When AI Replacement Goes Wrong
Klarna fired 700 people. Shopify's partners called it a "dumpster fire." Duolingo removed the human soul. Then came the u-turns.
The Problem Everyone Saw
"We need to cut costs with AI."
In 2024, the mandate was universal: deploy AI to reduce headcount, automate cognitive labor, look modern to investors. Klarna's AI handled the work of 700 agents. Shopify funneled support through bots. Duolingo fired its translators.
The metrics were seductive. Costs plummeted. The AI was available 24/7 in 35 languages. Leadership declared victory.
The Problem That Was Actually There
Category Error: They treated Complex problems as Simple ones.
The Cynefin Diagnostic
| Domain | Example | AI Fit | What Went Wrong |
|---|---|---|---|
| Simple | Password reset | High | Nothing — this worked fine |
| Complicated | API debugging | Medium | Shopify's AI hallucinated technical policies |
| Complex | Calming an angry customer | Low | Klarna treated financial anxiety as data query |
| Chaotic | Crisis management | Zero | The AI made it worse by looping |
AI is the engine of average. It produces the most likely, mediocre response. Great service is an outlier — a support agent cracking a joke or bending a rule. These companies "averaged out" their service quality, eliminating the peaks of delight that create loyalty.
The Unexpected Solution
The Centaur Model: Human + AI beats AI alone.
Klarna's Correction
Scrambled to rehire human agents. The AI now provides the memory (recalling every transaction) while humans provide the conscience (knowing when to bend rules).
The Sort That Should Have Happened
- Simple tasks: Full automation ✓
- Complicated tasks: AI assists, human decides
- Complex tasks: Human only, AI banned
The Feedback Loop They Severed
Support isn't a cost center — it's a signal. By blocking the path between customer pain and product teams, they lost the early warning system.
What You Can Steal
1. Run the Cynefin Sort Before You Automate
Break every role into atomic tasks. Apply the domain test. Only automate Simple. For Complex tasks, design an immediate escalation to humans.
2. The Pre-Mortem
Before launch, gather the team and say: "It's one year from now. Our AI support has been a disaster. Why?" Generate the list of reasons. Build guardrails for each one.
3. Measure Augmentation, Not Deflection
Stop measuring how many tickets the bot "handles." Start measuring how much more value your humans create because the AI is helping them.
4. The Mom Test for AI
Don't ask "Would you like an AI assistant?" (They'll say yes). Ask "Tell me about your last frustrating support experience." The answer reveals whether AI will help or hurt.
5. Keep Humans on Genesis Tasks
Use Wardley Mapping. Automate Commodities (data retrieval). Keep humans on Genesis (cultural nuance, creative problem-solving). Duolingo automated their Genesis — and lost their soul.
The Verdict
The path forward is not to rollback AI and return to 2019. It's to advance to a Human-in-the-Loop architecture where AI handles the drudgery of the Simple domain, and humans are unleashed to master the nuance of the Complex domain.
True innovation is not the replacement of the human; it is the elevation of the human to higher-order problems, supported by the infinite leverage of the machine.
Thinking about AI automation?
Don't make the same mistake. Let's run the Cynefin Sort on your operation first.
Start the ConversationSources
This analysis draws on reporting from CX Today, The Economic Times, Reddit community threads, McKinsey, Brookings, and company filings from 2024-2025.