Klarna Reversed. Duolingo Reversed. Both Within Twelve Months. The Lesson Is Not That AI Failed: It Is That Agencies Are Pricing Their Service Wrong

In February 2024, Klarna CEO Sebastian Siemiatkowski publicly stated that AI was doing the work of 700 customer service agents. By October 2025, Siemiatkowski had reversed publicly. His direct quote: "From a brand perspective, I just think it's so critical that you are clear to your customer that there will always be a human if you want." Klarna is now hiring humans back at $41 per hour.

Duolingo announced an "AI-first" strategy in 2025 and started tracking employee AI tool usage in performance reviews. On April 20, 2026, the company dropped the AI performance tracking after employee pushback. CEO Luis von Ahn publicly reversed: "I do not see AI as replacing what our employees do."

Two of the most public AI-first business strategies of the last 18 months both ended in reversal. The Forrester Research 2026 Future of Work report estimates that 55% of employers regret AI-driven layoffs. A Careerminds survey of 600 HR professionals found that two-thirds of organisations that laid off staff for AI in the past year have already rehired some.

The standard interpretation of these reversals is that AI failed to deliver on its promise. That interpretation is wrong, and it leads agencies to draw the wrong lesson.

What Actually Happened at Klarna and Duolingo

The Klarna AI customer service deployment did exactly what it was designed to do. The AI system handled high volumes of routine customer interactions: billing questions, account status queries, payment confirmations, at scale and at acceptable quality. The cost reduction relative to the previous human-staffed model was real. The volumes the AI handled were genuine.

What Klarna discovered, twelve months in, was that the customer service agents who had been replaced were not only doing routine ticket handling. They were doing several other things at the same time, and those other things were not visible in the operational metrics that justified the AI deployment.

The agents were holding institutional knowledge about edge cases that recur infrequently but matter when they do. They were carrying brand voice in interactions where the customer's emotional state mattered as much as the technical resolution. They were managing escalations where the AI's scripted responses created more friction than they resolved. They were creating the trust signals that, in aggregate, shaped customer perception of Klarna as a brand.

When the role was reframed as "thing AI does," the surrounding functions disappeared. They had not been priced into the original analysis because they had not been visible in the operational metrics. The cost of losing them only became apparent six to twelve months later, when customer satisfaction scores deteriorated, complex cases generated more complaints, and brand perception measurably declined.

Duolingo's reversal followed a similar pattern. The "AI-first" performance tracking measured AI tool usage as a proxy for employee productivity. What the metric did not capture was the judgment work: the decisions about which AI-generated content was on-brand and which was not, which translation choices were correct in context, which gamification mechanics actually motivated learners. Employees who used less AI but applied more judgment looked unproductive on the metric while actually delivering more valuable work. The performance review system created the wrong incentives, and the company eventually backed away from the framework.

The Implication for Agency Pricing

The Klarna and Duolingo reversals are not stories about AI capability. They are stories about how organisations price work.

When work gets bundled, when one role does both visible task execution and invisible institutional knowledge maintenance, the bundle becomes vulnerable to disruption from any tool that can do the visible part. The economic logic of replacing the bundle looks attractive on the metrics that capture the visible part. The hidden cost of losing the invisible part only appears later.

Sydney marketing and recruitment agencies face exactly this pricing structure in their own businesses.

A typical agency retainer bundles many things into a single monthly fee. Strategic direction. Execution work. Account management. Reporting. Quality assurance. Relationship management. Industry knowledge maintenance. The retainer is priced as a single number, and the client buys it as a single product.

When the client decides to "move this to AI," they are making the same decision Klarna and Duolingo made. They are looking at the visible parts of what the agency delivers, the execution, the deliverables, the reports, and concluding that AI tools can produce comparable outputs. They are not seeing the invisible parts because the invisible parts have never been priced separately.

Six to twelve months later, the same client will have the same boomerang experience. The deliverables get produced by AI tools. The strategic direction wobbles. The execution quality declines on edge cases. The client experience suffers in ways that are hard to attribute back to the agency cancellation. By the time the connection is made, the agency has lost the relationship and the client has built operational habits around the AI-first arrangement.

What Agencies Should Actually Do With the Klarna Lesson

The strategic move is to break the bundle.

The agencies that survive the AI transition are not the ones that prove AI tools cannot do the execution. That fight is already lost. AI tools can do the execution well enough at the price point the client cares about. The fight that is still winnable is to make the institutional knowledge layer visible and price it separately.

This requires several practical changes to how agencies present their service offerings.

First, the agency needs to identify which parts of its current bundle are execution (replaceable by AI) and which are institutional knowledge, judgment, or relationship work (not replaceable). Most agencies have not done this exercise explicitly. The exercise typically reveals that 40-60% of the bundled retainer is execution and 40-60% is the higher-value layer.

Second, the pricing has to be unbundled. Execution work gets priced as execution work, often at lower margins reflecting the AI competition. Strategic and judgment work gets priced separately, often at higher margins reflecting the lack of AI substitution. Clients see two line items where they previously saw one.

Third, the value of the higher-margin work has to be made visible in client deliverables. If the strategic direction layer never produces a visible artefact, if it is just embedded in the execution work, the client will not value it correctly. Quarterly strategic reviews, written direction documents, and explicit decision rationale all serve to make the invisible work visible.

Fourth, the agency has to be willing to lose the execution work to AI competition while keeping the strategic layer. This is psychologically difficult because execution work is often the larger portion of current revenue. But the alternative is to bundle the strategic layer with the execution layer and lose both when the client cancels the bundled retainer.

The Boomerang Window

The Klarna and Duolingo reversals create a specific window of opportunity for agencies that have made this transition.

Clients who have implemented AI-first strategies in 2024-2025 are now in the six to twelve month window where the hidden costs are becoming visible. The customer satisfaction scores are deteriorating. The brand consistency is wobbling. The institutional knowledge gaps are creating expensive edge case failures.

These clients are open to conversations about reintroducing the human layer, but not as a return to the previous bundled retainer. They want the human layer for the work that genuinely requires it, priced at a level that reflects the value, separate from the execution work the AI tools are now handling.

The agencies that can have that conversation, that can articulate which parts of the work require human judgment and price those parts explicitly, are the boomerang partners those clients are looking for. The agencies still selling bundled retainers are the ones the clients are still avoiding.

Klarna proved the bundle is dead. The agencies that internalise that lesson before their clients explicitly state it are positioned to win the next phase. The agencies that wait for clients to make the request will discover that by the time the request is made, the client has already chosen the partner.