Artificial intelligence (AI) has evolved and improved significantly recently, highlighting its ability to serve in conversational and task-oriented capacities. As large language models (LLMs) have become more commonplace, their daily applications have expanded. The traditional use of LLMs for recognizing and generating text has now been adapted to work alongside other LLMs, improving efficiency and use cases. 

LLM agents are an emerging area of this technology that expands on LLM capabilities, creating an LLM that can use tools, systems, and other LLMs to improve response capabilities. Applications of LLM in retail have enhanced retailers’ ability to work efficiently, research and predict market trends, optimize operations, and enhance product demand forecasts. Retailers should continue to explore this emerging technology for potential improvements to their customer or employee experiences. 

Features of a LLM Agent in Retail 

A LLM agent is an LLM that leverages additional features to differentiate itself from traditional language models. These features include memory, plan formulation and reflection, and additive tools. Additionally, the agent’s behaviors to stimuli (user interaction, task planning, etc.) can be influenced through prompts or instruction, commonly known as a persona, which can help define how the agent uses the tools it has.  

Memory 

Memory assists the agent in holding a record of completed tasks. Short-term memory involves the current conversation and critical details, but long-term memory holds lessons learned to improve future interactions.  

The benefits of memory include:  

Becoming conversational. Agents can retain crucial details from the current conversation and reference previous conversations to inform its responses. An example of this kind of agent is eCommerce chatbots. Memory allows the chatbot to follow the immediate conversation and rely on past conversations to navigate appropriate responses. 

Improving repeated tasks. As agents learn from past interactions, they become more effective in completing repeat tasks. For example, asking an agent to run a report on upcoming fashion trends across continental markets may not produce the outcome you were ideally expecting. However, the agent can improve the output based on feedback provided by user interaction or conversing with other LLMs to better understand what you may have been expecting. 

However, as an agent stores more information in its memory, it’s possible that it may hallucinate incorrect or inaccurate information. Due to possible hallucinations, it’s important to always check the output and ensure that the information provided is accurate.  

Plan Formulation and Reflection  

Plan formulation within an LLM agent can help solve the presented problem without requiring additional input from the user. The LLM creates step-by-step plans, breaking large tasks into smaller sub-tasks and allowing more accurate and complete outputs. This also lets the LLM agent use additive tools (more on that later) at different stages of executing the plan, creating new capabilities and increasing accuracy. 

For example, say a retailer wants to order hoodies in prepacks that could be available across 7 sizes and 6 colors. The merchant has bought this type of item before but isn’t sure how to optimize the size-color breakdown. With plan formulation, a simple prompt from the merchant with the product, colors, and sizes could be broken down by the LLM agent to: 

1. Research sales trends and previous season results for similar products 
2. Research sales trends and previous season results for these colors 
3. Research sales trends and previous season results for these sizes 
4. Combine the above and come up with a recommendation for the size/color breakdown of the prepacks 
5. Create a report for the results and recommendation  

That said, plan formulation allows a user’s simple request to be transformed into a plan that the LLM executes. 

After executing the plan, LLM agents can review their actions, get feedback, and learn how to improve. This is called plan reflection. Feedback can come from the agent itself as it learns or from external sources like other LLMs or human interactions. 

Expanding on the example above, after the LLM executes the plan and provides the report, it may ask, “how did I do?” The merchant could point out that the LLM should look for localized purchasing patterns or expand the color comparison to similar hues. The next time the LLM agent creates the plan, it could incorporate that feedback and add those steps to the output. 

Additive Tools 

Additive tools are external resources that the agent connects with such as subject-specialized databases, other more domain-specific LLMs, APIs, and more. Additive tools increase the functionality of the agent by providing it with separate capabilities that can be used to respond to prompts. When prompted, agents can use these additive tools to perform functions that normal LLMs cannot. 

For instance, a retailer could create an LLM agent with access to a weather API and a historical sales database. The store ops team could ask the LLM agent to model historical sales against the weather and predict store sales for the next week based on the previous year’s sales and weather trends.  

These additive tools can also help avoid the limitations of training data. Let’s say an LLM was only trained with data up to 2023. When asked about supply chain disruptions in 2024, this LLM, limited by its training data, would be unable to provide that information. However, if this LLM was an agent connected to a web API, it could simply look up this information and expand its capabilities beyond the training limitations. 

The benefits of additive tools for LLM in retail are far-reaching. These can allow the LLM agent to perform many functions, expanding the possibilities of what tools the LLM can be connected to. Proprietary databases, web APIs, and other LLMs allow agents to do comparative analyses and leverage specialized LLMs and external information to create, format, and return wide-ranging results. 

Getting Started with a LLM Agent in Retail 

LLM agents in retail are the latest iteration of LLMs, and they work to streamline results by coordinating with other LLMs to improve on past tasks and adjust with provided feedback. Retailers can significantly benefit from using LLM agents in internal and customer-facing use cases. Adding this improvement to their business’s lifecycle can reduce bottlenecks and problem areas. 

For help getting started, connect with our retail experts today. 

Contributions from Joseph Tang