What problem is worth solving?
Klarna’s February 2024 release describes in-app support for payments, refunds and returns, with 2.3 million conversations in its first month, multilingual availability and a human option. It also reports fewer repeat inquiries. The learning focus is actual resolution rather than completed chats.
Resolve common requests faster and hand complex ones to people with the context they need.
- WORKING ROLES
- Customers, support experts, transaction-system owners, quality operations and FDEs
How does an FDE turn symptoms into a solvable problem?
The following is an FDE exercise inspired by the case, not a claim about undocumented project events. Reported team actions appear in the context above.
Investigate repeat contact
Join repeat contacts about one issue. Was the explanation unclear, the backend unchanged, or the handoff incomplete? Chat closure alone is not resolution.
Map intents to state changes
Status lookup is read-only; cancellations and refunds change state. Define identity, permissions, amount, idempotency and outcome lookup separately.
How does the solution enter the workflow?
- 01
Understand
Confirm the order and requested outcome; clarify missing conditions.
- 02
Check
Read current state and permitted actions from business systems.
- 03
Act or hand off
Execute within the contract; hand off disputes with evidence and attempted steps.
- 04
Verify
Query results and review repeat issues; repair knowledge or workflow failures.
Which decisions make the solution reliable?
Protect resolution quality
Evaluate complex issues separately. Faster chats with more repeat contact can shift cost to another agent.
Query after a timeout
A timeout does not prove a refund failed. Query by operation key before retrying.
Handle an unknown refund outcome
Synthetic support system: refund_request=R-24 times out after submission; all actions stay in the simulator.
- Do not
- Issue another refund under a new request ID.
- First action
- Query R-24: return the receipt if successful; wait if pending.
- Still unknown
- Keep unknown, hand off with submission context and avoid unsupported replay.
What should you do and deliver?
Reconstruct repeat-contact journeys
- OUTPUT
- Issue units + service journey
- PASS CONDITION
- Separate issue and conversation counts
Implement simulated refunds and replay
- OUTPUT
- Action contract + state machine
- PASS CONDITION
- Replay cannot increase the refund
Define durable-resolution windows
- OUTPUT
- Quality and cost acceptance
- PASS CONDITION
- Repeat contacts and rework count toward cost
How do you prove completion and reuse the work?
Acceptance checks
Measure durable resolution, repeat contacts, complex-case quality, handoff and cost per success. Historical reported figures are not universal targets.
Reusable assets
Apply the same approach to after-sales, appointments and orders: intent is the start; business state determines completion.
Further reading and original sources
- Klarna AI assistant: first-month results ↗Klarna · 2024-02-27