Unlocking Agents Powered Automation Excellence with FMEA and Jidoka

Featured article by Jayachandran K R

Toyota Production System (TPS) is one of the highly acclaimed methodologies that revolutionized the automobile industry. Over the years, TPS and its avatar such as lean manufacturing has been adapted for many industry verticals. The origin of TPS is based on the philosophy of “making someone’s work easier.” The two critical pillars of TPS are Jidoka (automation with a human touch) and Just-in-time.

It all started with Sakichi Toyoda closely observing his mother working on manual looms to make fabric. His objective was to bring some intervention which would bring some level of automation into the process. This led to the invention of Power Loom that are equipped with a weft-breakage automatic stopping device. While the fabric is being made, whenever the thread is over or is broken, the machine would stop, and the operator would intervene before restarting the process. Though there was some level of automation, it required a person to monitor the machine so that the thread can be provided to the machine, and the process can be continued. He brought in more automation with the invention of the Type-G Toyoda Automatic Loom, the world's first automatic loom with a non-stop shuttle-change motion. This reduced continuous manual supervision of the machines and that led to improved productivity. These inventions were the bedrock of Jidoka which emphasizes the concept of “stopping immediately when abnormalities are detected” to prevent defective products from being produced and improving productivity to eliminate the need for people to be simply watching over machines. It all started with “making someone’s work easier” and productivity was a byproduct of that.

The second pillar of TPS is Just-in-time concept that was pioneered by Kiichiro Toyoda, Eiji Toyoda and Taiichi Ohno. It helped in revolutionizing the manufacturing process by “Making only what is needed, only when it is needed, and only in the amount that is needed”. Consistent and continuous application of these methodologies eventually led to the creation of the world acclaimed TPS process.

Failure Mode Effect Analysis (FMEA) is another key approach adopted widely in Manufacturing. It is a systematic and proactive approach for identifying, assessing, and mitigating potential failure points in a product, process, or system before they occur. The roots of FMEA can be traced back to the mid-20th century, particularly in the aerospace and military sectors. During World War II, engineers and analysts were faced with the challenge of ensuring the reliability of complex systems such as aircraft, vehicles, and weaponry. The U.S. military played a significant role in the early development of FMEA. It involves a structured team-based approach to:

• Identify potential failure modes: What could go wrong?
• Analyse their potential effects: What are the consequences of each failure?
• Assess the severity (S), occurrence(O), and detection(D) of each failure mode.
• Calculate the Risk Priority Number (RPN): RPN = Severity X Occurrence X Detection High RPN values indicate the need for urgent action.
• Prioritize and Address Risks: Focus on failure modes with the highest RPN. Develop actions to eliminate or mitigate risks.
• Develop and implement corrective actions.

We are currently in the? age of AI and Agentic AI will transform businesses at an unprecedented scale. Agents, our new coworker is entering the workforce at a rapid pace. Large scale automation led by robots have automated manufacturing processes. There is a lot we can learn and apply from there to Agents powered automation of the future. The following are some of the learnings from Jidoka and FMEA to the Agentic AI World.

Applying FMEA for Agent-Powered Automation

AI Agents typically consists of multiple modules such as perception, planning, tools, memory, reasoning, interaction, feedback mechanism, learning, self-reflection, action, adherence to guidelines etc. Failures can happen in any of the modules in an agent. Failures can be due to invalid tool selection, invalid input, invalid data format, inefficient or delayed processing, wrong LLM output, hallucinations, security vulnerabilities, algorithmic bias, ethical concerns and many more. Even if one of the modules gives an invalid output, the agent will fail to provide expected results. If we are planning to include multiple agents for automating a workflow, then it can compound the error. For example, if there are 10 agents in a workflow and if each agent is 95% accurate, then the accuracy of overall workflow is only 60%! This may not be an acceptable outcome for enterprise automation. This is where techniques like FMEA are immensely helpful. Each failure mode in an Agent is assessed for its severity, occurrence and detection and a Risk Priority Number is computed so that the high-risk failure modes can be prioritized for action.

Applying Jidoka for Agent-Powered Automation

Jidoka, or "automation with a human touch," emphasizes quality, continuous improvement, and human oversight in automated systems. It ensures processes are stopped when an abnormality is detected so that defects are prevented from flowing to the next process. Following are key best practices and their application to agent-led digital automation:

·?????? Detect and Stop Abnormalities: In Manufacturing, machines stop automatically when they detect errors, or operators stop processes manually to address issues. On similar lines, build anomaly detection into agents to flag and halt operations when data or process irregularities occur (e.g., workflow errors, invalid inputs). Apply simple, low-cost Poka-Yoke (foolproof) mechanisms that prevent errors from occurring in the first place. This would prevent cascading errors and ensures process reliability.

·?????? Real-Time Monitoring and Alerts: In manufacturing, Visual Signals or Andons (e.g., lights, alerts) notify operators of abnormalities in the process. It enables quick intervention and minimizes downtime. On similar lines, create notification systems and dashboards to provide real-time updates on agent performance. This would enable quick intervention and minimizes service disruption.

·?????? Continuous Improvement (Kaizen): Regularly analyse agent performance data to identify areas for model retraining or process optimization. This would enhance the accuracy, efficiency, and adaptability of AI agents.

·?????? Human-in-the-Loop (HITL): Integrate human oversight into critical decision-making processes where the agent's confidence is low. Conduct regular audits, perform quality control, and provide feedback to agents for continuous learning. This further reinforces the concept of “automation with a human touch.”

·?????? Root Cause Analysis: Automate logging and diagnostics for agents to identify root cause of failures or inefficiencies. This reduces recurring errors and supports long-term reliability.

·?????? Build for Scalability and Maintenance: Ensure agents are designed with modular, reusable components to facilitate updates and for scaling.

Establish Clear Success Metrics:

To effectively monitor and improve agent-powered automation systems that use FMEA and Jidoka, it is crucial to track relevant metrics such as:

• Performance Metrics: Accuracy: How often does the agent make correct decisions? Efficiency: How quickly and efficiently does the agent complete tasks? Throughput: How many tasks can the agent process within a given period? Error Rate: How often does the agent make mistakes or produce incorrect outputs?
• System Reliability Metrics: Mean Time Between Failures (MTBF): How long does the system operate without experiencing a significant failure? Mean Time To Repair (MTTR): How long does it take to identify and resolve a system failure? System Availability: How often is the system available for use?
• Human-in-the-Loop Metrics: Human Intervention Rate: How often do human operators need to intervene to correct agent actions or resolve issues? Human Override Rate: How often do human operators override agent decisions? Human Feedback Rate: How often do human operators provide feedback on agent performance?

By regularly tracking these metrics, organizations can identify areas for improvement, refine agent-powered automation systems, and ensure that they are operating effectively and reliably.

In summary, FMEA helps to identify potential problems before they occur, while Jidoka provides a framework for building quality and responsiveness into the agent-powered automation system. By combining FMEA and Jidoka in building Agentic Systems, organizations can minimize the risk of failures, ensure quality and reliability, promote continuous improvement, drive operational excellence, build transparency, trust, and confidence. Let us imbibe the spirit of TPS and apply agents for “making someone’s work easier” and reimagine the enterprise of the future.