Machine learning made its way into business in the early 1990s with simple, narrow tasks: assessing credit risk from loan applications, sorting mail by reading handwritten zip codes on an envelope. The pattern from those early use cases still holds today, even as the technology has become dramatically more capable. The tasks that automate well are high volume, repeatable, and governed by rules that can be made explicit. The tasks that resist automation are the ones requiring judgment, emotional nuance, or handling a situation nobody has quite seen before.
AI agents push that pattern a step further than earlier automation did. Instead of only classifying or predicting something, an agent can take an action: issuing a refund, updating a CRM record, drafting a document, escalating to a human when its own confidence drops. What follows are real, named deployments, including one very public case where the automation went further than it should have, because the failure mode is just as instructive as the success.
Customer service at Klarna: the clearest large scale case study available
In February 2024, the buy now pay later company Klarna launched a customer service assistant built on OpenAI’s models, integrated directly into its app with authenticated access to each customer’s purchase history, payment status, and account data. In its first month, the assistant handled 2.3 million conversations, about two thirds of all customer service chats, and cut average resolution time from 11 minutes to under 2. Klarna reported the assistant was doing the equivalent work of roughly 700 full-time agents, that repeat inquiries dropped 25 percent, and that customer satisfaction scores matched human agents, with an estimated $40 million profit improvement projected for 2024.
The part of the story that gets left out of most headlines is the correction that followed. By May 2025, CEO Sebastian Siemiatkowski told Bloomberg that Klarna had pushed automation too far, that cost had become “a too predominant evaluation factor,” resulting in lower quality on the harder, more emotionally charged cases: disputes, fraud, financial hardship. Klarna began re-recruiting human agents so customers could always reach a person. The task mix Klarna automated well (refunds, returns, payment plan changes) followed the same rule that has held since the 1990s: high volume, repeatable, rule-governed. The task mix that needed to stay human was exactly the part that did not fit that description.
Sales and support at Salesforce, testing its own product on itself
Salesforce runs its own Agentforce platform internally, a practice it calls being “Customer Zero.” After one year, the company reported its service agent had handled more than 1.5 million support requests on its help site, the majority resolved without a human. Its SDR (sales development rep) agent worked over 43,000 leads and generated $1.7 million in new pipeline from accounts that had gone cold, and Agentforce deployed inside Slack gave internal teams back an estimated 500,000 hours over the year by handling routine tasks automatically.
What is notable about this case is the shape of the automation: the agent is not replacing the sales or support function, it is working the portion of the funnel that was previously too low priority to staff, dormant leads, routine help site questions, and freeing people for the interactions that actually need a human judgment call.
Internal knowledge work at Moderna: hundreds of narrow tools, not one big system
Moderna’s approach looks different from the other two cases, and that difference is itself a useful lesson. Rather than building one large customer facing agent, the company gave employees access to ChatGPT Enterprise and let teams build their own narrow, purpose-built GPTs. Within two months, employees had created roughly 750 custom GPTs internally. Examples included a “Policy Bot” that answers HR and safety questions by searching internal documentation, a “Contract Companion” that summarizes legal documents in plain language for non-lawyers, and a “Dose ID” tool used to evaluate vaccine dosing options against clinical criteria, explicitly positioned as an aid to a human decision, not a replacement for one. The legal team reported 100 percent adoption, and Moderna’s Chief Legal Officer noted the tools let lawyers “focus our time on those matters that are truly driving impact for patients,” while regulatory writing that used to take weeks was cut to minutes for the first draft.
Retail at IKEA: prioritization as the actual hard problem
Ingka Group, IKEA’s largest franchisee, offers a useful contrast to the other three cases because its Chief Digital Officer, Parag Parekh, is candid about the real bottleneck not being technology. In a June 2026 interview with McKinsey, Parekh described sorting AI initiatives into a four quadrant framework, customer-focused versus coworker-focused, growth-oriented versus cost-oriented, and deliberately stack-ranking the portfolio rather than letting use cases multiply. His reasoning: the risk at this stage is doing everything and therefore accomplishing very little.
The results so far back that discipline. IKEA’s strongest ROI is still coming from a traditional AI use case that has had years to mature: goal-based order allocation across more than 400 fulfillment and shipping points, which meaningfully improved the company’s shipping cost ratio. On the generative AI side, a room design consultation that used to take a coworker about six hours and cost the customer roughly €70 now takes approximately 30 minutes per room, using a LiDAR room scan combined with a conversational design assistant. Scaling any of this across a 160,000-person organization required more than deployment, Ingka has run AI literacy programs for 40,000 coworkers so far, treating adoption as a measured outcome rather than an assumption.
The same pattern shows up across other industries too
The Klarna, Salesforce, Moderna, and IKEA cases are not outliers. IBM’s own research on agentic AI cites a Thomson Reuters legal assistant that cut average time spent on document review, research, and drafting by about a third, and reports that its internal coding agent, used by 10,000 IBM developers, increased their productivity by 45 percent. On the financial services side, a World Economic Forum analysis of agentic AI pilots at banking institutions found they collectively saved an estimated 30,000 workdays, with productivity on administrative and documentation tasks improving between 20 and 59 percent. The specifics differ by industry, but the shape of what works is consistent with everything above: a bounded task, a measurable outcome, and a clear line for where a human needs to stay involved.
A useful frame for why some tasks automate and others do not
This TED Talk from Anthony Goldbloom, founder of the data science platform Kaggle, is one of the clearer explanations of the underlying pattern behind all three cases above, and it holds up well even as the models involved have become far more capable since it was recorded.
What separates the successful pattern from Klarna’s overcorrection
Comparing the three cases side by side surfaces a consistent set of conditions:
- Clear escalation logic. Klarna’s assistant worked when it correctly identified which two thirds of tickets it could safely resolve, and the friction started precisely where that boundary was pushed further than the underlying task complexity actually supported.
- A defined, bounded action space. Issuing a standard refund or updating a CRM field is a reversible, well scoped action. Salesforce’s SDR agent working dormant leads and Moderna’s Policy Bot answering HR questions both stayed inside a narrow, well understood scope.
- Measurement that goes beyond speed and cost. Klarna’s own postmortem pointed to cost being weighted too heavily against quality as the specific mechanism behind the reversal, a direct, named example of what happens when the wrong metric drives the rollout.
- Starting narrow before scaling. Moderna’s model of many small, purpose built tools, each solving one specific job, scaled faster and with fewer failure modes than a single large system trying to cover every use case at once.
- Deliberate prioritization, not enthusiasm. IKEA’s four-quadrant ranking of initiatives is a direct answer to the same trap Klarna fell into from a different angle, chasing too much scope without a disciplined way to say no.
Frequently asked questions
What is the difference between an AI agent and traditional automation (RPA)?
Traditional robotic process automation follows fixed, predefined steps and breaks when the input varies from what it expects. An AI agent interprets information and applies judgment within defined boundaries, which is why Klarna’s assistant could handle a much wider range of phrasing and account situations than a scripted chatbot could.
Did Klarna’s AI actually replace 700 employees?
No. The “700 agents” figure was a workload equivalence, 2.3 million conversations divided by average human throughput, not a count of people laid off. Klarna’s headcount reduction over that period came mostly from a hiring freeze and attrition, not targeted layoffs tied to the launch.
Why did Klarna reverse course if the initial results were strong?
According to CEO Sebastian Siemiatkowski, the company optimized too heavily for cost and resolution speed, which held up well on routine tickets but produced lower quality outcomes on complex, emotionally charged cases like disputes and financial hardship.
Is it better to build one large AI agent or many small ones?
Moderna’s experience suggests many narrow, purpose-built tools scale faster and fail more gracefully than one system trying to cover every use case, because each tool’s scope stays small enough to reason about and govern.
How do large organizations avoid spreading AI efforts too thin?
IKEA’s approach is a useful model: rank every AI initiative against a small number of clear criteria (customer versus internal impact, growth versus cost) and deliberately concentrate on a short list rather than letting use cases multiply across the organization.
Conclusion
Across Klarna, Salesforce, Moderna, and IKEA, the businesses getting durable value from AI agents were not the ones automating everything at once. They picked specific, well understood tasks, gave the agent a clear boundary and an escalation path for anything outside it, measured more than just speed and cost, and ranked their initiatives instead of chasing all of them. Klarna’s experience remains the clearest evidence available that skipping that discipline is where the real risk lives.
Learn more:
Gen AI Studio