The 2024 deflection wave, revisited
In 2024, "deflection" and "containment" took over every support vendor's homepage. Here's what the resolution data — and the walk-backs — have shown since.
In early 2024, two words took over every support vendor's homepage: deflection and containment. The pitch was clean — let AI answer the ticket, keep it away from a human, book the savings. The genre had a poster child. On 27 February 2024, Klarna announced its OpenAI-powered assistant had done in one month what a small army of people used to do.
Every deck that spring carried some version of that slide. Two years of follow-up reporting later, it is worth asking the plain question the marketing skipped: did the customers' problems actually get solved? Because deflection and resolution are not the same event — and the 2024 wave sold the first as if it were the second.
Deflection was always a cost metric
Containment rate counts the tickets that never reached a person. That is all it counts. It says nothing about whether the answer was right, whether the issue came back, or whether the customer left angrier than they arrived. Optimise it and you optimise for avoidance: a harder-to-reach human, a more insistent bot, a help centre engineered to wear the customer down. The number that never made the billboard was the one measuring the other side of the transaction.
Read those two figures together. Almost everyone tries self-service; roughly one in seven walks away actually resolved. The rest were "contained." Their problems still exist — now with added friction.
The resolution data since
To their credit, some vendors took the criticism and started reporting resolution instead of deflection. Progress — except the definitions stayed generous. Intercom's Fin, one of the more disciplined, still scores a customer who simply stops replying within 24 hours as an autonomous resolution.
Deflection tells you how many customers you kept away from a human. Resolution tells you how many you helped. The 2024 wave sold the first as if it were the second.
That 76% is real by its own definition, and it is a self-report, not an audit. Independent field aggregates across support programs put the tier-one automation median far lower — the gap between the billboard and the queue. For the full anatomy of one of these claims, see what "AI resolved 70% of tickets" actually means.
The walk-back
The clearest verdict came from the poster child itself. In May 2025, Klarna's CEO told Bloomberg the cost-cutting had "gone too far," and the company began recruiting human agents so customers could always reach a person — "really investing in the quality of the human support is the way of the future for us." The AI still handled two-thirds of chats. Klarna had simply learned what that other third was worth.
It was not alone, and the analysts caught up.
Read as a timeline, those three lines are the whole arc: a 2024 forecast of mass automation, a 2025 measurement showing most of the cuts never happened, and a 2026 prediction that half the cuts that did happen will be reversed. The org chart after AI turned out to bend, not break.
What the wave actually taught
None of this means the automation failed. Klarna's bot is productive; Fin resolves plenty; deflection is a perfectly good capacity-planning signal. The error was epistemic. The industry put a cost number on the exec dashboard, gave it a target, and called it customer experience.
Deflection tells you how many customers you kept away from a human. Resolution tells you how many you helped. When you can only cleanly measure one, measure the second — and treat any figure where a customer going silent counts as a win as exactly that: a hopeful guess, not a result. The 2024 wave did not fail because AI cannot help customers. It stumbled where it let the easy number stand in for the true one. For the fuller case against the metric that started it all, see deflection is a vanity metric.