What Klarna wrote on 27 February 2024
On 27 February 2024 Klarna published a press release about the AI assistant in its customer service. Every figure in it covers the first month:
- 2.3 million conversations, two-thirds of all customer service chats;
- errands resolved in under 2 minutes, against 11 before the assistant;
- 25% fewer repeat inquiries;
- 23 markets, more than 35 languages, 24/7;
- $40 million in profit improvement in 2024. That is the company’s own estimate, not a year-end result.
And the line that was retold more than any other: the assistant is doing “the equivalent work of 700 full-time agents”.
The word “fired” is not in the release. Nor is any claim that 700 people lost their jobs. An equivalent is a volume converted into full-time posts: how many people it would take to handle the same flow of chats. How Klarna arrived at 700, the release does not say. In the retelling the conversion became a layoff, and that is the version that spread. Headlines like “Klarna Replaced 700 Agents With AI and a Year Later They Were Rehiring Humans” are still being written.
Look at what the figures are about. Two concern quality: resolution time and the share of repeat inquiries. One is about money. One converts work into people. The last one is what people remembered.
May 2025: the CEO says cost weighed too heavily
On 8 May 2025 Bloomberg published an interview with Klarna’s CEO, Sebastian Siemiatkowski. The original is behind a paywall. We take his words from Customer Experience Dive, which quotes them verbatim: “As cost unfortunately seems to have been a too predominant evaluation factor when organizing this, what you end up having is lower quality.”
In the same interview he says the company’s future lies in investing in the quality of human support. Klarna started recruiting people again: through a pilot programme, with remote work and flexible hours.
Nobody dropped the AI, though. In the same CX Dive piece, company spokesperson Clare Nordstrom said the assistant still handles two-thirds of customer service chats, response times are down 82%, and repeat issues are down 25%.
The picture that emerges does not fit in a headline. The assistant kept its patch, and the company is putting money back into people for the sake of quality. The version “Klarna rehired all 700” has no source.
There is something odd here. The repeat-inquiry figure in May 2025 is the same as in February 2024: down 25%. Yet the CEO says quality fell. The match could mean a new measurement or the old number repeated. Either way, the drop in quality the CEO describes does not show up in that figure. What exactly he had in mind cannot be worked out from public sources.
CBA: 45 roles, and more calls
In July 2025 the Commonwealth Bank of Australia (CBA) announced 45 redundancies in direct banking. They were linked to a new voice bot on its inbound phone line. The Finance Sector Union challenged the cuts at the Fair Work Commission, Australia’s workplace relations tribunal.
On 21 August 2025, as ABC News reports, the bank backed down. CBA admitted that when it announced the redundancies it had not properly weighed the relevant business factors, and apologised to the staff affected. After a review, the roles turned out not to be redundant. The workers were offered the choice to stay, move to another role or leave.
What happened to call volumes is known from the union and staff. Volumes rose; management offered overtime and put team leaders on the phones. The bank itself calls the episode a mistake in its assessment, not a failure of the bot.
The mechanics here are simpler than at Klarna. A decision to cut jobs is easy to make before anyone sees the real load after launch. If the bot does not resolve the question, the customer calls again, and the queue grows. A metric of “how many posts we remove” counts people, not the queue, so the rise in calls never shows up in it.
Salesforce: 9,000 → 5,000, and no reversal
Not everyone is reversing course, at least not as of September 2025. On The Logan Bartlett Show podcast, Salesforce CEO Marc Benioff said of his support organisation: “I’ve reduced it from 9,000 heads to about 5,000, because I need less heads.” Fortune reported the remark on 2 September 2025.
According to the same article, AI agents handle about half of interactions, and support costs have fallen 17% since the start of 2025. As CX Today reports, Benioff added that AI agents had handled 1.5 million queries and humans about as many, with satisfaction scores almost the same. That is Benioff’s own claim. We have not seen any independent check.
Here a quality measure, customer satisfaction, sits next to the cut. It cannot be checked from outside. Klarna’s quality numbers also looked good in 2024, and a year later its CEO said quality had dropped. Still, Salesforce at least says what it measures.
Economists described the mechanism in 2019
Five years before Klarna’s release, Daron Acemoglu and Pascual Restrepo published a paper on automation and new tasks in the Journal of Economic Perspectives. Their conclusion: “it is not the ‘brilliant’ automation technologies that threaten employment and wages, but ‘so-so technologies’ that generate small productivity improvements.”
One of the examples they give is automated customer service. In their assessment it has displaced human service representatives but is generally deemed low quality, and so is unlikely to have produced large productivity gains.
What follows is our conclusion, not theirs. A technology like this is good enough to remove people and not good enough to make the work better. The saving appears in the payroll line. Productivity barely moves. The difference lands on the customer: they wait longer, write again, phone instead of using the chat.
Siemiatkowski’s words in 2025 describe the same mechanism, only from inside the company.
What to measure instead of saved headcount
Saved headcount is visible at once and fits on one line. Lost quality shows up later and is spread across other lines: repeat inquiries, calls instead of chats, customers who leave. That is why a pilot measured by headcount can easily look like a success in its first month.
For a support pilot we would write down at least this before the start:
- time to resolution, not time to first reply;
- the share of repeat inquiries on the same topic, say within 7 days;
- the share of conversations handed to a human, and how the handover ended;
- the customer’s rating after the conversation;
- total inquiry volume across all channels, to see whether the load has moved from chat to phone, as it did at CBA.
You can count saved headcount too, but only after these measures have held for more than a month. Otherwise you end up in the scenario from Forrester’s predictions for 2026: its analysts expect half of the layoffs blamed on AI to be quietly reversed, with the jobs coming back offshore or on lower pay. That is a forecast published on 12 November 2025, not a measurement. How to write a metric down before work starts, we covered separately.
We do not promise to replace staff. Valli, our auto-responder for Telegram Business and a website widget, is designed as a first line and a night shift. The owner stays in the loop: by default, notifications about new customer messages go to their private chat, and if they reply in a chat themselves, Valli stays silent in that chat for two hours. In a two-week pilot we write the metric down before the start and put a quality measure next to response speed, for example the share of repeat inquiries, the same figure Klarna itself reported. We have no pilot numbers of our own yet.
What we don’t know
All of Klarna’s February 2024 figures are the company’s own statements about the assistant’s first month. The $40 million is its own estimate. The Bloomberg interview is paywalled: we take Siemiatkowski’s words from CX Dive, and Fortune carries the same wording. What exactly got worse in quality, and how many people Klarna hired back, the public material does not say.
The rise in CBA’s call volumes is known from the union and staff; we have not seen figures from the bank. Salesforce’s customer satisfaction is a company claim, and we have not followed what happened to its support after September 2025. Forrester’s line is a forecast. A figure of 55% of employers who supposedly regret AI-driven layoffs is often attributed to Forrester. We could not find it at Forrester itself and do not use it.
Where an AI agent can handle support and where the conversation should go to a person is covered on AI agents for business.
If you are working out which part of support to hand to a machine, start with one process and one quality metric. We do the breakdown for free, in 48 hours, with no intro call: the questionnaire on the main page, and back comes what to take off people first and how to check it.
