• Written by Azhar Aliyar, Director

A common objection to implementing agentic AI workflows in chemicals is that the complexity of the process and the variability of the inputs make it impossible for an agent to take effective action. However, this objection is based on a misunderstanding of what an agentic workflow is and how it can be designed to accommodate the unique challenges of the chemicals industry.

Chemicals industry challenges

The chemicals industry presents several challenges that make it harder to implement agentic AI workflows. Yield drift, where the output of a process changes over time due to factors such as equipment wear and tear, can make it difficult to predict and optimize process outcomes. Feedstock variability, where the input materials have different properties and characteristics, can also affect process yields and product quality. Additionally, plant-to-plant recipe differences, where different manufacturing facilities have different processes and equipment, can make it challenging to develop and implement standardized workflows.

These challenges require agentic AI workflows to be highly adaptable and able to respond to changing process conditions in real-time. By using data from sensors and other sources, an agent can monitor process conditions and make adjustments as needed to optimize yields and product quality.

Agent capabilities

What separates an agent from a chatbot is its ability to use tools, maintain state, and work towards a goal. In the context of the chemicals industry, an agent might use tools such as control systems and sensors to monitor and control process conditions. The agent maintains state by keeping track of the current process conditions, including the status of equipment and the properties of the input materials. The agent's goal is to optimize process outcomes, such as maximizing yields and minimizing waste.

An agent's ability to use tools and maintain state allows it to take multi-step actions, rather than simply answering a single question. For example, an agent might adjust the temperature and pressure of a reactor, and then monitor the resulting changes in the process conditions to determine if further adjustments are needed.

Human approval steps

Most people believe that a human approval step is necessary to ensure that an agent's actions are safe and effective. While it is true that human oversight is important, removing the human approval step is not usually premature. In fact, an agent can often respond faster and more effectively to changing process conditions than a human operator.

However, there are situations where a human approval step is necessary. For example, if an agent is proposing a major change to a process, such as switching to a new feedstock or changing the operating conditions of a reactor, human approval may be necessary to ensure that the change is safe and effective. By linking the agent to a human operator, it is possible to provide an additional layer of oversight and approval.

Agent failure modes

An agent fails differently from a model, in that an agent's failures are often related to its ability to take action in the world. If an agent is unable to use the tools at its disposal, or if it is unable to maintain state, it may be unable to achieve its goals. For example, if an agent is trying to optimize a process, but is unable to adjust the control systems, it may be unable to achieve the desired outcomes.

It is therefore important to log an agent's actions and outcomes, so that any failures can be identified and addressed. This might involve logging the agent's decisions, as well as the resulting changes in the process conditions. By analyzing this data, it is possible to identify areas where the agent may need additional training or support.

Limitations of agentic AI workflows

Agentic AI workflows are not a solution to every problem in the chemicals industry. There are situations where the use of an agent may not be appropriate, such as in situations where the process conditions are highly unpredictable or where the stakes are very high. In these situations, a more traditional approach to process control may be necessary.

If you are considering implementing an agentic AI workflow in your organization, it is important to carefully evaluate the potential benefits and limitations. You can contact us to discuss your specific needs and determine if an agentic AI workflow is right for you. By understanding the capabilities and limitations of agentic AI workflows, you can make informed decisions about how to use this technology to improve your operations.