Klarna’s ‘700 people’ was workload, not layoffs

Can you report that “we cut the work of 700 agents”?

When reviewing a customer-support AI proposal, attention-grabbing numbers tend to come first: “the work of 700 agents,” “81% of chats handled,” or “$58 million saved annually.” Put those figures on one line and it is easy to conclude that AI replaced people and that much payroll disappeared.

But Klarna’s official announcements and SEC filings say something different. The 700 figure was not the number of people actually laid off or rehired; it was an FTE-equivalent estimate describing the work Klarna said its AI handled, equal to 700 full-time support agents. The original announcement did not disclose a specific conversion formula. prnewswire

Public materials do not establish how many people Klarna actually let go because of AI, or how many it later rehired. So the value of this case is not a twist that “AI was rolled back,” but how to read FTE equivalents, company-estimated savings, ledger-recorded spending reductions, and service quality as different numbers.

“700 people” is FTE-equivalent workload, not headcount

In February 2024, Klarna announced that its AI assistant handled 2.3 million conversations in its first month. It described that as two-thirds of customer-service chats and work equivalent to 700 full-time agents. However, it did not explain how it derived 700, such as agent throughput or productive hours per agent. prnewswire

A subsequent Form F-1 gives separate, newer figures and a calculation basis. It says that after launch, AI handled 62% of customer-service chats and performed work equivalent to more than 800 FTEs. This estimate of more than 800 was calculated from how much the average monthly volume of chats and phone conversations handled by full-time agents declined after the AI launch. That methodology cannot simply be applied retroactively to the original 700 figure. sec.gov

In its original announcement, Klarna said repeat inquiries fell 25%, resolution time dropped from 11 minutes to under two minutes, and customer satisfaction was comparable to human agents. But it did not disclose the absolute repeat-contact rate or denominator, case mix, or the satisfaction sample and control group. These can be signals of quality improvement, but not interpreted as independently run randomized-experiment results. prnewswire

$40 million, $39 million, and $58 million are different figures

The $40 million in the first 2024 announcement was an expected profit improvement for that year. Later, in its Form F-1, Klarna stated that the AI assistant generated about $39 million in cost savings in 2024. It did not disclose the formula or whether model, integration, operations, and quality-assurance costs were included. So $39 million is also a company-disclosed cost-savings estimate, not necessarily net ROI verified in the ledger. prnewswire +1

Klarna’s Q3 2025 investor materials list 28 million conversations since launch, an 81% share of chats handled, 853 FTE equivalents, and estimated annual savings of $58 million. The $58 million was an annualized estimate based on average monthly reductions in human work and human chat-handling costs observed from October 2024 through September 2025. It should not be simply added to the 2024 $39 million or presented as actual cash savings. sec.gov

Connecting people does not mean the AI strategy was abandoned

The option to choose a human agent is not easily understood as a feature restored later. Klarna stated from its first 2024 announcement that customers could choose a live agent if they wished. prnewswire

In a May 2025 comment letter, the SEC cited concerns about AI service quality reported in outside articles and Klarna’s human-agent policy, and asked it to enhance disclosure on how its AI approach had changed as experience accumulated. This document did not independently verify worse quality or provide numbers for new hires or rehires. sec.gov

Automation rates and human agents are not opposites. You can design together which work AI handles alone, which work people handle, and the conditions for handoff. In that design, handoff rate is not automatically a failure metric to minimize; it is a quality measure that prevents risky answers and unfinished handling.

Calculate your first ROI row yourself

You do not need to install separate analytics tools. In a browser, open the AI Automation ROI Benchmark Report 2026 to see how expected savings, annualized savings, recovered capacity, and actual cost reduction differ. The report is public and requires no separate authentication. Then sign in to your current spreadsheet and create a blank sheet. alicelabs

The figures below are hypothetical inputs to explain the calculation format, not Klarna results.

1. Fix one task and its completion condition.
   For example, choose “payment-dispute handling” and define completion as delivery of a final response that complies with policy. If case type and difficulty differ before and after rollout, split by type or normalize to the same mix.
2. Define the path and final outcome separately.
   For each inquiry ID, record one handling path: AI only / human from the start / handed off from AI to human. This is separate from whether it was resolved. Record the final outcome separately as resolved / unresolved. The three paths and the two outcomes must each sum to total intake. Sum time directly spent by people and distinguish it from wait time.
3. Guarantee seven days of observation for the repeat-contact cohort.
   For each completed case, retain the original inquiry ID and completion time, and include only cases with a full seven days after completion in “completed cases with complete 7-day observation.” Count a repeat contact only when the new inquiry’s original ID links to that cohort, and count original IDs with repeat contacts without duplicates. Exclude cases completed near the end of the measurement period, before seven days have passed, from both denominator and numerator.
4. Enter before-and-after data in two rows.
   The hypothetical example below compares status at the end of each eight-week intake period, then observes cases completed by then for a further seven days. The first table covers handling paths and effort time; the second covers resolution outcomes and repeat contacts. “AI-only resolved” is a subset of resolved cases, so do not add it again to total resolutions.

In the hypothetical example, human effort per case before rollout is 2,000 hours ÷ 10,000 cases = 0.2 hours. Assuming equal workload and case mix, recovered time is 0.2 hours × 10,000 post-rollout cases − 700 hours = 1,300 hours.

At 40 paid hours per week, one agent has 8 weeks × 40 hours = 320 hours over eight weeks, so the FTE equivalent for the same period is 1,300 hours ÷ 320 hours = about 4.06 FTE. If you use net productive hours excluding leave, training, and meetings as the denominator, document the deduction formula and apply the same basis before and after. 4.06 is recovered workload, not a claim that 4.06 people were laid off.

5. Keep every ratio’s denominator in a cell.
   The AI-only resolution rate is 6,000 ÷ total intake 10,000 = 60%, and the human handoff rate is 3,000 ÷ AI-first intake (7,000 + 3,000) = 30%. For repeat-contact rate, use only completed original cases observed for a full seven days: 970 ÷ 9,700 = 10% before rollout and 630 ÷ 9,000 = 7% after. The numerator is the number of original inquiry IDs in that cohort with one or more repeat contacts. Count an original only once even if it has multiple repeat contacts, distinguishing this incident rate from a count that can exceed 100%.
6. Separate FTE from money.
   Next to FTE equivalents, record actual layoffs, avoided hiring, and reduced outsourcing costs in separate rows, linked to HR records, contracts, and ledgers. Put projections, annualized figures, company-disclosed estimates, and ledger reductions in separate columns too. alicelabs +2
7. Review net impact and quality together.
   Subtract model, integration, operations, and QA costs from ledger-recorded payroll and outsourcing reductions, and review unresolved cases, repeat contacts, human handoffs, and quality samples side by side.

If you want to dig deeper

Klarna AI assistant handles two-thirds of customer service chats in its first month shows the original wording for 2.3 million conversations, work equivalent to 700 people, resolution time, and the $40 million projection. prnewswire.com

Klarna Group plc Form F-1 explains that the later estimate of more than 800 FTEs was based on the decline in full-time agents’ average monthly chats and phone conversations, and includes the 2024 cost-savings claim. sec.gov

Klarna Group plc Q3 2025 presentation provides the observation period and calculation basis for the 853 FTE and $58 million annual estimate. sec.gov