output is a baseline you trust, a realistic automation split, and a fixed build price, so the two biggest unknowns in the formula are replaced with measured figures. You can see how this fits into a project on our how we work page.

A pilot. Take one workflow to production with success metrics agreed in advance: straight-through rate, review time per item, error rate and cycle time. Run it in shadow mode first, alongside the existing process, then on live volume for several weeks. Compare the results with the baseline and the business case. Expand only if the pilot meets the agreed thresholds.

For finance teams, invoice processing and reconciliation are common first pilots because the volume and error costs are easy to measure. Our finance operations page shows what those systems include.

#Mistakes to avoid

• Using estimated handling times. Ask people how long a task takes and the answer is usually a guess. Time a sample.
• Assuming everything can be automated. Exceptions, poor-quality inputs and judgment calls remain. Plan for them.
• Counting freed hours as cash. Hours become savings only when staffing, overtime or hiring plans change.
• Leaving out running costs. Model usage, hosting, monitoring and maintenance continue every month for the life of the system.
• Choosing a low-volume workflow. A task that takes ten hours a month rarely justifies a custom build, however tedious it is.
• Forgetting the cost of new errors. An automated mistake can repeat at scale. Validation rules, approval gates for sensitive actions and an audit trail are part of the cost and part of the protection.

Automation is not always worth it. If a workflow is low volume, changes every month or depends on judgment that is hard to write down, the honest result of this exercise may be "not yet." That is a useful answer too.

If you want a measured estimate for one of your workflows, book a strategy call. We will look at the workflow, the volumes and the systems involved, and tell you whether an audit is worth doing. You can also see how we approach workflow automation.

#Frequently asked questions

#What is a good ROI for AI automation?

There is no universal figure. It depends on volume, labor cost, error cost and how much of the work can run straight through. A useful test is payback: many teams want the build cost recovered within the first year of running, with clear savings in year two.

#How long does AI automation take to pay back?

It varies with the workflow. In the illustrative example above, payback falls in month 10 after go-live, because of ramp-up and running costs. High-volume workflows with expensive errors pay back faster, and low-volume ones may never pay back.

#Should we include employee salaries as savings?

Only when the freed time changes what you spend: avoided hires as volume grows, reduced overtime or temporary staff, or people moved to work you would otherwise pay for. Otherwise, report it as capacity gained, separate from cash savings.

#What running costs do people usually forget?

Monitoring and maintenance are the most common omissions, followed by human time for reviews and exceptions. Keeping a system accurate as inputs change usually costs more than model usage.

#How do we check the estimate before spending on a full build?

Run a short audit that times a real sample and tests a prototype on historical items, then a pilot on one workflow with agreed success metrics. That replaces the assumptions in your business case with measured numbers.