Why Most AI ROI Calculators Produce Garbage

There are hundreds of tools online that want to work out whether your AI project pays off. Most of them are not calculators. They are wishful thinking with an input form. You type in a few optimistic numbers, and at the end a percentage lights up that is big enough to carry into the next management meeting. The problem: that number does not survive a single critical question. And you find out only when the CFO takes a closer look - usually in front of everyone.

This is not a matter of taste, and it is not theory. Whether an ROI calculator is any good shows up in a few concrete places - at the edges, where the inputs get uncomfortable. In normal operation every calculator looks equally competent. The difference appears only when you push them to the limit. That is exactly where most of them fall apart. Here are the places where you can see it, with the mechanics behind each one.

1. The calculator that thinks "infinity" is an answer

Give a typical ROI calculator a use case with a clear saving, but leave the cost fields empty - because you do not know the licence costs yet, for instance. Many tools stubbornly compute savings divided by costs. And since you cannot divide by zero, the mathematical result is "infinity". Some then actually display an astronomical percentage, others crash quietly and show nothing at all, others still jump to zero without a word.

All three are wrong. There is no such thing as a project without costs, and an honest tool says "not calculable" at this point, or explicitly flags that an input is missing - instead of producing a fantasy number that looks like a jackpot. The point is not the edge case itself, which you rarely hit in exactly this form. The point is what it reveals about the build: whoever catches this case cleanly has thought about the edges. Whoever ignores it poured a formula into a web page and hoped nobody would step off the path.

2. The return that appears out of nowhere

A little more subtle, but from the same family: metrics like the internal rate of return, the IRR, can only be computed meaningfully if the cash flow changes sign at least once - money goes out at the start, money comes in later. Without that sign change, the result is simply undefined in mathematical terms. There is no right answer, only the honest statement "the inputs do not support this".

Bad tools produce a number anyway. Any number. They conceal that the basis for this metric is missing and sell you a false precision that sounds like hard financial mathematics. That is dangerous, because IRR values carry weight when investments are compared - an invented IRR can lead to a weaker project beating a stronger one, purely because the tool had the nerve to produce a figure in the wrong place.

3. The risk nobody answered - and that still comes out "medium"

Most ROI calculators ignore risk entirely. The few that include it often make a different, subtler mistake: if you skip the risk questions, they quietly insert an average and show you a reassuring "risk: medium" at the end. You answered nothing, and the tool pretends it assessed you.

That is more dangerous than no answer at all. An empty field signals "something is still missing here". An invented "medium" signals "all clear" - and that exact signal then travels unchecked into the decision paper. Your sales or investment team pushes a never-assessed project into the pipeline with a solid average attached, and nobody notices the assessment never happened. An honest tool distinguishes cleanly between "risk assessed, result medium" and "risk not answered" - and says exactly that in the second case, even if it looks less finished.

4. The ROI that changes when you reclassify a number

Here is a test you can run yourself in two minutes. Take a cost item - say CHF 12,000 for implementation - and enter it once as a one-off cost. Note the ROI. Then enter the same amount as twelve monthly instalments of CHF 1,000. In business terms this is identical money over an identical period. In many calculators the ROI figure still jumps noticeably, depending on which field you typed the amount into.

This happens because they use a naive formula - monthly saving divided by monthly licence, for example. Move an item from "monthly" to "one-off" and the denominator changes, tipping the whole metric. A robust model instead uses total cost over the period in the denominator: initial outlay plus twelve months of running costs. Then it makes no difference whatsoever how you classify the amount - the number stays stable. This is not a technical detail for accountants. It is the difference between a metric that tells you something about your project and one that mostly tells you something about your data-entry habits.

5. The calculator that does not know what its number depends on

The last and most underestimated point: a single ROI figure without context is worthless, however precise it looks. 180 percent sounds great - but if those 180 percent depend entirely on your assumption of 500 transactions per month, and the real figure turns out to be 300, the number is standing on one leg. A serious tool shows you which of your assumptions moves the number most, through a sensitivity view for instance. It does not just say "180 percent", it says "and that hinges mainly on your volume assumption, which you should check again". Leave that out and you have handed over a number without a package insert.

The garbage test: four questions, 30 seconds

You do not need a business degree to expose a bad ROI calculator. Four questions are enough, and you can run them against any tool in half a minute:

1. What does it show at zero cost? A huge number or a crash means: failed. "Not calculable" or a clear warning means: someone thought about it.
2. Does it factor in risk - and does it notice when you have not answered? A blanket "medium" without your inputs is a warning sign, not a convenience feature.
3. Does the ROI change when you move a cost item from one-off to monthly? If it does, the tool is measuring your typing habits, not the business case.
4. Does it tell you which of its assumptions moves the number most? Without that, you get a number without a package insert.

If a tool buckles on any one of these, you cannot trust its result. It is that simple - and that rare for a calculator to pass all four.

Why we are writing about this at all

Because we had to handle exactly these cases explicitly in our own calculator - and while building it we noticed how many tools simply do not. An ROI model only becomes useful once it stays honest at the edges: with missing inputs, with zero costs, with unanswered questions, with shaky assumptions. Everything else is a number that looks good in the boardroom and disintegrates in the first serious discussion - taking you with it.

That is exactly why AIghty20 Elevate thinks and calculates in business terms before talking about technology. An AI initiative is an investment, and you assess an investment with numbers that survive scrutiny, not with optimism that lasts until the first follow-up question. If you want to see what a calculator looks like that passes the four questions above: ours handles every one of these edge cases openly - including the uncomfortable ones others prefer to hide.

→ Run your own use case