Whoever shows you only the ROI on an AI investment picked the answer before you asked the question. Not because the ROI is wrong, but because on its own it says too little. An investment has four faces, and each answers a different question. Show only one and you hide the other three - usually not out of malice, but because the one shown happens to look best. And the one shown is almost always the ROI, because it produces the biggest percentage and needs the least explaining.
Here are the four numbers, what each can do and where each on its own leads you astray.
ROI: fast, but blind to time
Return on investment is the tabloid of metrics: a percentage, everyone grasps it instantly, it fits in a headline. Saving minus costs, divided by costs. The problem: ROI knows nothing about time. A franc you save in two years counts the same to it as a franc today.
For a quick plausibility check it works - "is this even worth thinking about?". As the sole basis for a multi-year investment decision it is too crude. A project with 150 percent ROI over three years can be worse than one with 120 percent over one year, because in the second case the money goes back to work far sooner. ROI cannot see that difference.
NPV: does the maths on time
Net present value repairs exactly that weakness. It discounts future cash flows back to today - a franc in two years is worth less than today, because you could have deployed the money elsewhere in the meantime. The discount rate, often around 10 percent a year, reflects your cost of capital.
NPV answers the question: is this project worth it in absolute terms, once I honestly account for the time value of money? A positive NPV means yes, the project creates more value than the capital would earn elsewhere. But it does not tell you how good it is against a specific alternative - an NPV of CHF 80,000 is an absolute number, not a comparative measure. And it reacts sensitively to the discount rate: set 8 instead of 12 percent and you get a noticeably different figure. That assumption should not stay hidden.
IRR: makes projects comparable
That is what the internal rate of return is for. It tells you what annual return the project generates from itself - a number you can hold directly against other investments and against your cost of capital. Two AI projects with a similar NPV but different IRR? Then you know which one puts your capital to work more efficiently, and that is exactly the question when you have to prioritise between several initiatives.
IRR does have its pitfalls. It can only be computed cleanly if the cash flow changes sign at least once - out at the start, in later. Without that change it is simply undefined. A tool that produces a number anyway is lying. And with irregular cash flows there can even be several mathematically valid IRR values - another reason never to read it in isolation.
Break-even: the number for the gut
The break-even month answers the most human question of all: when do I have my money back? It is not a return metric, but it is the number that calms a committee. "We are in the black in seven months" is understood by everyone - including those who mentally check out at discount rates and internal returns.
As a feel for risk it is indispensable: a break-even after 4 months carries a different risk from one after 26 months, even if both deliver the same ROI in the end, because more can happen in the longer window - market shifts, new tools, strategy changes. As the sole basis for a decision it is too simple, because it says nothing about how large the gain after break-even actually is.
The same project, four stories
Take a concrete case: a document bot, CHF 30,000 initial outlay, 500 transactions a month, 12 minutes saved per transaction. Set adoption to "instantly full" and every one of the four numbers looks rosy - triple-digit ROI, break-even within a few months, NPV and IRR cleanly positive.
Now put the same calculation on a realistic, ramping adoption curve. What happens? Break-even shifts back by months, NPV falls because the saving arrives later and gets discounted harder, IRR drops - and the ROI over twelve months changes most of all, because there is simply less saving in the first year. Same inputs, one single more realistic assumption, four markedly different stories. That is the core of it: one metric alone can be massaged by clinging to an optimistic assumption. Four together cannot all be flattered at once without the contradiction showing.
Which measure when
Take ROI for the quick check on whether thinking about it is worthwhile at all. NPV for whether the project creates value in absolute terms, with the discount rate disclosed. IRR when you have to prioritise between several initiatives and want to know which makes your capital work hardest. Break-even to ground the committee's sense of risk and gauge the exposure over time. Only all four together produce a picture that survives a critical question - and in the boardroom that question always comes.
Our AI ROI calculator delivers all four in one report - including clean handling of the cases where a metric is mathematically undefined, instead of inventing a number that looks good and is not true.
→ All four metrics in one report
