How to calculate the cost of a manual process: a four-part formula, a loaded-rate method, a two-week tally, and the three costs owners leave out.
The office manager at a 25-person firm knows invoice reconciliation eats her coordinator's mornings. She tells the owner it is "probably a day a week." The owner hears a day a week, does salary times one fifth in his head, and decides it is not worth fixing. Both of them are guessing, and that kind of guess usually lands well under the counted figure because it uses the wage instead of the loaded rate and skips half the touches. Knowing how to calculate the cost of a manual process is what turns "it takes a while" into a number the owner can put next to a quote. This article gives the formula, the method for counting, and the three lines almost every owner leaves out.
Quick Answer: The annual cost of a manual process is the loaded hourly rate of the person doing it, times touches per week, times minutes per touch, times working weeks, plus rework, plus the cost of delay. Add three lines owners forget: interruptions, someone checking the work, and cover when the person is off.
You calculate the cost of a manual process by pricing every touch, not the task. Count how many times a week someone picks the process up, time a sample of those touches, multiply by the loaded hourly rate, and annualise. Then add the cost of redoing work that failed and the cost of the work being late.
The formula has four terms. Most pages on this topic stop after the first one.
Annual cost = direct labour + rework + delay + hidden lines
where:
Two words in that formula do the work. "Touches" instead of "tasks": an invoice is one task but it gets picked up four or five times (received, coded, matched, approved, posted), by different people at different rates. If you price it as one 10-minute task you miss the approval chase and the second person's time. "Loaded" instead of "salary": what an employee costs per hour is well above the hourly wage, and the next section shows how far above.
Working weeks is 50 for a full-time role in most firms, not 52. If the process runs monthly rather than weekly, count touches per month and multiply by 12 instead.
A loaded hourly rate is what an hour of an employee's time costs the business, not what it pays them. It includes payroll taxes, health and retirement contributions, paid leave and insurance. The quick method: divide annual salary by 0.70, then divide by 2,080 hours. A $52,000 coordinator costs about $36 an hour, not $25.
The 0.70 is not a rule of thumb. The Bureau of Labor Statistics measures it every quarter. In its Employer Costs for Employee Compensation release for March 2026, private-industry wages and salaries averaged $32.60 per hour worked (69.9 percent of total compensation) and benefits $14.01 (30.1 percent). Dividing salary by 0.70 gets you to that average. Your own ratio sits somewhere either side of it. If the bookkeeper can pull total payroll cost divided by wages in five minutes, use that; if not, keep 0.70.
Three refinements matter for a cost calculation:
| Refinement | What to do | Why |
|---|---|---|
| Use the rate of whoever does each touch | Price a manager's approval at the manager's rate, a coordinator's coding at the coordinator's rate | Approval touches are short but expensive; blending everything at the junior rate hides them |
| Treat 0.70 as a floor | Add the per-seat software and desk cost if you want the full figure; leave it out if you want a number nobody can argue with | For a budget request, a conservative number that survives scrutiny beats a bigger one that gets picked apart |
| Divide by 2,080 hours, not hours actually worked | Keep the standard year even though nobody works it | The cover line (below) accounts for the weeks the person is away; dividing by fewer hours here would count that twice |
Rounded examples: $36 for a $52,000 accounts payable coordinator, $52 for a $75,000 property manager, $55 for an $80,000 office manager.

Your likely bottlenecks, and the AI solutions worth doing next.
You count touches with a two-week tally sheet and time a sample of ten. The person doing the process marks each time they pick it up and what kind of touch it was. At the end of two weeks you have touches per week by type. Then time ten touches of each type and use the median, not the average.
The method has three rules, and each one exists because skipping it produces a wrong number.
Tally for two weeks, not one. Most back-office processes have a cycle. Invoices bunch at month end, payroll queries bunch before payday, owner statements bunch in the first week. A single week lands either in the peak or the trough and misses the other. Two weeks catch one full cycle for anything monthly. For anything quarterly, tally two ordinary weeks and one peak week, and weight them.
Tally by touch type, not by process. A tally sheet with one column ("worked on invoices") gives you a total you cannot price, because the touches run at different rates and different people do them. Five columns (received and coded, matched, allocated, chased for approval, posted) give you five counts that each map to a person and a rate. Those same columns are the ones a matching system takes over; which of them stay with a person is laid out on the invoice reconciliation solution page.
Time ten touches and take the median. Three timings are noise. Ten is enough to see the shape. Use the median because manual work has a long tail: one invoice with a wrong purchase order number takes 40 minutes and drags an average up, but that invoice belongs in the rework line, not the direct line. If the spread across your ten timings is wider than three to one (fastest three minutes, slowest ten), the touch is two touch types wearing one name. Split it.
Estimates from memory run low. The person doing the work has stopped noticing the process, and the question "how long does an invoice take?" gets answered with the happy path. The tally sheet exists because it is the only way to see the approval chase and the re-keying that the person has stopped counting as work.
Owners price a manual process at direct minutes and stop. Three lines are missing from almost every estimate that arrives at an audit: interruptions to the person doing the work, a second person checking it, and cover for the weeks the person is away. Together they routinely add 25 to 40 percent to the direct figure.
These three are the section an owner cannot write from the outside, because each one is invisible to the person paying for it. The per-process price is the first thing an audit builds. In the law-firm audits that each turned up more than ten automation candidates, it is what let the partners rank them in an order they could defend rather than the order in which people complained.
A manual process that involves other people (vendors, tenants, clients, colleagues) generates its own interruptions: the vendor calling to ask where the payment is, the property manager asking which owner an invoice was allocated to. Each one costs the call and the re-entry. The University of California, Irvine study of fragmented work by Mark, Gonzalez and Harris found that people spent an average of 11 minutes on a piece of work before being switched away, and that when an interrupted task was picked back up the same day it took an average of 25 minutes to get back to it. You do not need to charge the full 25 minutes, since some of that gap is other useful work. The conservative rule: count only the interruptions that break a touch longer than 10 minutes, and price each at the call time plus 10 minutes of re-entry. If your tally sheet has an "interrupted" column, you already have the count.
Somebody signs off on the payment run, reviews the report before it goes to the client, or spot-checks the data entry. That person is usually more senior and more expensive, and their time never appears in the doer's estimate because it is not the doer's time. Ask one question: who would be blamed if this went out wrong? That person is checking it, whether or not the process document says so. Price their minutes at their rate.
The process does not stop when the person is off. According to BLS data on paid leave for March 2025, a private-industry worker averages 11 vacation days after one year of service, 15 after five, plus 7 sick days at any tenure. That is three to four weeks a year during which someone else does the work. The cover person is more senior (so more expensive), slower at a process they do not run every day, and lets a backlog build that then costs a catch-up week. Price cover as: cover weeks x doer hours per week x 1.5 x the cover person's rate. The 1.5 is a working assumption for someone competent but out of practice; if the cover person has never done it, use 2.
None of these three is a reason to automate on its own. They are the reason the owner's number and the real number disagree, and they are the first thing a quote gets challenged on if you leave them out.
Once the annual cost exists, two more steps turn it into a decision. The four-trait test for what to automate in your business tells you whether software can collect the cost at all, since a rule-based process gives up most of its touches and a judgment-heavy one gives up few. Then the return is arithmetic: cost removed each year against the one-time build and the monthly run cost, which is covered in how to calculate the return on an AI audit. Price the process first; both of those steps are guesses without the number.
Priced one process and want to know which of the others is bigger? That is the question our free AI assessment answers. Six questions, about two minutes, and you get a personalised preview of where the manual cost sits in your firm. Start your free AI assessment.