How to measure the ROI of office automation: minutes per item, items per person

"It saves hours" is not an ROI number

Every automation pitch says it saves time. Most owners have heard "it'll save your team 10 hours a week" and had no way to check it a month later. The problem isn't that the claim is wrong. It's that nobody measured the starting point, so nobody can prove the ending point.

Office work comes in units. An order. An invoice. A purchase order. A quote. That makes it easier to measure than most people think. You don't need a consultant or a dashboard. You need a stopwatch, a sample, and two numbers. This guide covers which two, how to collect them, what the research says about getting it wrong, and where the skeptics are right.

Why you can't trust a guess, even your own

In July 2025, the research group METR ran a randomized trial with 16 experienced software developers working on 246 real tasks from their own projects. Each task was randomly assigned to "AI allowed" or "no AI." Before starting, the developers expected AI to make them 24% faster. With AI, they took 19% longer. And afterward, they still believed AI had sped them up by 20%.

What developers believed vs. what happened

Effect of AI tools on time to finish a task. METR randomized trial, 16 developers, 246 tasks, 2025. Positive = faster.

1. Expected before starting24% faster
2. Believed afterward20% faster
3. Measured19% slower

Source: METR, July 10, 2025.

Be careful with this study. It's software development, not order entry. METR itself calls it a snapshot of early-2025 tools in one setting, and says it does not show that AI fails to speed up most people. The lesson for an office isn't "AI makes you slower." It's that people are bad at judging their own speed, in both directions. A team that likes a new tool will feel faster. A team that's nervous about it may feel slower. Neither feeling is a measurement.

Self-reported time is shaky in general. A 2011 BLS Monthly Labor Review study found people overestimate their work hours by 5 to 10 percent compared with time diaries, mostly among people who report long hours. A 2014 follow-up in the same journal argued the gap is much smaller for "hours last week" questions. Researchers disagree on the size. They agree a clock beats a memory.

If you didn't time it before, you can't prove it after. Measure the baseline first.

The two numbers to track

• Minutes per item. How long one order, invoice or PO takes from the moment it arrives to the moment it's done in your ERP. Start the clock when someone opens it. Stop it when it's entered and checked.
• Items per person per day. How many a person finishes in a normal day, counting the interruptions, phone calls and lookups that come with the job. This catches what the stopwatch misses.

Add one more as a guard: errors per 100 items. A wrong price, a wrong ship-to, a credit memo later. If minutes go down and errors go up, you haven't saved anything. You've moved the work to someone else's desk.

What the benchmarks say about cost per item

The best public benchmarks come from APQC, a nonprofit research group. Its 2016 report on sales order processing was sponsored by Esker, which sells order automation software, so keep that in mind. Across 608 organizations, the total cost to manage sales orders was $5.11 per order for top performers, $24.21 at the median, and $40.87 for bottom performers. In APQC's words, bottom performers spend "almost 8 times more" than top performers. The report names labor as the biggest cost driver.

Total cost to manage sales orders, per order

APQC benchmark, 608 organizations, 2016 (Esker-sponsored).

1. Top performers$5.11
2. Median$24.21
3. Bottom performers$40.87

Source: APQC, Cutting the Costs of Sales Order Processing, 2016.

The same report split cost by channel. Orders through what APQC calls traditional "paper-laden methods" cost a median $12.60 each, in samples of about 750 organizations. Orders through digital channels, such as a cloud service that handles the steps on one platform, cost a median $4.00.

Payables shows the same pattern. An APQC blog post from 2018 says top performers in accounts payable process five times as many invoices per person as bottom performers, at about a fifth of the cost. That's "items per person" in action.

Use these as a sense of the range, not a target. They're 2016 and 2018 figures, they include costs beyond labor, and your business isn't the median. The point is that the gap between good and bad is large, and most of it is people's time.

What an office minute really costs

Price the minutes at full cost, not the paycheck. In its Employer Costs for Employee Compensation release for June 2026, BLS put private industry compensation at $46.89 per hour worked. Wages were $32.82 of that (70%). Benefits were $14.07 (30%). Use your own payroll numbers if you have them. If you only know wages, that 30% is a reasonable reminder that the true cost is higher.

How to run the measurement

1. Pick one process. Order entry, AP invoices or purchase orders. Not all three.
2. Time 50 items before. Spread them across the week and across people. Include the messy ones: the PDF with handwritten notes, the customer who uses their own part numbers. Write down minutes, and whether it needed a fix later.
3. Count items per person per day for the same week, from your ERP's entry log if it has one.
4. Change one thing. Turn on the automation for that process only.
5. Wait two to four weeks, then time 50 items again the same way. The first week is learning, not results.
6. Do the math. Minutes saved per item, times items per month, times your full hourly cost. Then subtract what the errors cost before and after.

In this example, 100 hours a month come back. At BLS's $46.89 average, that's worth about $4,700 a month in time. Your numbers will differ. The method is what matters: real minutes, real volume, full cost, and errors counted.

Where the skeptics have a point

Saved time isn't saved money. If nobody's hours change and the freed time goes to longer lunches, the books don't move. MIT's NANDA project, in a 2025 report (preliminary, based on interviews, copy hosted by a law firm), found that companies getting value from AI mostly saw it without cutting staff. The savings came from spending less on outside help, such as outsourcing and agencies. Decide ahead of time where the hours go: more orders without a new hire, faster quotes, or work that never got done.

Being watched changes the work. People often work differently when they know they're being timed. A 2014 review of the "Hawthorne effect" found these effects exist but that their size is hard to pin down. Time both samples the same way so the effect hits before and after equally.

Benchmarks come from vendors. Much of the public data, including the APQC reports above, was paid for by companies that sell automation. The figures are real benchmarks, but they were chosen to make a case. Your own 50 orders are better evidence than any of them.

Small volume may not pay. If you process 15 invoices a week, saving five minutes each is about an hour. That may not cover the effort to set anything up. Measuring tells you that too, and that's a good outcome.

What to show your CFO

One page. The process, the before and after minutes per item, items per person per day, errors per 100, monthly volume, and your full hourly cost. Where the hours went. No vendor estimate, no "industry average" savings claim. If the numbers are good, they'll speak for themselves. If they aren't, you'll know early and cheaply.

For how automation fits a real order desk, see our order processing build, and if you're still deciding where to start, read what to automate first.