• Written by Sonali Subhadarsini, Product Manager

A shift supervisor in a retail warehouse checks the inventory levels of a popular product, noticing that it is running low, and wonders if they should reorder or wait for the next automatic replenishment cycle. This is a common scenario in e-commerce and retail, where assortment decisions, returns, and promotion effectiveness can make or break a business.

Challenges in e-commerce and retail

E-commerce and retail present unique challenges when it comes to making informed decisions. Assortment decisions, for instance, involve choosing which products to stock and in what quantities, taking into account factors like seasonality, customer demand, and supplier lead times. Returns and their root causes are another area of concern, as they can eat into profit margins and damage customer satisfaction. Promotion effectiveness is also crucial, as it can make or break a product's success. The category manager, for example, needs to carefully plan and execute promotions to maximize sales and minimize waste.

The fast-moving nature of e-commerce and retail means that decisions need to be made quickly, often with limited data and under intense pressure. This is where agentic AI workflows can help, by providing a framework for taking multi-step actions and making decisions based on real-time data. However, most people get wrong the idea that agentic AI workflows can simply replace human decision-making, which is not the case. Human judgment and oversight are still essential in many areas, particularly when it comes to complex decisions that involve multiple stakeholders and competing priorities.

What separates an agent from a chatbot

An agent is distinct from a chatbot in that it has the ability to use tools, maintain state, and work towards a goal. In the context of e-commerce and retail, an agent might use tools like inventory management systems, customer relationship management software, and data analytics platforms to inform its decisions. It might also maintain state by keeping track of inventory levels, customer interactions, and sales data over time. The goal of an agent in this context might be to maximize sales, minimize returns, or optimize inventory levels.

A chatbot, on the other hand, is typically limited to answering a single question or providing a simple response. It does not have the ability to use tools, maintain state, or work towards a goal in the same way that an agent does. This is why agentic AI workflows are often more effective in e-commerce and retail, where complex decisions need to be made quickly and with limited data. For more information on how AsscherAi can support agentic AI workflows, visit our website.

Human approval and oversight

While agentic AI workflows can automate many decisions, human approval and oversight are still essential in many areas. The shift supervisor, for example, may need to approve a decision to reorder a product or override an automatic replenishment cycle. This is because human judgment and expertise are still necessary in many areas, particularly when it comes to complex decisions that involve multiple stakeholders and competing priorities. Removing the human approval step prematurely can lead to errors, inconsistencies, and unintended consequences.

In fact, most failures of agentic AI workflows can be traced back to a lack of human oversight and approval. This is why it is so important to carefully design and implement agentic AI workflows, taking into account the needs and limitations of both humans and machines. By doing so, businesses can reap the benefits of automation while still maintaining the flexibility and adaptability that comes with human judgment and expertise.

Failure modes and logging

An agent fails differently from a model, in that it can fail in multiple ways and at different stages of the decision-making process. For example, an agent might fail to gather accurate data, fail to apply the correct rules or logic, or fail to execute the desired action. This is why logging and monitoring are so important in agentic AI workflows, as they allow businesses to identify and troubleshoot issues quickly and effectively.

Logging should include information about the agent's actions, decisions, and outcomes, as well as any errors or exceptions that occur. This can help businesses to identify patterns and trends, and to refine and improve their agentic AI workflows over time. For support with logging and troubleshooting, contact our team at AsscherAi.

Limitations and wrong choices

Agentic AI workflows are not a silver bullet, and there are many situations in which they are not the right choice. For example, in situations where the decision-making process is relatively simple and straightforward, a chatbot or other automated system may be sufficient. Similarly, in situations where the data is incomplete, inaccurate, or unreliable, an agentic AI workflow may not be effective.

In general, agentic AI workflows are best suited to complex, dynamic environments like e-commerce and retail, where multiple factors need to be taken into account and decisions need to be made quickly. Even in these environments, however, agentic AI workflows should be carefully designed and implemented, with clear goals, rules, and oversight mechanisms in place. By doing so, businesses can reap the benefits of automation while still maintaining the flexibility and adaptability that comes with human judgment and expertise.