Klarna automated their customer service with AI. Resolution times dropped. The metrics looked fantastic. Then the CEO publicly admitted that quality had collapsed, and they started rehiring humans.
What happened? Customer service looks like explicit work. Scripts, policies, resolution steps. But the experienced reps carried something the scripts never captured: when to bend a rule, when to escalate before being asked, when a frustrated customer needed to feel heard before they would accept a solution. That knowledge was tacit. It lived in their heads, not the manual. And nobody realized it was gone until the numbers started lying.
This is the mistake organizations keep making. They decide what to automate based on how repetitive the work looks, not on what kind of knowledge the work actually requires.
Two questions that change the decision
Every task sits somewhere on two axes. First: is the knowledge explicit (documented, rule-based, teachable in a training manual) or tacit (judgment-based, pattern-driven, built through years of experience)? Second: if the AI gets it wrong, what does it cost you?
Four zones. Each one gets treated differently.
Explicit knowledge, low cost of error. Scheduling, data entry, standard formatting, first-draft communications. Hand it over. Check occasionally. This is where most organizations start, and it is the right place to start. Just make sure the error cost is actually low.
Tacit knowledge, low cost of error. Brainstorming, early ideation, exploring design alternatives. The knowledge is hard to articulate, but mistakes are cheap. Let AI generate options. Let humans choose. The value here is range, not precision.
Explicit knowledge, high cost of error. Compliance checks, financial calculations, contract review. AI does the work. But nothing ships without a human sign-off, because the rules are clear and the penalties for breaking them are not abstract.
Tacit knowledge, high cost of error. This is the zone that gets people in trouble. Strategic decisions, complex negotiations, crisis response. AI can pull data and model scenarios, but the call belongs to a person. These are tasks where the knowledge required to be right has never been written down, and the cost of being wrong is real. This is where Klarna was operating without knowing it.
The misclassification problem
Most teams automate by ease, not by risk. They put AI on high-stakes processes without building verification. They keep humans doing low-value explicit work out of habit. Both are costly. But the most dangerous mistake is the one Klarna made: classifying tacit work as explicit because the process looked documentable.
Here is the uncomfortable truth: most of what your experienced people do is less explicit than you think. The senior analyst who "just knows" when a number is off. The project manager who senses a stakeholder problem before anyone says a word. If you automate their work based on the process manual, you are automating the 60% that is writable and losing the 40% that actually matters.
The hard part is being honest about which zone you are actually in.
The bottom line
Before you automate anything, stop asking "can AI do this?" and start asking "what happens when it's wrong, and does the knowledge live in a manual or in someone's head?" The organizations getting this right are not the ones automating the most. They are the ones who know where their expertise is tacit, where the stakes are high, and where a human still needs to make the call.