The moment an AI investment feels like progress is the moment the invoice arrives: licence booked, access set up, the tool is running. Number in the budget, box ticked. And this is precisely where most people are mistaken. The licence is not the investment, it is the cheapest line item of them all. What an AI strategy really costs appears on no quote - and what many people fear does not end up costing them at all.
This is not rhetoric, it is measurable. Anyone who separates the real costs from the imagined ones spends less and gets more back. Anyone who confuses them lands in the large majority of projects that tie up a lot of money and move very little. So let us start where the money is actually lost.
Why do 95 percent of organisations see no return - and what does that say about the costs?
The report "The GenAI Divide: State of AI in Business 2025" out of MIT's NANDA project labels its own results preliminary findings and gives its base as 52 organisations, 153 surveyed senior leaders and more than 300 publicly disclosed AI initiatives. Fortune's write-up of 18 August 2025, which we relied on here first, states a different methodology: 150 leaders interviewed, 350 employees surveyed, 300 public AI deployments analysed. The finding sits on page 3 of the report: 95 percent of organisations get no measurable return from GenAI, and only around 5 percent of integrated pilots are extracting millions in value. And the decisive sentence: the culprit is almost never the quality of the model, but a "learning gap" - the tool is not embedded in the actual workflows.
This is the most expensive insight of all, because it shows where the money burns. Not on the model you buy, but on the integration you skip. A generic AI tool quickly achieves something for a single person, yet stalls inside a company because it knows nothing about your processes, your customers and your data. The cost of an AI strategy therefore does not sit where the invoice records it.
What does an AI strategy really cost?
Three things that never appear in the licence price: cleaning up your data, defining your processes, and one person who owns the whole thing. In the Bitkom AI study 2026 (604 companies with 20 or more staff), 33 percent report that costs came in higher than expected - and the biggest hurdles are a lack of employee skills (53 percent) and integration into existing processes (39 percent). These are not software costs. They are people and order costs.
Concretely: before an AI can say anything useful about your customers, it needs clean first-party data - data that genuinely belongs to you and fits together. Before it can speed up a workflow, that workflow has to be defined in the first place. And before anything runs permanently, it needs a human who maintains the rules. Those three items are the real price - and the reason two equally priced licences deliver completely different results.
And what does it explicitly not cost?
Not your own data science team, not a self-trained model, and not a six-figure budget to start. The MIT report is unambiguous here: those who buy ready-made, specialised solutions and integrate them with an external partner - such as an AI agency - succeed in roughly 67 percent of cases - in-house builds only about a third as often. The expensive bespoke route is statistically the worse one.
For small and mid-sized teams this is good news: you do not have to invent AI, you have to embed it. An AI agent that takes over one clearly defined step in your customer journey - pre-sorting enquiries, answering follow-up questions, setting up appointments - delivers more than an ambitious in-house project that never ships. The fear of the big budget stops many people from taking the small, cheap first step.
How do you know you are spending money too early?
You buy a tool before you know the use case. That is exactly what is happening on a mass scale right now: according to the Voice of the Enterprise study by S&P Global 2025 (1,006 professionals surveyed), 42 percent of companies have abandoned most of their AI initiatives - the year before it was only 17 percent. On average, 46 percent of all pilots end up in the bin before they ever go into production.
This is not a technology problem but a sequencing problem. Anyone who buys first and then wonders what for is paying licences for questions they never asked. The test is simple: can you say in one sentence which concrete result this AI initiative is meant to improve, and how you will measure it? If not, every euro is spent too early. Strategy comes before the tool, not the other way round.
What is the cheapest first step?
Choose a single process, make it clean, and apply AI exactly there - measurably and with an owner. Not the whole customer journey at once, but the one point where you lose the most time today or where the most enquiries slip through. A narrow, well-chosen use case is cheaper, goes live faster and delivers the data with which you justify the next step.
The charm of this approach is that it rules out the expensive mistakes before they happen. You do not automate a gap, because you close it first. You do not buy a tool on suspicion, because you know the case first. And you do not tie up budget in a prestige project, because you start small and provable. That is how "AI costs too much" becomes a sentence you no longer need.
Three levers for an honest AI cost calculation
Budget for the integration, not just the licence. The biggest cost block is invisible: clean data, defined processes, an owner. Anyone who only budgets for the software plans the most expensive part out of the calculation - and ends up among the 33 percent whose costs were "higher than expected".
Buy in rather than build. Ready-made, integrated solutions succeed three times more often than in-house builds. Your own model, your own team, your own system are rarely the cheaper and almost never the faster choice. For a look at possible partners, see our comparison of AI agencies in Vienna.
The case first, then the tool. If you cannot name the goal measurably in one sentence, it is too early to buy. A clearly defined first use case is the cheapest entry point there is.
An AI strategy costs less than most people fear - but in a different place than they think. If you want to know where your first euro moves the most, just drop us a line. 💡