Fujitsu's Self-Learning AI Agents Cut Need for Manual Expert Adjustments
Fujitsu has developed multi-AI agent technology that continuously adapts to changing business conditions without requiring specialists to manually reconfigure systems. The agents work together as a team, analyze their own failures, and adjust operations based on business results, human feedback, and policy changes.
The problem the technology addresses is straightforward: regulations shift, systems get updated, and workflows evolve constantly. Conventional AI agents execute instructions but cannot independently diagnose why they fail. When something goes wrong, experts must manually adjust prompts, search strategies, and evaluation criteria-a repetitive task that drains specialist resources.
Fujitsu's architecture flips this dynamic. The agents identify reasons for success and failure while performing tasks, extract actionable knowledge, and adapt their operations accordingly. This transfers work that previously demanded continuous manual maintenance by specialists onto the agents themselves.
28-Point Accuracy Gain in Fujitsu's Own Language Model
Fujitsu tested the technology on Takane, its business-specific language model. Multiple AI agents managed the entire optimization cycle: data selection, training parameter adjustment, evaluation, and improvement. Each agent proposed changes based on actual business results, but only proven improvements were implemented.
The result: Takane's accuracy improved by an average of 28 points compared to the pre-specialized version. The model was optimized for manufacturing, healthcare, financial services, and government sectors.
In healthcare, Takane now extracts structured information from unstructured medical records in consistent format-diagnoses, disease progression, and treatment policies. The system handles work that previously required manual data processing.
From Document Search to Autonomous Improvement
A second application shows how the agents learn from repeated tasks. Fujitsu uses the technology for document searches across design specifications for its electronic health record systems and municipal software solutions.
Previously, determining how legislative changes affected software required experts with deep knowledge of regulations, business processes, and system architecture. The new agents learn from previous search results, failures, and human corrections. They adopt search techniques-such as consulting adjacent documents or including seemingly irrelevant files from the same domain-that were once the exclusive domain of experienced specialists.
Fujitsu plans to integrate the self-learning agent technology into its Kozuchi AI platform and offer it as core technology for enterprise-specific AI development. The company is also collaborating with Carnegie Mellon University on a lighter version designed for on-premises and edge environments with limited memory and power consumption.
For operations teams managing complex workflows and regulatory compliance, this approach reduces the operational overhead of maintaining AI systems as business conditions change. Rather than waiting for specialists to adjust configurations, agents adapt themselves based on performance data.
Learn more about AI Agents & Automation or explore how AI can optimize operations management.
Your membership also unlocks:
