What Is the Real ROI of an AI Employee?
Question — — by Mahmoud Zalt
Payback period, not percentages. What is worth counting, the two numbers that make every model lie, and a worked calculation you can copy.
Somebody has probably shown you a calculator with a 340 percent ROI on it. You closed the tab, because you know what a number looks like when it was reverse-engineered from a conclusion.
The good news is that a real calculation fits on one page and takes fifteen minutes. It will not produce an impressive percentage. It will produce something better, which is a number you can defend to yourself in three months.
Sistava charges a monthly plan fee with a credit allowance included, so the cost side of your model is one predictable number plus usage you can watch and cap. That is the easy half. The benefit side is where every model goes wrong, and it is where the rest of this article lives.
At a Glance
- 4
- Things actually worth counting in the model
- 2
- Numbers that make every ROI model lie
- 1 month
- Before you have data worth trusting
- $200
- The third tier, per month
What is the real ROI of an AI employee?
It is the value you can trace, minus the time you spend making it work, divided by what you paid. In practice the honest answer for a well-chosen job is a payback period of a few weeks, and for a badly chosen job it is negative forever. Which of those you get is decided by the job, not by the software.
That is why a general ROI figure for AI employees is meaningless. The same subscription pays back in nine days for someone chasing 200 quotes a month and never pays back for someone automating a quarterly report. The variable that matters is not the price, it is what you pointed it at.
Count four things, and refuse the other two
There are exactly four inputs worth putting in the model, two on each side. Everything else people add is there to make the number bigger, and it is the reason those calculators feel dishonest even when the arithmetic checks out.
| Side | Count this | Do not count this |
|---|---|---|
| Benefit | Hours you genuinely stopped spending | Hours a tool 'could' save if used perfectly |
| Benefit | Revenue you can trace to work that now happens | Revenue you hope for from better focus |
| Cost | Plan fee plus credit usage | Nothing, this half is easy |
| Cost | Your setup hours and ongoing review time | Nothing, but almost everyone omits it |
| Never | Productivity multipliers, morale uplift, strategic value |
The two fake numbers are worth naming. The first is capacity you did not use: if you got four hours back and spent them on email, you did not convert them into value and you should not count them. The second is anything with the word potential in front of it.
Why payback beats an ROI percentage
A percentage hides the size of the bet and the time it takes. A payback period tells you both at once, in a unit you already think in. If the answer is three weeks, you can act on it today. If the answer is fourteen months, you know it is a maybe rather than a yes.
The calculation is one line. Take your monthly cost, take the monthly value you recovered, and divide the first by the second. A result under one means the thing pays for itself inside a month. A result over three means look harder before you commit.
There is a second reason to prefer payback. It forces you to state the monthly value in a real number, which is the step most people skip. It is easy to say a task is a huge time sink and much harder to say it costs eleven hours and forty minutes a month, which is what the model needs.
Worked calculation: 140 inbound signups a month
Dev runs a B2B software product with one other person. Every month about 140 people sign up for the free version. He always meant to research the promising ones and send a personal first message, and in practice he got to about 30 of them before the week ran out.
He hired an AI employee to research every signup against a short set of fit criteria, pull public company details, and draft a first message for him to approve. He kept the send button. That distinction matters to his customers and it matters to the model, because his review time is a real cost.
| Line | Dev's number | How he got it |
|---|---|---|
| Signups per month | 140 | Product analytics, three-month average |
| Signups he used to research | 30 | Honest count from his sent folder |
| Manual research time | 11 minutes each | He timed eight of them |
| Hours he was spending | 5.5 per month | 30 signups multiplied by 11 minutes |
| Signups now researched | 140 | All of them, drafted overnight |
| His review and send time | About 2.5 hours per month | Roughly 65 seconds per draft, measured |
| Net hours recovered | About 3 per month | 5.5 hours minus 2.5 hours |
| Extra demos booked | 3 per month | Compared against his previous three-month baseline |
| Demo to customer rate | 25 percent | His own historical figure |
| New customers per month | About 0.75 | 3 demos multiplied by 25 percent |
| Average first-year contract | $2,400 | His actual average |
| Traceable monthly value | About $1,800 | 0.75 customers multiplied by $2,400 |
Look at what happened to the hours. He only recovered three a month, because reviewing 140 drafts costs real time even at sixty-five seconds each. If hours were the whole story this would be an unremarkable result.
The value came from the 110 signups he was never going to reach. That is the pattern in most honest ROI models: the prize is not the work you were already doing faster, it is the work that simply was not happening. His payback fraction is his plan fee divided by roughly $1,800, and at that ratio the plan tier stopped being an interesting question.
What ruins an ROI calculation
Four mistakes account for nearly every model that looks great on a Tuesday and embarrassing in month three. They are easy to avoid once you have seen them named.
- Counting gross hours instead of net. Review time never goes to zero. Subtract it every single time, and measure it rather than guessing at it.
- Valuing recovered hours at your billing rate. Hours are only worth that if you actually sold them. Otherwise they are worth whatever you did with them, which is often less.
- Using planned volume instead of last month's. Planning numbers flatter everything. Use the volume you had, from a system that counted it for you.
- Forgetting setup hours entirely. Three to five hours in the first fortnight is normal and it belongs in the first month of the model, not nowhere.
There is a fifth, subtler one. People model the good month. Run your numbers on an average month and a bad month too, because a payback that only works during your busy season is a seasonal decision rather than a permanent one.
When the answer is negative, and why that is useful
Sometimes you run the model and it says no. That is a good outcome, delivered cheaply, and it usually means one of three things: the task does not repeat often enough, the review cost is nearly as high as the manual cost, or the work needed judgment you cannot write down.
In every one of those cases the right answer is to leave the task alone or to hire a person for it. A model that never returns a negative is not a model, it is a sales tool, and you should be suspicious of any calculator that cannot tell you not to buy.
The category where a person's return is plainly higher is worth stating directly. Work that needs accountability, work that is a relationship, work that changes shape weekly, work that needs a legally binding signature, and anything physical. Those are not underserved by software, they are outside what software does, and hiring for them is the highest return decision on the list.
Build your own payback model in fifteen minutes
Seven lines on one page
- Name one job and count last month's volume — From a system that counted for you: your CRM, your helpdesk, your analytics. Not from memory.
- Time yourself on five real instances — Average them, multiply by volume. That is your current monthly hours, and it is usually higher than you thought.
- Write down the work not happening at all — The follow-ups you skip, the leads you never research. This is where most of the real value hides.
- Run it for a month, then measure review time — Time your own checking. Subtract it from the hours saved. What remains is your net hours.
- Trace revenue against a real baseline — Compare to the three months before, not to a feeling. Only count what you can point at.
- Add setup hours to month one — Three to five hours is typical. Put them in the first month so your payback figure is honest rather than flattering.
- Divide monthly cost by monthly value — Under one pays back inside a month. Over three, go find a better job to automate before you commit.
FAQ
What is the real ROI of an AI employee?
It depends entirely on the job, which is why any general figure is worthless. Calculate payback instead of a percentage: divide your monthly cost by the monthly value you can actually trace. A well-chosen high-volume job often pays back in weeks. A quarterly task with unclear rules never pays back at all, and the model should be able to tell you that.
How long before I know if it worked?
One month of real running, and you should not trust anything earlier. The first two weeks include your setup time and your least accurate output, so they distort the picture in both directions. Compare month one against the three months before it, using numbers a system counted rather than numbers you remember.
Should I value recovered hours at my hourly rate?
Only if you actually sold those hours. If you got four hours back and billed them, use your rate. If you spent them on admin, they are worth much less, and inflating them is the single most common way an ROI model ends up wrong. Be honest about what the recovered time turned into.
What is a good payback period?
Under one month is a clear yes and requires no further analysis. One to three months is fine for a job you are confident repeats. Beyond three months, look for a better candidate first, because a task with a long payback is usually one with low frequency or high review cost, and both of those tend to get worse rather than better.
Where does most of the value actually come from?
Usually from work that was not happening at all, rather than from doing existing work faster. The follow-ups you skip, the leads you never research, the reports nobody has time to write. Recovered hours are the visible half and often the smaller one, so make sure your model counts the missing work as well as the existing work.
When is hiring a person the better return?
When the work needs someone accountable for the outcome, when the value is a long-term relationship with a customer, when the process changes shape too often to write down, when a signature carries legal weight, or when someone has to be physically present. In those cases hiring is the highest return decision available, and no payback calculation changes it.
A real ROI number is smaller and more useful than the one in the brochure. It has your volume in it, your review time, and your setup hours, which is exactly why you will still believe it in three months.
Pick the job with the highest volume and the least judgment, run it for a month, and do the division. Whatever comes out is worth more than any percentage anyone else can calculate on your behalf.