Automation vs. Augmentation
Where AI should replace humans, where it shouldn't — and what the Klarna reversal reveals about the future of work.
The debate "AI replaces humans" vs. "AI supports humans" is wrong. The right question: Which tasks to automate, which to augment? Klarna shows what happens when you automate too aggressively — then have to reverse course. GitHub Copilot shows how augmentation works: 46% of code AI-generated, developers 55% faster. The winners of the AI revolution are "centaurs" — humans who use AI as an amplifier, not a replacement.
The Klarna lesson: Automation has limits
At the end of 2022, Klarna had 5,527 employees. In February 2024, CEO Sebastian Siemiatkowski celebrated the biggest AI success in the industry: An AI assistant took over the work of 700 customer service agents, handled 2.3 million conversations in its first month, and was scaled to 800 FTE equivalents. Siemiatkowski called AI "capable of replacing all our jobs — mine included."
May 2025: Klarna was down to ~2,907 employees — nearly halved. Then the public reversal. Siemiatkowski admitted: AI responses were generic, customer satisfaction had collapsed, software developers and designers had to answer customer inquiries. The new model: "Uber-Style" hybrid workforce with flexible remote agents. Siemiatkowski's insight: "From a brand perspective, it is absolutely crucial that the customer knows: there is always a human if they want one."
What research shows
net new jobs by 2030 according to the WEF Future of Jobs Report 2025. 170 million new roles emerge, 92 million are displaced — 22% job disruption. Most affected roles: data entry, accounting, clerical work. Fastest growing: AI/ML specialists, data analysts, sustainability experts. McKinsey (November 2025): 57% of US work hours are technically automatable — 44% by AI agents, 13% by robots. But: 39% of worker skills will need to change by 2030.
The augmentation model: Human + AI > AI alone
The most convincing results come not from full automation — but from combination. Harvard, Wharton, and MIT Sloan studied 758 BCG consultants and identified the "Jagged Technological Frontier": Inside the AI frontier, teams solved tasks 25% faster and 40% better in quality, completing 12% more tasks. Outside the frontier, AI actively worsened outcomes. Crucially: the lowest-performing consultants benefited most; top performers saw little or even slightly negative quality effects. The NBER study confirms: AI in customer service boosts productivity by 13.8% — for the least experienced workers, by 35%.
The "Centaur Model" (named after chess teams of human + computer that beat both pure humans and pure computers): The most effective division of labor isn't "human OR AI" but "human WITH AI." GitHub Copilot: 46% of code AI-generated, but human-reviewed. Radiology: AI screening + human diagnosis = 31% fewer misdiagnoses than AI or human alone. Legal research: AI finds precedents in minutes instead of hours, lawyer evaluates relevance and strategy.
The decision matrix: Automate or augment?
Automate (AI takes over completely)
Rule-based, repetitive tasks with clear input/output: document processing, data entry, standard emails, invoice verification, spam filtering, scheduling, standard reports. Characteristics: High predictability, low variability, no emotional component. Here, full automation is right — and a human in the loop would be waste.
Augment (AI supports, human decides)
Complex tasks with discretionary judgment: strategic decisions, customer relationships, negotiations, creative work, people management, crisis management, ethical evaluations. Characteristics: High variability, emotional/social component, context dependency. Here, AI is a tool — not a replacement.
What research shows
faster — that's how much developers work with GitHub Copilot, the world's most successful augmentation example. 46% of code is AI-generated, 88% is retained in the final code. But: Copilot doesn't replace developers. It shifts their focus from typing to thinking, from implementation to architecture. The most productive teams aren't those with the most AI-generated lines of code — but those using AI strategically as an amplifier.
Change management: The underestimated success factor
80% of workers need reskilling by 2030 (PwC Global Workforce Survey 2025). But only 6% of companies have begun upskilling "at meaningful scale" — despite 89% acknowledging the need. The most successful companies counter this not with training alone, but with a cultural framework: transparency (which tasks will be automated, which won't), participation (involving employees in the AI pilot phase), quick wins (visible work relief before efficiency optimization), and upskilling (not "learn AI tools" but "strengthen decision-making competence").
In Germany and Austria, there's an additional factor: the works council (Betriebsrat). §87 Para. 1 No. 6 BetrVG: co-determination for technical systems monitoring behavior/performance — AI systems fall under this. §95 Para. 2a BetrVG: works council approval for AI-based personnel selection guidelines. Additionally: Germany's Qualifizierungsgeld (qualification allowance) pays 60% (67% with children) of net salary during AI training (min. 120h). Works councils engaged early become allies of AI transformation.
Our approach at Radical Innovators
We don't believe in "AI replaces humans." We believe in "AI frees humans for higher-value work." At Radical Innovators, we guide companies not just technically but organizationally: Where to automate, where to augment? How to communicate AI transformation internally? How to design works agreements? Our network includes not just AI experts but also change management specialists and labor lawyers.
The Klarna lesson is simple: If you automate everything, nothing human remains. And people want to talk to people — especially when it matters. The winners are companies that deploy AI where it excels, and humans where they're irreplaceable.
— Martin Kocijaz, CEO Radical Innovators