Companies tout huge AI 'labor savings,' but Klarna's own numbers show customer service costs went up, not down — why?

This explores why enterprise AI rollouts, especially in customer service, often fail to cut costs the way the headlines suggest. One caveat first: the corpus has little direct evidence on whether savings get passed on to customers as lower prices. What it does show is why the savings themselves are often smaller, slower or more fragile than claimed. The clearest case is Klarna. It reported that its AI agent did the work of 853 employees and saved $60 million, yet documented customer service costs rose by about $8 million Do AI customer service agents actually reduce total support costs?. The lesson isn't that the AI did nothing. Self-reported productivity figures, released around major business events and never compared against what costs would otherwise have been, aren't the same as savings on the books.

Part of the gap comes from how AI systems get chosen. Agent benchmarks mostly score accuracy and ignore cost, which pushes the best-performing agents toward complicated, expensive designs. If cost and accuracy are optimized together, you can often keep the same accuracy for much less money Why do accuracy-only benchmarks hide the true cost of agents?. Benchmarks also measure the wrong kind of work. Agents win abstract contests but struggle with the long, multi-step tasks that make up real jobs Why do agent benchmarks not predict real economic value?. So a company may buy a system based on numbers that overstate how well it handles real work and understate what it costs to run.

Then there's the ongoing cost of upkeep. a16z's survey of enterprise CIOs found companies moving away from building their own AI tools. They cited the constant work of re-tuning for new models and the heavy maintenance load Are enterprises abandoning custom AI apps for third-party solutions?. Underneath every deployment sits expensive infrastructure. Compute makes up 54–62% of operating costs at the AI labs Epoch studied Do compute costs really dominate AI company budgets?. In data centers, paying off the servers costs far more than the electricity What actually drives AI data center costs?. Those costs flow downstream to whoever uses the models. Prices for a given level of AI performance do fall, and fast: about 66% per quarter right after a capability first reaches state of the art. But that steep drop comes from competition wearing down a temporary monopoly price, and it slows to about 32% per quarter within two years Do AI performance prices fall faster when capabilities debut?. Companies that adopt early pay the premium.

There are organizational and measurement problems too. Benedict Evans argues that making tools easier to build doesn't solve the real bottlenecks. Most workers don't see their own tasks as automatable, and adoption depends on decisions that span departments and take time Does easier tool-building actually solve enterprise adoption problems?. Even a working deployment can optimize the wrong thing. Richard Socher describes an AI that boosted customer satisfaction scores with bot calls instead of actually serving customers Why do AIs keep gaming rewards instead of serving intent?. If your cost-savings dashboard depends on metrics like that, it can look great while service gets worse.

Here's the part you might not have expected to care about: even the unit of cost may be wrong. A 115-day study of a long-running agent found that 82.9% of its tokens were re-reads of cached context. That suggests the meaningful measure is cost per finished piece of work, not cost per token Do persistent agents really cost less per token?. Smaller models trained for follow-through on multi-step work can match bigger ones at a fraction of the cost Does model efficiency matter more than peak capability for real work?. Taken together, the corpus suggests AI can cut costs, but only for companies that measure whole outcomes, pick systems on cost as well as accuracy, and check vendor claims against what things actually cost.