how it happened
- whoAI assistantreplaced support staff
- the gapquality unmeasuredcomplex cases suffered
- the resulthumans brought back"efficiency over quality"
what everyone assumed
The February 2024 figures were striking, and the reasoning behind them was sound. The assistant resolved conversations faster than the human queue, was available in dozens of languages around the clock, and was projected to add tens of millions of dollars in profit. If a single system could carry two-thirds of the load that well, the conclusion seemed to follow: support could be a machine, and the team behind it could be much smaller. It was one of the most quoted AI deployments of the year.
what happened
The volume held, and the volume was the headline. What the volume measured was throughput: chats opened, chats closed, queries deflected. It did not measure the texture underneath, the complex case handled adequately rather than well, the customer who got a fast answer that was not quite the right one. Those did not show up in the deflection rate. For a stretch the efficiency story stood on its own, because the metric that defined success was the one thing the system was genuinely good at.
step by step
- AI takes over customer service chatsVolume and deflection become the headline.
- the metric counts throughput, not qualityComplex cases handled adequately rather than well do not show up.the assumption breaks here
- 2025Klarna concludes quality sufferedThe CEO says efficiency was pushed too far.
- humans return for complex casesAI keeps the routine tier.
how it surfaced
The quality showed up where the throughput numbers could not see it. By 2025 Klarna had concluded that leaning so hard on automation had cost it on service quality, and it moved back toward a model that paired AI with people. Its CEO put it plainly: the company had focused too much on efficiency and cost, the result was lower quality, and that was not sustainable. AI kept the high-volume, routine tier; humans returned for the complex and high-value cases where the machine had not held parity. The headline had been replacement; the working answer was a division of labor.
A system can carry most of the volume and still not be the whole job.
The cases that do not fit the script are a smaller share of the count and a larger share of what service actually means.
picture it at a company like yours
The company here is made up. The other two paths are not: they happened, and each links to the story.
- in this storythe day it mattersCustomers with real problems get fast answers that are not quite right.what it costs youSatisfaction drops quietly, and you end up hiring people back.
- or it goes like thisit happenedMoffatt v. Air Canada, 2024howA customer-facing bot promises something your policy does not offer.what leavesmoney you now owewho ends up with itthe customer, backed by a tribunal
- or it goes like thisit happenedDPD chatbot, 2024howA bot is talked into mocking your company and the chat is shared.what leavesyour reputation, in screenshotswho ends up with ita million strangers in a day
sources
your turn
what you can do about it
this week
Before you size a team around what a tool can deflect, ask what falls outside the script and who is left to handle it well.
Somewhere in your company there is a function someone is tempted to hand to AI outright, on the strength of how much of the routine load it can carry. The volume it handles is real, and it is also the easy part to measure. The hardest cases are usually the smallest number and the largest part of the reason the function exists.
go deeper
if you're not sure where to start
Run the offboarding checklist on one leaver. It covers every account and token to close when someone leaves.
open the checklist →