AI-First: Better strategies than the hype cycle

10.02.2026

7 min read

Key Takeaways

Between 80 and 95 percent of all AI projects in companies never reach productive operation – despite increasing budgets.
“AI-first” systematically leads companies to buy technology and then search for suitable problems.
Successful companies focus on vertical prioritization: solving one concrete problem completely before moving on to the next use case.
Augmentation instead of replacement generates more business value – as shown by the counterexample of Klarna, whose radical AI course led to decreasing service quality.
A “problem-first” filter with three core questions before each project start reduces the error rate and increases the impact of AI investments.

A significant portion of documented AI implementations in companies do not deliver the promised results. The number varies depending on the source, but whether you read from MIT or industry analysts: Between 80 and 95 percent of projects stagnate, are quietly abandoned, or never reach productive operation. Meanwhile, most executives are increasing their AI budgets for 2026. This is not a contradiction – it is a warning signal. Because hope is not a strategy.

The automotive industry provides vivid examples. Volkswagen’s software subsidiary Cariad burned through billions before the company pulled the plug – not because the technology did not exist, but because the organization had not defined a clear, prioritized problem. Instead, everything was supposed to be transformed at once. The result: lots of infrastructure, little output. Cariad is not an isolated case. It is the pattern.

Why “AI-first” Systematically Fails

The error begins with the framing. “AI-first” sounds like determination and a willingness to innovate. But in operational reality, it usually means: A company buys AI tools and then looks for problems that fit them. This is the reverse of any functioning innovation logic. Technology does not solve problems that no one has formulated.

The first symptom is the solution without a problem. Departments receive budgets to “do something with AI.” They evaluate tools, build demos, present at internal innovation days. But no one has asked beforehand: Which specific process do we want to improve? By how much? And how do we measure that? The result: Pilot projects that work technically but do not generate any measurable business value.

The second symptom is the horizontal PoC flood. Many companies start ten, twenty, or more proofs of concept in parallel in different departments. Each team works with different data, different providers, different success criteria. The result is a portfolio of half-finished experiments, none of which clears the scaling hurdle. Not because the technology fails, but because no one has done the organizational groundwork – data infrastructure, governance, change management.

The third symptom is the lack of real success metrics. “We are now using AI” is a press release, not a KPI. As long as companies measure the success of AI projects by their mere existence rather than by measurable improvements in throughput time, error rate, or cost reduction, the investments remain an act of faith. And faith does not scale.

What successful companies do differently

The few companies whose AI projects demonstrably work share a characteristic: they don’t treat AI as a strategy, but as an engineering discipline. The difference sounds subtle, but it’s fundamental.

Vertical prioritization instead of horizontal dispersion. Instead of introducing AI everywhere at once, these companies identify a specific process with high pain potential and high data availability. They solve this one problem completely – from data cleansing to model training to integration into the operational workflow. Only then comes the next use case.

Clear business cases before the first prompt. Before even a single API is connected, there’s a business case that quantifies the expected benefit. Not in vague categories like “efficiency improvement,” but in concrete numbers: reduction of processing time from 48 to 12 hours. Lowering the error rate from 8 to 2 percent. Saving 200,000 Euro per quarter. Anyone who can’t name these numbers doesn’t have an AI project – they have a hypothesis.

Augmentation instead of replacement. This ties in with the Klarna debate, which sparked waves again in mid-2025. Klarna had chosen “AI-first” as its company motto and consistently cut jobs. The result: declining service quality, growing customer frustration, a CEO who had to publicly admit that the math didn’t add up. The counter-model are companies that use AI to augment human work. A clerk who processes twice as many cases with AI support generates more value than a bot that answers every second case incorrectly.

The problem-first framework: Three questions before every AI project

Anyone who wants to reduce the error rate of their AI initiatives doesn’t need a new tool or another consultant. They need discipline with three questions that must be answered before every project start.

Question one: What specific, measurable problem does this project solve? If the answer is “we want to use AI,” that’s not a problem, but a means. Back to the start.

Question two: What is the current baseline value, and what’s the target? Without a baseline, there’s no measurable progress. If you don’t know how long a process takes today, you can’t judge whether AI makes it faster. Establishing the baseline is often more laborious than building the model – and that’s exactly why it’s skipped. Fatal error.

Question three: What happens if the AI is wrong? Every AI system has an error rate. The question isn’t whether errors occur, but whether there’s a fallback position. In insurance claims processing, a misclassified case might be an annoyance. In medical diagnostics, it’s a catastrophe.

These three questions aren’t an innovation framework for workshops. They’re a filter. If you answer them honestly, you’ll find that out of ten planned AI projects, maybe three pass the test. That’s exactly the point. Three well-thought-out projects that go into production beat thirty pilots that end in slide presentations.

Checklist: Ready for AI projects or just for AI announcements?

Five questions for an honest assessment in two minutes:

Can you name a concrete, measurable target value for every ongoing AI project?
Is there a defined process owner for every project who’s responsible for success – not just for the technology?
Do you have a documented data baseline for the processes that AI is supposed to improve?
Is there a fallback strategy in case the AI system fails or delivers incorrect results?
Is the success of your AI initiatives measured by business metrics – or by the number of projects started?

Anyone who answers at least four of these questions with yes is already working problem-first. Everyone else should not expand their AI roadmap, but sharpen it. Fewer projects, more impact. Less vision, more engineering. Less “AI-first,” more “problem-first, AI-enabled.” That’s not a brake on innovation. It’s the prerequisite for innovation to actually land.

Frequently Asked Questions

Why do so many AI projects fail in companies?

The most common cause is a lack of problem definition. Companies acquire AI tools without first formulating a concrete, measurable use case. This is compounded by poor data quality, lack of governance, and unclear success metrics.

What does “Problem-first, AI-enabled” mean?

What is the difference between Augmentation and Replacement in AI?

Augmentation means that AI supports and enhances human work – for example, through faster data analysis or suggestions. Replacement completely replaces human labor with AI. Studies and practical examples like Klarna show that Augmentation often delivers better results.

How do I create a business case for an AI project?

What role does data quality play in AI implementation?

Data quality is the crucial foundation. Without cleaned, structured, and complete data, no AI model can work reliably. Establishing a data baseline and building data infrastructure is often more labor-intensive than the actual model training.

What can companies learn from the Klarna example?

How many AI projects should a company pursue simultaneously?

Less is more. Successful companies prioritize vertically: they solve one problem completely before tackling the next use case. Three well-executed projects with measurable impact are more valuable than thirty parallel proof-of-concepts.

Source of the title image: Unsplash / Scott Graham