• Written by Azhar Aliyar, Director

A purchasing manager at a manufacturing firm is deciding this week which supplier to use for a critical component, and they currently base it on a manual review of the purchase order table and the supplier performance log. To start using agentic workflows, they need to choose a first case that will allow them to compare the results of the workflow against a known answer.

Choosing a first case

The first case should be a simple and well-defined problem that the purchasing manager is already familiar with, such as determining the best supplier for a specific component. This will allow them to compare the results of the agentic workflow against their current manual process and verify that the workflow is producing the correct answer. A good first case should also have a clear and measurable outcome, such as a list of approved suppliers or a recommendation for a specific supplier.

For example, a purchasing manager at a manufacturing firm might choose to start with a case where they need to determine the best supplier for a critical component. They would use the agentic workflow to analyze the purchase order table and the supplier performance log, and then compare the results against their current manual process. This would allow them to verify that the workflow is producing the correct answer and make any necessary adjustments before moving on to more complex cases.

What separates an agent from a chatbot

An agent is different from a chatbot in that it uses tools, maintains state, and has a goal. In the case of the purchasing manager, the agent would use the purchase order table and the supplier performance log as tools to analyze the data and make a recommendation. The agent would also maintain state by keeping track of the current suppliers and their performance, and it would have a goal of recommending the best supplier for the critical component.

This is in contrast to a chatbot, which would simply provide a response based on a set of pre-defined rules and would not have the ability to use tools or maintain state. The agent's ability to use tools and maintain state allows it to provide more accurate and relevant recommendations, and its goal-oriented approach ensures that it is always working towards a specific outcome.

Where a human approval step belongs

A human approval step is often necessary in agentic workflows, particularly in cases where the outcome has significant consequences. In the case of the purchasing manager, a human approval step would be necessary to ensure that the recommended supplier is acceptable and that the purchase order is accurate. This step would allow the purchasing manager to review the recommendation and make any necessary adjustments before finalizing the purchase order.

Removing the human approval step is usually premature, as it can lead to errors or inaccuracies in the outcome. While the agent may be able to provide accurate recommendations, it is still important to have a human review the outcome to ensure that it is correct and acceptable. This is particularly important in cases where the outcome has significant consequences, such as in the case of the purchasing manager where the wrong supplier could lead to delays or cost overruns.

What to stop doing if the comparison fails

If the comparison between the agentic workflow and the manual process fails, it is necessary to stop and re-evaluate the workflow. This may involve reviewing the data and the tools used by the agent, as well as the goal and the state maintained by the agent. It may also involve adjusting the workflow to better match the manual process, or providing additional training to the agent to improve its accuracy.

For example, if the purchasing manager finds that the agentic workflow is not providing accurate recommendations, they may need to review the purchase order table and the supplier performance log to ensure that the data is accurate and up-to-date. They may also need to adjust the goal and the state maintained by the agent to better match the manual process. By stopping and re-evaluating the workflow, the purchasing manager can identify and correct any errors or inaccuracies, and improve the overall accuracy and effectiveness of the agentic workflow.

When to seek help

If the purchasing manager is having trouble getting started with agentic workflows, or if they are experiencing difficulties with the workflow, they can seek help from the support team. The support team can provide guidance and assistance with setting up and configuring the workflow, as well as troubleshooting any issues that may arise. They can also provide additional training and support to help the purchasing manager get the most out of the agentic workflow.

What this does not do

Agentic workflows are not a replacement for human judgment and decision-making. While they can provide accurate and relevant recommendations, they are not a substitute for human review and approval. Additionally, agentic workflows are not a solution for all problems, and they may not be the best choice for every situation. For example, if the purchasing manager is dealing with a complex or nuanced problem, they may need to use a different approach, such as manual analysis or consultation with experts.

It is also important to note that most people get wrong the idea that agentic workflows are a simple solution that can be implemented quickly and easily. In reality, getting started with agentic workflows requires careful planning and configuration, as well as ongoing maintenance and evaluation. By understanding the limitations and challenges of agentic workflows, the purchasing manager can use them more effectively and get the most out of their investment.