In February 2024, Klarna announced its AI assistant handled 2.3 million customer service conversations in a month, about two-thirds of its chats, and claimed it was doing the equivalent work of 700 full-time agents.
By 2025, Klarna announced it was rebuilding human capacity in customer service, with the CEO acknowledging quality issues after pushing too far towards an AI-heavy service model. Whatever the exact internal metric was, the external signal was that absorbing volume is not the same as holding a service standard.
That’s the real point of the story. The question is not whether AI can produce output at scale. The question is what makes AI-assisted work dependable in the moments that matter.
The mistake is treating headcount reduction as the operating model
Some tasks will be automated and some roles will shrink. Those are outcomes that most organisations will see.
But labour substitution becomes fragile when it becomes THE strategy. Because then the organisation tries to remove the very layer that currently makes probabilistic systems workable: judgement, verification, escalation and clear ownership.
AI can be fluent and wrong at the same time. So reliability is not a property of the model alone. It is a property of the surrounding system: how work is framed, what gets checked and who is accountable when it goes wrong.
Once you accept that usage will happen and errors will happen, the next question becomes sharper: What happens when you remove the people who would normally catch those errors before they become customer reality?
“Speed versus safety” is usually the wrong framing
Leaders often talk about AI as a trade-off: move fast or stay safe. In practice, that framing pushes organisations towards brittle extremes.
If you try to govern AI before people have enough lived competence, you get one of two outcomes. Either the rules block delivery or the rules are so abstract that teams ignore them.
People do not stop working because leadership is uncertain. They route around friction to meet deadlines.
The UK National Cyber Security Centre makes the same point about shadow IT: workarounds indicate that policy and user needs are misaligned. They go on to say that the fix is to refine the rules and address the underlying need so that activity can be brought above board.
To put that into context, Salesforce reports that more than half of generative AI adopters use unapproved tools at work, tied to unclear policies and guidance. So the first leadership task is not “write stricter rules”. It’s “create a safe lane that is easier than the workaround”.
Governance that holds starts with empowerment
Many AI programmes build governance artefacts and mistake them for governable behaviour. Policies, approvals, catalogues and training completion can all exist while real work stays uneven.
You’ll see one team using AI carefully, another using it sloppily and a third staying quiet because disclosure feels politically risky. That variance is where quality drops and risk goes quiet.
AI empowerment is the alternative posture. Build human capability in a real work ‘playground’ first, then lock it in through systems and governance so the organisation can rely on it.
This is the congruence point that many programmes miss: governance becomes executable only when it is embedded in how work is actually done. Until then, “governance” either becomes theatre or friction.
The unit of durable change is the workflow
Tool rollouts create activity, but they do not create organisational reliability.
If AI stays a personal productivity trick, you get pockets of excellence and pockets of failure. Effectively, capability fragments instead of compounding.
Durable change happens at the workflow level. A workflow has defined inputs, a quality standard, explicit gates and an accountable owner. That is where verification stops being an individual virtue and becomes a built-in step.
This is also where the “substitute first” narrative often collapses. Work that looks automatable from a distance often contains edge cases, judgment calls and accountability moments that only become visible when you map the workflow properly. If you automate before you understand the workflow, you scale misunderstandings instead of outcomes.
Workflow design is not bureaucracy. It is how you make AI-assisted work repeatable, defensible and safe under load.
The practical deliverable is an approved pathway
Organisations need more than a manifesto. They need an approved way to use AI that is faster than the workaround.
An approved pathway removes guesswork. It answers where AI is encouraged and where it is refused, what must be verified, who owns the output and what happens when stakes are high or uncertainty appears.
This pathway is not a guardrail bolted onto the side of work. It is the fastest route to consistent output because it gives teams a default operating framework that already includes checks and accountability.
It also restores leadership visibility. When the approved path is usable, people disclose usage instead of hiding it and managers can reinforce standards instead of improvising local rules.
This is why shadow AI data matters, but only as a diagnostic tool. If large numbers of people are using unapproved tools, the system is telling you the sanctioned path is unclear, too slow or not fit for real work. An approved pathway addresses that by meeting both the delivery need and the risk need simultaneously.
“But automation is inevitable”
Yes, some automation is inevitable. The question is whether you automate from strength or from fragility.
The sustainable sequence is simple: empower first, then automate what is stable. Stability means the workflow is understood, quality gates are explicit and owners are accountable. Until that is true, automation increases your blast radius instead of reducing your burden.
That is the operating lesson from Klarna. AI can absorb volume, but business is more than volume. It is about trust, continuity and the ability to handle exceptions without compromising quality. Those outcomes do not appear by subtracting the control layer. They appear by designing it.
The principle that survives contact with reality
The future is not “AI everywhere” or “humans everywhere”. It is whether AI-assisted work becomes dependable.
The decision rule is the one most organisations try to skip: automation is downstream of empowerment. If you cannot rely on how AI is being used by people today, scaling it will scale inconsistency tomorrow.
The organisations that do well will not be the ones that pursue headcount reduction fastest. They will be the ones that build an approved pathway, embed AI in workflows and make governance executable in lived work. That is what “AI through people” means. And it’s why empowerment beats substitution as a strategy.
