Klarna and Duolingo reversed AI-first before the EU adopted it
Ursula von der Leyen announced "AI first" in October 2025.
By then, the startups who pioneered it had already reversed course.
Klarna cut 700 customer service roles, announced AI would replace them. 18 months later they're rehiring humans because quality dropped so badly customers noticed.
Duolingo declared "AI-first transformation" in early 2025, phased out contractors. By mid-year they faced sustained user revolts—people deleting apps and abandoning years-long streaks in protest.
The companies that moved fastest are now moving backwards.
But the consulting industry didn't get the memo.
They're packaging AI-first for mid-sized companies right now. Creating FOMO environments where acting fast matters more than understanding consequences.
The result: Most AI deployments fail to deliver P&L impact.
The technology works. Nobody owns the outcome.
AI-first compresses complex decisions—value creation, risk ownership, compliance obligations, human expertise—into a single direction: move faster with AI.
That creates momentum. It doesn't create accountability.
The companies that build durable AI systems ask a different question:
"Who owns the outcome once this system is operational—when it works, when it fails, when vendors change terms, or when you need to exit?"
Not "which AI tool should we buy?"
I call this Consequence-Driven AI Integration.
It means defining accountability structures before procurement:
- Describing the problem without mentioning AI (if you can't, you don't have a business case)
- Fixing processes before accelerating them (speeding up broken workflows creates mess at scale)
- Building compliance into architecture from day one, not cleaning it up later
- Calculating the total cost of ownership including legal reviews, audits, and exit scenarios
- Optimizing for reliability with self-monitoring, correcting and human-in-the-loop mechanisms
- Defining what stays human and what gets automated before deployment, not letting scope emerge organically
- Co-creating with the people who will operate and use the system and whose performance is connected to the system
- Assigning a single owner with P&L responsibility for outcomes
Generic AI pilots skip these questions. They assume you'll figure it out later.
By then, you've accumulated sunk costs that make optimization expensive.
The companies that succeed treat AI as an accountability decision, not a procurement decision.