I was chatting with a course mate about how companies analyse their technology investments. And always, always, always we arrived at the same thing: they use the ROI models of a lifetime. The ones that worked well when we bought servers, when we migrated to the new ERP or when we changed the network infrastructure.
The problem is that now they’re trying to apply exactly the same calculator to Artificial Intelligence projects. And that’s where things go wrong.
Because it turns out AI doesn’t work like a classic technology project. Not by a long shot. In investment committees, brilliant initiatives get rejected because “the numbers don’t add up” with traditional metrics. And AI projects get approved that later fail because nobody really understood what they were buying.
AI introduces variables that classic models don’t contemplate: data quality, model training, human adoption, continuous improvement, constant learning. And, above all, a completely different life cycle from any tool you’ve bought before.
That’s why I think it’s important to talk about this. Because if we keep measuring AI with the wrong ruler, we’re going to keep making the wrong decisions.
A traditional ROI seeks efficiency. AI seeks transformation
When you buy traditional software, the goal is usually crystal clear: reduce costs, replace an obsolete tool, automate that process that’s been killing your team for years, improve operational times. The impact is direct and relatively easy to measure. You’ve invested X, you’ve saved Y, you do the maths and that’s it.
AI doesn’t work like that. An AI system doesn’t just optimise what you were already doing: it often completely changes the way your organisation operates. It can modify decision processes that hadn’t been touched in decades, redefine the relationship with your customers, optimise resources in real time in ways that were previously impossible, or generate capabilities that simply didn’t exist in your company.
In other words: while traditional ROI measures efficiency, AI ROI also has to measure transformation capacity. And that doesn’t appear in any standard spreadsheet.
In classic projects the system is stable. In AI, it evolves
A traditional application behaves relatively stably after its rollout. You install it, configure it, train your people, and that’s it. From there, basic maintenance and little else.
An AI system needs continuous training. Supervision. Retraining. Data maintenance. Constant adjustments. Because AI learns, evolves and —this is what many don’t understand— it can also degrade if you don’t look after it properly.
We’re going to see AI systems that work wonderfully for the first six months and then start to fail because nobody took care of maintaining the quality of the data feeding them.
That’s why, in AI, ROI isn’t calculated on the initial investment. It’s calculated over the model’s entire life cycle. And that cycle can be very, very long.
Data becomes a central part of the investment
In a traditional project, data is usually a secondary element. Nobody does an exhaustive analysis of data quality before buying a CRM, for example. It’s taken for granted that “we already have the data” and that’s that.
In AI, data is the core of the project. The whole project. The quality of the return you’ll get depends directly on the quality of your data, its availability, its structure and its ability to feed the models correctly. And here comes the surprise many companies get: the biggest cost isn’t in developing the AI.
The biggest cost is in preparing the data so the AI can work. There are projects where 70% of the budget goes on cleaning, structuring and normalising data. The AI model itself is almost the cheapest part. But that didn’t appear in the initial ROI presented to the board.
Classic ROI is usually immediate. AI ROI is progressive
In traditional projects, the return usually appears shortly after rollout. You install the system, start using it, and within weeks or months you’re already seeing benefits.
In AI this rarely happens that way. I’d say never. There are usually several phases: initial investment (which includes data, models, infrastructure), then a learning and adoption phase (where the system starts to understand your business), then optimisation (where you tune and improve) and, finally, scaling (when you really start to see the value).
This means AI ROI must be analysed over much broader time horizons. And with much more realistic expectations. If someone promises you immediate results with AI, be suspicious.
Human adoption carries far more weight in AI
In a traditional project, people usually adapt to the system. Like it or not, you end up learning to use the new software because there’s no alternative.
In AI, people have to trust the system. And that completely changes the game. Even if the model works technically well, if users don’t understand the recommendations, distrust the results or simply don’t integrate the tool into their daily operations, the real ROI can collapse.
I’ve seen projects where everything works perfectly from a technical point of view. The algorithms are brilliant. The data, impeccable. But the system ends up being switched off because nobody used it. Because people don’t trust it.
That’s why the cultural and organisational component carries far more weight in AI projects than in conventional technology projects. And that, again, doesn’t appear in classic ROI metrics.
AI generates value that’s hard to measure
Classic ROI works with clear financial metrics: savings, productivity, cost reduction, revenue increase. Everything can be put into an Excel sheet and the break-even calculated.
AI can also generate that kind of benefit. But it also brings strategic value that’s very hard to quantify: better decision-making, risk reduction, predictive capability, greater scalability, competitive advantage over your competition.
And often that intangible value ends up being the most important. But, since it doesn’t appear on the ROI spreadsheet, it isn’t taken into account when approving the project.
What I would do (and what I think you should do)
If you’re evaluating an AI project, don’t use only the classic ROI models. Not because they’re bad, but because they’re not designed for this.
AI evolves, learns, depends on data, requires human adoption and generates both operational and strategic value. That means you need to change the approach: measure efficiency, of course, but also measure transformation. Measure adaptability. Measure future competitive advantage. Measure the value of making better decisions. And, above all, measure with realistic time horizons.
Because in Artificial Intelligence the real return isn’t always in saving more. Sometimes it’s in being able to do things that were previously impossible.
And that is priceless. Or, rather: it has a price we don’t yet know how to calculate with the tools we’ve always used.