Kronehoefstraat 85
5612 HL Eindhoven
BTW NL863902509B01
Building On AI, Or Building Yourself Into a Corner?
qw
vc
mshm
.svg)
The biggest AI risk for a leadership team isn't which model you pick, it's the invisible dependency a custom project leaves behind. Start small, build model-agnostic, and prove value before you scale, and you avoid spending three to four hundred thousand euros on something existing models could already do.
Everyone in a leadership or finance meeting has an opinion on the best model. Which foundation model is most capable, which vendor improves fastest.
That debate feels important, and that's exactly the problem. The real risk isn't which model you choose. It's the dependency you build up without noticing while you're having that debate: on one vendor, on one way of working, on code that only holds up with that one provider.
An investment you can't reverse isn't an investment, it's a gamble. That insight doesn't come from a report. It comes from a conversation with a leadership team that lived through exactly that.
We see that pattern often: a leadership team debates which model evolves fastest, while nobody asks who holds the keys once the terms change. That question is exactly where a good AI strategy starts, not with the technology.
A company in our network put three to four hundred thousand euros into a custom AI solution. A year later, it still wasn't working properly. Meanwhile, existing models on the market, with little adjustment, could already handle a large part of that same task.
Nobody in that project had bad intentions. The vendor built what was asked, and leadership trusted expertise they didn't have in-house. One step was missing: a short, independent check on whether existing options already sufficed before the contract was signed.
The question nobody asked beforehand: are we building something new, or wrapping a layer around something that already exists? The answer determines whether you spend three hundred thousand euros on a working product, or on reinventing something a vendor already offers.
This isn't an isolated incident. Research by MIT NANDA (2025) into more than three hundred corporate AI implementations found that 95% of pilots delivered no measurable financial return, even as companies invested $30 to $40 billion combined. Not because AI doesn't work, but because the approach didn't match what already existed.
Building model-agnostic doesn't mean you never choose. You do optimize for one model: you tune prompts, context, and workflow to that specific model's strengths, a form of fine-tuning of your application, not of the model itself.
The difference is where that choice gets locked in. Hard-code the connection into standalone scripts that only work with that one API, and you're stuck. Build the same connection into a layer you can decouple and replace, and you get optimization without dependency.
That distinction sounds technical, but the bill is a business one. A leadership team that doesn't know the difference signs the wrong two-year contract without a single alarm going off.
Platform research confirms this difference costs real money in practice. Atlan (2026) points to migrations that regularly exceed one million dollars, and switching projects that take months to years at more complex organizations.
"A better, cheaper model" is only a meaningful choice if you know what "better" means for your task. Without measurement, switching is a gamble, exactly the problem the first section opened with.
So record three things for every task a model performs: the prompt, the model used, and the outcome. Score that outcome on a fixed scale, for example on accuracy, completeness, and how often the model makes something up that isn't true (a hallucination).
This takes a few hours of setup, not a new team. A spreadsheet with a hundred sample tasks, their outcomes, and a score from one to five is often enough to make a clean comparison between your current model and an alternative.
With that data you can put a new model side by side with the old one before you switch. Without a scorecard, "the new model is cheaper" isn't a saving, it's a hunch.
Frontier models are expensive to build. Training costs for the heaviest models on the market climbed toward hundreds of millions of dollars per model in 2025, and the expectation is that this will keep rising (Galileo AI, 2025). That money comes from investors who want a return.
At the same time, cheaper, often open models are closing the gap fast. Stanford HAI (2025) measured that the performance gap between the best closed and the best open model shrank within a year from more than eight percentage points to just a few. SemiAnalysis confirms this pace is accelerating with each model generation.
For your business case, this means two things. Today's premium model is a commodity in two years, and a three-year contract based on today's price ratios won't hold up by then.
For a leadership team deciding on a three-year contract, this matters more than which model scores best today. The question isn't whether the market will change, it's how fast you can move when it does. Read more about what this shift means in the state of the AI economy.
A vendor asking three to four hundred thousand euros for custom work is often still pricing it at development costs from two years ago. Organized well, comparable custom work now costs a fifth to a tenth of that amount.
That saving doesn't happen on its own. It comes from two things being present together: a product owner with real knowledge of the process being automated, and a developer who builds model-agnostic instead of around one API. Miss either one, and you still pay the high price, just for less result.
So don't just ask a vendor for the price, ask about the assumptions behind it: are they still pricing at rates from two years ago, or have they already built the drop into their offer?
The underlying costs are falling fast. Epoch AI (2025) measures that the price to achieve a fixed level of performance per task drops 9 to 900 times a year, with much of that drop happening in the past year alone. Anyone who passes that difference on in their rate builds the way the example in building as an AI-native company describes.
Starting small often feels like delay to a leadership team. It's the opposite. A proof of concept (POC) with a fixed, limited budget proves whether an approach works before you multiply the spend.
Define upfront what "it works" means, in numbers you set at the start. Only scale once those numbers check out. That's not a brake on ambition, it's the only way to avoid spending three to four hundred thousand euros on something an existing model could already do.
A fixed budget here acts as a brake on scope creep, not on ambition: it forces the question "what does this need to prove" before the first hour of work, instead of after.
Scoping tightly, building model-agnostic, and proving value within a fixed budget: that's where an AI strategy that doesn't get in your way begins.
How much of your current AI spend would you still defend today against your own numbers?
It means your application isn't locked to one AI vendor. You deliberately choose one model to work with, but the connection runs through a layer you can replace without rebuilding the whole application. That keeps you free to switch without starting a new project.
No. Existing models on the market can already handle a large part of many tasks, often at a fraction of the price. Custom work only becomes the right choice once you've first tested that an existing model can't handle the task well enough, not before.
Ask what happens if you want to switch to a different model. Do you have to rebuild, or just change a setting? If you don't get a clear answer, the dependency is probably bigger than you think.
That varies a lot by task and vendor, and we won't name a fixed figure here. More important than the exact amount is the condition: a pilot with a fixed, limited budget and success criteria set in advance, before you decide to scale.
Record the prompt, the model used, and the outcome for every task, and score that outcome on a fixed scale. Without that data, any comparison between models is a gamble instead of a decision.
Competition between vendors and more efficient techniques are pushing the cost per task down fast. Research by Epoch AI (2025) shows drops ranging, per performance level, from nine to several hundred times a year.
First look at what assumptions the contract rests on. A multi-year contract based on today's prices risks having you pay well above market rate for the same task a year from now.
Start with a short audit of your existing AI plans: which ones are locked to one vendor, and which would work just as well with a middle layer? That determines where you adjust first, before you free up new budget.
Scope tightly, build model-agnostic, and prove value within a fixed budget. That's where a strategy begins that won't get in your way two years from now.
.svg){kind=icon}