How to Calculate Automation ROI Honestly

The standard automation ROI calculation — hours saved multiplied by an hourly rate — usually overstates the gain by a factor of two or more. It does this by assuming the automation handles every case, that no time goes into supervising it, that saved minutes reassemble into usable hours, and that a salaried person's freed time is cash. None of those hold by default.

Here is a calculation that survives contact with reality, worked all the way through with one example.

The four places the naive formula leaks

1. Coverage is never 100%

Automations handle the shape of work they were built for. Real inboxes contain PDFs that are photographs, vendors who changed their invoice layout, and the customer who replies to the confirmation email instead of filling in the form. Whatever fraction of volume falls outside the happy path still gets done by a person.

Worse, exception handling is often slower per item than the original manual process, because someone has to notice the failure, find the record, understand what the automation did or did not do, and finish it. Budget a surcharge, not just a pro-rata share.

2. Automations need watching

Someone checks that the run happened. Someone investigates the error email. Someone replays the failed batch. This is real recurring time and it belongs in the model as an ongoing deduction, separate from the annual maintenance budget. It also carries a platform cost on some meters — Zapier's documentation notes that steps replayed during full Zap runs consume tasks (Zapier, How is task usage measured, accessed September 2026).

3. Fragmented minutes are not whole hours

If you remove four minutes from a task performed fifteen times a day, you have not created an hour of productive capacity. You have created fifteen four-minute gaps between other interruptions. Some of that converts into output; some evaporates. If the work was a single Friday afternoon block, conversion is high. If it was scattered, it is not.

There is no published constant for this, and anyone who quotes you one is making it up. What you can do is state your assumption explicitly, defend it, and show the answer at more than one value.

4. Cash and capacity are different currencies

Freed hours are cash only if you stop paying for them: a contractor's invoice that shrinks, overtime you stop approving, a seasonal hire you do not bring back, a vacancy you decide not to fill. If a salaried employee keeps the same salary, you have bought capacity. Capacity is genuinely valuable, but it does not show up in the bank account, and a business case that quietly treats it as though it does will be wrong when someone checks.

The formula

Annual gross value = (baseline weekly hours − exception hours − oversight hours) × recapture factor × 52 × fully loaded hourly cost

Annual cost = subscription + (maintenance hours × your loaded rate), plus, in year one, (build hours × your loaded rate)

For the loaded hourly cost, use federal data rather than the wage. BLS Employer Costs for Employee Compensation for March 2026 puts total compensation per hour worked at $46.60 for private industry workers overall, and $37.36 in private industry establishments with 1–49 workers — $27.68 wages plus $9.68 benefits (BLS ECEC March 2026; Table 6, by establishment size). Small employers should use the smaller number.

A worked example

Invoice data entry at a 20-person company. The person doing it reports spending six hours a week. The automation reads incoming invoices, extracts the fields, and writes them to the accounting system.

Assumptions, all stated so they can be argued with:

• The automation handles 75% of invoice volume end to end.
• The remaining 25% takes 1.8 hours a week — a 20% surcharge over the 1.5 hours those invoices used to take, because exceptions cost extra to find and finish.
• Monitoring, error triage and reruns take 0.5 hours a week.
• Recapture factor: 70%. The work was scattered across the day, so not all of the freed time becomes output.
• Loaded labor cost: $37.36 per hour worked.

Now the costs

Subscription: $300 a year, which comfortably covers a single workflow on a mid-market platform — Make's Core plan is $12/month for 10,000 credits as of September 2026 (Make pricing). Build: 20 hours. Maintenance: 18 hours a year. Both valued at $60 an hour, which is what this company would pay to have the work done.

• Year one: $300 + $1,200 + $1,080 = $2,580
• Year two onward: $300 + $1,080 = $1,380

Year-one net: $5,032 − $2,580 = $2,452, a return of about 95% on the year-one spend. Year two: $5,032 − $1,380 = $3,652, about 265%. Payback: $2,580 ÷ ($5,032 ÷ 12) = 6.2 months.

That is a good project. It is also a much less exciting number than the one the naive formula produces, and it is the one that will still be true in twelve months.

What to do about partial time savings

Three workable approaches, in order of preference.

1. Split the volume. Count only the share the automation truly completes without a human, and model the exception share as unchanged work with a surcharge. This is the method used above and it is the most honest.
2. Present a range. Run the model at 60%, 75% and 90% coverage and show all three. A range invites scrutiny of the assumption instead of hiding it inside a point estimate.
3. Count only the hard cases. Where the freed hours are unambiguously cash — a contractor invoice you will stop paying — count those at full value and count everything else at zero. Conservative, easy to defend, and useful when the decision is contested.

Things worth counting that are not hours

Some real value never appears in an hours model. Count these separately, and only if you have a measured baseline:

• Error reduction, valued at the measured cost of the errors you were actually having — not an industry figure.
• Cycle time, where faster matters commercially: quotes out the same day, invoices issued sooner and therefore paid sooner.
• Headroom, where the automation lets you take 30% more orders without adding people. This is often the largest item and the hardest to prove; state it as a scenario, not as a saving.
• Key-person risk, where a documented workflow replaces a process that lives in one person's head.

Four traps

• Double-counting the same hour. Three projects each claiming two hours a week from the same person means that person now works six fewer hours. Track claims against a single roster of hours.
• Counting build time as free. It is the largest line item in year one. If you would not do it for free for someone else, do not book it at zero here.
• Ignoring the decommission case. Some automations should be switched off. Their value is not permanent, and a five-year projection on a workflow tied to a tool you may replace next year is fiction.
• Using the vendor's calculator. Platform ROI calculators are marketing instruments and set coverage, recapture and maintenance to values that flatter the product. Use them to learn which variables the vendor thinks matter, then build your own sheet.

Bottom line

A defensible automation ROI has four haircuts applied before you multiply by a rate: coverage, exception surcharge, oversight, and recapture. Then subtract build and maintenance at a real hourly cost, and label each freed hour as cash or capacity. Projects that survive that treatment are worth doing. Projects that only look good before it are the ones that quietly disappoint a year later.