The interesting agentic implementations are not necessarily the most autonomous. They are the ones that change a real workflow and report a result we can inspect.
Three cases illustrate different forms of leverage: customer-service resolution at Klarna, software delivery at Rakuten, and advisor follow-through at Morgan Stanley. They also illustrate why a faster task is not automatically a higher-margin business.
Klarna: service capacity, with a quality warning
In its February 2024 announcement, Klarna reported that its AI assistant had handled 2.3 million conversations in its first month—two-thirds of customer-service chats. It described the workload as equivalent to 700 full-time agents, reported a 25% reduction in repeat enquiries, and said resolution time fell from 11 minutes to under two minutes. The assistant covered tasks including refunds and returns. These were Klarna's own figures, not an independent audit. 1
The same announcement estimated $40 million in profit improvement for 2024. An estimate is not a realised financial result. Work “equivalent” to 700 employees is not proof that 700 jobs were eliminated or that their full cost was removed.
The later context is essential. In September 2025, Reuters reported that Klarna's chief executive said the company had gone too fast with AI and was shifting the emphasis from cost cutting to growth. The simple story of permanent human replacement misses the subsequent reassessment. 2
What matters operationally: resolution time and repeat contacts belong together. A shorter interaction is useful only if the customer gets an acceptable outcome. Keep a reliable route to human service and measure quality alongside cost.
Rakuten: an agent that works through a coding task
Anthropic's Rakuten case study reports seven hours of sustained autonomous coding on a complex open-source refactoring project. It also reports a reduction in time to market for new features from 24 days to five days—79%. These are results reported in a vendor-published customer story, not a controlled study across all of Rakuten's engineering work. 3
This is closer to the practical meaning of an agent: a system working through a multi-step task, rather than merely suggesting the next line of code. But a long autonomous run is not, on its own, the economic result. The relevant outcome is usable software after testing, review and integration.
Faster delivery could let a team release more improvements with the same resources or bring a revenue-producing feature to market earlier. Neither follows automatically. The public case does not establish a 79% reduction in engineering expenditure, a similar increase in revenue, or a particular improvement to operating margin.
What matters operationally: measure accepted changes, review time, defects and production outcomes—not generated code or hours the agent remained active. Give the system a bounded task and a testable definition of completion.
Morgan Stanley: a useful contrast to full autonomy
Morgan Stanley's Debrief summarises advisor meetings, identifies action items, drafts follow-up emails and saves notes. The advisor reviews and edits the output. In the firm's announcement, one advisor team estimated it saved about 30 minutes per meeting on note-taking. That is an attributed user estimate, not a firmwide controlled time study. 4
This is an assistant workflow, not evidence of an autonomous financial advisor or a system executing trades. It is included because the boundary is instructive: leaving judgement and client communication with a person can still remove substantial administrative work.
The result may be more time with clients, faster follow-through or better records. Those are valuable possibilities. They are not proof that payroll fell or that margins expanded.
What matters operationally: a human approval step is not a failure of agentic design. In a regulated or high-trust workflow, it can be the appropriate operating boundary.
Read the result at the right level
There are three different claims an implementation can make:
- Workflow result: fewer minutes per task, more resolved cases or faster accepted delivery.
- Operating result: lower cost per correctly completed outcome or more output from a stable cost base.
- Financial result: improved contribution or operating margin after deployment and oversight costs.
The cases above provide useful evidence at the first level. They do not establish a clean, independently audited causal bridge to the third. Mixing the levels creates a stronger headline and a weaker investment decision.
For an owner or operator, the lesson is to borrow the workflow, not the headline number. Start with a process you understand. Define quality, permissions and escalation. Establish a baseline. Then ask whether the gain survives the full cost of running the system.
These are external examples, not Rebellion Capital portfolio results. Metrics refer to the reporting periods and scopes of the cited publications; they are not claims about current performance.