Finance company Klarna learned an expensive lesson about AI agents: automate entire jobs, and you’ll likely rehire those same workers months later. After pursuing aggressive automation, the company reversed course, bringing back human employees in areas where AI agents simply couldn’t deliver.
It’s a pattern playing out across industries. McKinsey’s analysis of over 50 AI agent implementations found that while some companies enjoy early successes, many more struggle to see value from their investments—with some even retrenching and rehiring people where agents have failed. The difference between success and failure often comes down to one strategic choice: companies that succeed automate specific tasks within redesigned workflows, while those that fail attempt wholesale job elimination.
Why Task Automation Beats Job Replacement
Yet the landscape is shifting rapidly. Henrik Landgren, Co-founder and CPTO of Swedish finance company Gilion, points out that decisions made six months ago may already be outdated. “You have to talk about how you use agentic AI [AI agents] differently from how the world talked about generative AI [powered by LLMs] or machine learning,” Landgren told me when I sat down with him.
Building Investment Analysis With 82 AI Agents
Gilion demonstrates what’s possible when companies rethink their approach from the ground up. The company, which uses data to help investors and lenders make better decisions, collects four types of information: payment, marketing, accounting and product usage data. Machine learning provides 12-month forecasts with 90% accuracy. But the introduction of generative and agentic AI transformed what Landgren believed was achievable.
“What has come in the last six months with generative AI and agentic AI—oh my God—it has helped us to exponentially improve our product experience and product value to customers,” Landgren said.
The shift changed Gilion’s entire product vision. Initially, Landgren assumed the company would focus solely on quantitative data analysis, leaving investors to handle qualitative research. Generative and agentic AI eliminated that limitation.
“With generative AI, and especially agentic capabilities, we found that we can create an interactive version of the investment memo that is being built by an army of agents, which we can control,” Landgren explained. The company now deploys 82 different AI agents working as part of a comprehensive investment analysis process.
The MECE Framework for Agent Orchestration
Landgren’s approach relies on a MECE (mutually exclusive, collectively exhaustive) framework to manage this agent army effectively. “It helps to break down one big problem into a MECE framework. So every part needs to be mutually exclusive, but together they give you all the answers to provide the answer to the main question,” he said.
Each AI agent receives specific instructions for its designated task. Agents report their findings to parent managers, which in turn report upward, ultimately producing detailed investment analyses. Another agent then condenses this output into readable credit memos or investment memos in an interactive format.
This structured approach allows banks and investors to define detailed investment analysis processes tailored to their specific requirements—something impossible when trying to automate entire analyst roles.
Data-Driven Decisions Reduce Downside Risk
For Landgren, who is also an investor, the data-driven approach serves a dual purpose. While it won’t necessarily identify all positive outliers, “it will definitely help you to do that, and also to reduce the downside” of investing in poor companies by eliminating human biases.
The system has surprised even its creator. Landgren admits the data has made him reconsider investments he had become emotionally attached to, where “I loved the founder, and I loved the idea,” but the analysis revealed fundamental flaws in the investment case.
Agentic AI opens new automation possibilities by managing multiple specialized tasks rather than attempting to replicate human judgment wholesale. Startups like Gilion have designed operations from scratch around what machines do well, then added human expertise where it matters most.
The challenge for established larger companies is more complex: how to implement similar task-based automation without the cycle of cutting staff, discovering gaps, and rehiring. The answer lies in treating AI agents as collaborators that enhance human capabilities rather than replacements that eliminate them—a distinction that separates successful implementations from expensive failures.