The projector in the meeting room is still on. A stack of flowcharts is spread across the table, and next to it, the sales support staff's laptop still has that Excel file open. The question the owner asks me has come in roughly the same version for years: once this system is in, how many fewer people do I need to hire?
It's a completely reasonable question. So reasonable that not answering it head-on would be dishonest. I'm Johnny Yang, Brightstream Technology, Taiwan. I've led teams doing implementation for seven years, across 252+ projects in 32+ industries. An owner who judges whether something is worth doing by headcount isn't being short-sighted. Labor cost is money he actually wires out every month. "Improved efficiency" is not. You can't pay salaries with efficiency.
Here's the full conclusion up front: no outsider can tell you how much labor cost AI automation will save, because the answer is set entirely by five variables in your own company — items processed per month, manual minutes per item, the share that can be automated, the human intervention rate after implementation, and your own unit labor cost. Anyone who tells you "you can probably cut half your staff" before asking you about those five things isn't estimating. They're selling. And even if you fill in all five honestly, the number usually still comes out optimistic, because most people leave out two things: first, the maintenance cost that keeps recurring every month after go-live; second, saved time that isn't explicitly assigned to something else gets absorbed by daily work, and not a single dollar shows up as saved on the books. This post doesn't give you the answer. It gives you a sheet you can fill in yourself, plus two corrections that pull it back from optimism to reality.
First, an admission: counting heads is the right instinct
Plenty of people will tell owners not to look at automation in terms of labor cost. Think bigger, they say, think long-term competitiveness. I disagree with that, at least as the opening line.
The owner's instinct is right. Labor is the most concrete and least flexible line on a small business's P&L. Equipment can be bought a year later. Marketing can be paused. Once a person is hired, it's a fixed cost every month, and it goes up every year. Using it as the yardstick means the owner is evaluating with real money, which is far more honest than "things feel smoother."
The instinct is right. The math is wrong. Among the failed implementations I've seen, one type is very typical: the owner quietly calculates in his head, "this person costs me this much a year, once the system is in I won't need her," holds that mental number up against the quotation, decides it's a good deal, and signs. The problem is that number was never broken down. Nobody asked how much of that person's day is actually spent on this task, what share of the task a machine can really take over, or whether someone still has to check the work afterward. Without breaking it down, what gets signed isn't an investment decision. It's a mood decision.
So the fix isn't telling owners to stop looking at labor cost. It's putting that mental number on the table and splitting it into five cells. Once you do, you'll find some processes come out surprisingly large and are worth acting on right away. Others are so small you'd be embarrassed to call a meeting about them. Drop those, honestly.
The framework: count hours first, then talk money
I deliberately use only letters in this framework, no amounts. The reason is simple: the moment I fill one in, you'll be tempted to treat it as a benchmark, and my numbers mean nothing for your company.
| Code | Variable | How to measure |
|---|---|---|
| A | Items processed per month | Pull last month's actual count from the system or Excel. Don't go from memory |
| B | Manual minutes per item | Stand behind the person who does it, watch her do one start to finish, and time it |
| C | Automatable share | How much of this process a machine can genuinely take over. Estimate conservatively |
| D | Human intervention rate | The share that still needs someone to check, fix, or handle exceptions after go-live |
| E | Unit labor cost | Your own company's cost per person-hour. Only you can fill this cell |
The formula itself is simple. Monthly manual hours are A times B divided by 60. The hours that could theoretically be automated are that result times C. But after the machine takes over, someone still has to look back over the work, so the gross hours actually saved, S, equal A times B divided by 60, times C, times 1 minus D.
Here's a set of sample numbers to make it concrete, and let me be clear first: the A, B, C and D below are all assumptions I made up so the formula is readable. They are not real data from any client or any project. The only thing you should do with them is replace them with numbers you measured yourself. Say A is 800 items and B is 6 minutes. That's 80 person-hours a month. Say C is 0.7, so 56 person-hours could theoretically be automated. Then say D is 0.2, meaning 20% still needs someone to handle it, so the gross saving S is 44.8 person-hours.
Stop here for a moment, because this is the step in the framework most likely to sober you up. 44.8 person-hours isn't even one person's monthly working hours. It doesn't mean "you can hire one fewer person." It means "you can remove a fraction of one person's workload." Automating most single processes saves a fraction of a person. To add up to one full head, you usually have to handle several processes at once. That's why I keep recommending you first think hard about which process to start with rather than doing everything at once — what you need is to stack several fractions together, not spend your whole budget on one small process.
As for E, I won't fill it in for you, and no amounts appear in this post. Multiply S by your own cost per person-hour, and that number is your answer. It might excite you. It might also show you this isn't worth doing at all. Both outcomes are good outcomes, because you finally know what you're deciding.
The first missing piece: maintenance, and it happens every month
Most spreadsheets stop at the step above, and that's where they're least honest: they assume the system keeps itself alive after go-live.
It doesn't. Automation is alive, and it's connected to other people. A supplier changes its report format, an e-commerce platform redesigns, invoice rules get adjusted, customers start ordering in a new way, your own company adds a product category — every one of these makes a process that was running fine start spitting out strange results. Someone has to notice, someone has to judge, someone has to fix it. Those hours don't disappear. They just shift from "doing the task" to "keeping the task running."
I call this M, monthly maintenance person-hours, and correct the formula to: net saving equals S minus M.
M usually contains four kinds of work. Exception handling is the heaviest. Every rule or model gets things wrong sometimes, those cases need a person to look at them, and they're often the most troublesome ones. Next is chasing upstream changes: when a format changes, you adjust to match. There's no schedule, only emergencies. Third is tuning — patches to rule thresholds, prompt wording, classification logic. In the first stretch after go-live, these change almost every week. Last is the people cost: new staff have to learn it, and the judgment calls in a departing employee's head have to be handed over. This one is most often ignored entirely.
Treating M as zero is the most common self-deception I've seen. In most of my projects it goes like this: M is especially high in the first few months, because every exception nobody thought of shows up then. Afterward it comes down, but it never reaches zero. If you really don't know what to put for M, give it a value that feels "a bit excessive," because in my experience this item is far more likely to be underestimated than overestimated. It's the same thing as the hidden costs I broke down in the implementation cost post: what's written on the quotation is the start, not the whole.
There's an even more critical question: who carries M? If the answer is "everyone will figure it out," then nobody owns it. When something breaks, it falls back to manual work, and one day you find the sales support staff back in Excel while the system is still billing you every month. Automation with no owner for maintenance won't survive the first upstream change.
The second missing piece: does the saved time actually go to something else
This piece is even less visible than maintenance, and more damaging.
Say you do the math honestly and, after maintenance, you net 30 person-hours a month. Those 30 person-hours are "time," not "money." For them to become money, a management action has to happen in between. Without that action, the time goes one of three ways.
The first way: the same person uses it for higher-value work — chasing a few more receivables, making a few more sales calls, counting inventory more accurately. That's a good outcome, but the return won't show up in the payroll line. It shows up in revenue or dead stock. If you convinced yourself to buy by telling yourself "this saves payroll," the books won't add up, even though the company actually got better.
The second way: the time is absorbed by the work itself. Work expands to fill the time available, and nobody has to slack off for that to happen. With a little slack, messages get answered more carefully, files get organized more neatly, meetings run longer. Three months later you ask her if she's busy, and she says yes, very. She's not lying. In this case the realization rate is zero. What you bought is a more comfortable workplace, not lower cost. That's not a bad result in itself, but you have to admit it honestly instead of continuing to write "30 person-hours saved per month" in your report.
Only the third way is cash: not refilling a position when someone leaves, overtime hours actually dropping, less outsourcing or part-time help, no more scrambling to hire temps in peak season. Only when it gets this far does saved time become something you can see on the P&L.
So the framework needs one more coefficient at the end: R, the realization rate. The final formula is: actual cost saving equals (S minus M) times R, and only then times your unit labor cost E. And R is the only one of these five or six variables that technology has no say in. You decide it. Whether you wrote down "what this time will be used for" before go-live, and whether you wrote it specifically enough to check three months later, decides whether R is 0.8 or 0.
The advice is direct: write that sentence down before the project starts, and make it specific enough to verify. Not "give our people more time for important work." Instead: "every Wednesday afternoon, sales support chases receivables more than 45 days overdue and produces a list each month." The first is self-comfort. The second is management.
When this framework tells you not to do it
A calculation framework that always says "worth it" isn't a framework. It's a sales pitch. So I need to be clear about when it tells you to stop.
When A is small — when there just aren't many items a month — almost no automation pays off, because the base is too small while M is nearly fixed. For low-volume processes, a clearly written form and a fixed time slot for checking usually work better than a system, and cost less. Same when B is short: if a task only ever took a minute, there's little to save, and you'd be taking on long-term maintenance for it.
You should stop even more firmly when the process is still changing every day. Automation, by nature, locks the current way of doing things in place. Lock in a process that hasn't stabilized and you've just made the mess harder to change. Let it run stably for three months first, then talk about automation. Doing it in the reverse order gets expensive.
When nobody is willing to own M, or you know full well R is going to be zero — you have no intention of changing staffing, and when someone leaves you'll hire a replacement anyway — then this shouldn't be done in the name of "saving labor cost." It can be done in another name. Fewer errors, faster delivery, cleaner data are all legitimate reasons. But do something that's bound not to save cost in the name of saving cost, and a year later in the review meeting, the most embarrassed person will be whoever approved it.
So what should you do now
You can start tomorrow, without any tools. Take a sheet of paper and pick the one process that bothers you most. Just one. First, pull last month's actual count. That's A. Don't go from memory; memory usually overestimates by 20 to 30 percent. Second, stand behind the person who actually does it, have her do one from start to finish, and time it yourself. That's B. Don't skip this step. The minutes an owner estimates and the minutes on the floor are often far apart, and where the gap is turns out to be important information in itself.
Then give C and D conservative values. If you think 70% can be automated, put 60%. If you think only 10% needs a human to look back over it, put 20%. You'll correct them later anyway, but people who start out optimistic usually don't get a second chance.
Then ask two questions, out loud. First: after this goes live, who maintains it every month, and where does their time come from? Second: what specific, checkable thing will the saved time be used for? If nobody in the room can answer these two, put the sheet aside, because when M and R have no answer, everything calculated before them is an illusion.
Fill in that sheet and you'll have an AI automation cost estimate that belongs to your own company, not a pretty number from someone else's slide deck. If after filling it in you're still not sure which process to start with, come talk to me in a free 30-minute diagnostic. No slides needed. Just bring the sheet.




