The AI Metabolism Gap

In February 2024 Klarna announced that its AI customer service assistant was doing the work of 700 human agents. The bot was handling 2.3 million conversations in 35 languages, and the number became the case study every consultancy referenced when they wanted to talk about what generative AI could really do. But then fifteen months later Klarna started hiring people back.

The story got read as an AI walk-back, but it wasn’t. The AI is still handling around two-thirds of customer inquiries and resolution times are 82% faster than before it shipped. What Klarna did was deploy fast and then reshape the deployment based on what they learned from how it ran. The simple stuff stayed automated, the complex cases got routed to humans. CEO Sebastian Siemiatkowski put it about as plainly as you can. “We focused too much on efficiency and cost. The result was lower quality, and that’s not sustainable.”

That’s an especially clean example of something happening more often than we might realise. Companies deploying AI, learning something specific from how it performs, and reshaping the deployment around what they found, rather than either celebrating it forever or burning it down. It’s also, in my experience, the trait that separates the businesses pulling real value out of AI from the ones running endless pilots that never settle.

The data backs this up as well. Nine in ten companies now use AI in at least one business function, according to McKinsey. But only a small fraction have moved past experimentation into something resembling sustained deployment. The gap between those two groups has very little to do with talent or budget. The vendors and models are largely the same. The published case studies are the same. What separates them is how readily an organisation can take in a new capability and adjust the work around it once they’ve seen what it’s good and bad at.

I think of this as organisational metabolism. Some businesses develop it almost without thinking. Others spend enormous energy on AI initiatives and barely move.

You can often spot it before any ROI report shows the tells. The signal is in how quickly teams move from “interesting idea” to “working prototype,” and how little ceremony sits between flagging a problem and putting something in front of it. A workflow bottleneck gets flagged on Monday and has an early-stage model tackling it by Friday. The output isn’t polished yet, but it’s running.

It also shows up in how teams talk about the work. In fast-moving organisations the language sounds like a product team in the middle of a sprint, iterative and conversational, adjusting as they go. In slower organisations AI gets discussed the way you’d discuss a multi-year infrastructure project. The conversation is large and abstract and never finishes scoping itself.

The hard part is that metabolism gets organically built into how an organisation works day to day. Clearer processes and teams that can take on new tools without ceremony are the things that let the next deployment land faster than the last one.

That’s why Syfre’s AI Readiness Workshops focus less on the tools themselves and more on the conditions around them. When the conditions are in place, the second AI deployment costs a fraction of the first, and the third costs less again. The companies that get there early are the ones doing what Klarna did, adjusting fast based on what the deployment told them, while their competitors are still on the first pilot.