Chatbot is not the answer: Practical LLM use cases in supply chain

LLM and AI are buzzwords, but over the last several months we’ve seen examples of LLM-powered chatbots really go off the rails. So, executives are becoming cautious about the future of AI. I argue that chatbots are not a good first avenue for a company that wants to embark on the AI journey. Instead, I suggest focusing on data cleanup tasks. As a supply chain analytics practitioner, I have first-hand knowledge of the daily challenges businesses face, and I describe three specific use cases in which LLMs can make a significant impact to the company's business processes, while being very low-risk. 

Even if you went rowing across the Atlantic last year, you could not escape talks about the new technological sensation—Chat GPT. It seemed to conquer the world, promising to make thousands of jobs obsolete and force leaders to reassess business processes. Numerous companies were actively implementing it, mostly as an LLM-enabled chatbot. The idea was clear: they wanted to cut costs by making some human assistants redundant and others more efficient.

Is it too early to adopt LLMs in business, specifically supply chain? I believe that the time is right, but we should not be looking at chatbots first.

However, about 1.5 years after the initial release, there are many examples of this technology going completely rogue, giving dangerous advice and “hallucinating” (providing outright false information). Some wild examples include playing with a car dealership center’s chatbot to get it to act against the interests of the dealership. In other words, LLM is a risk, and now executives are acting with caution, adopting a “wait and see” strategy.

Is it too early to adopt LLMs in business, specifically supply chain? I believe that the time is right, but we should not be looking at chatbots first. What follows are three specific examples of immediate data cleanup use cases where any supply chain professional can start leveraging the advanced capabilities of LLMs with very low risk.

Example 1: Order notes cleanup

If we were to sum up the time that supply chain professionals all over the world spend typing back the leading zeros that get automatically removed in a spreadsheet, we would probably boost the global GDP by at least a percentage point. Jokes aside—data cleanse is an integral part of most processes at most companies.

Let’s look at a classic supply chain example where we have a list of orders with notes, and the task at hand is to extract the name of the city. Since the order notes were obviously typed in manually, there will be errors and odd characters. The make of the text field might be different for every order. What if you have thousands of those? This means weeks of work for someone who’s equipped with a traditional spreadsheet, a few hours for someone who’s proficient at ETL-tools or programming language and a few minutes for someone with a basic LLM.

Table 1: Extracting city name from order notes

Example 2: Data extraction

In the second example, we have a list of tasks that a group of people have completed over a course of time. In order to apply any kind of analysis to this data, we need to extract the name of the person. While this can be performed using traditional spreadsheet formulas, it is considerably faster to use an LLM. An example prompt could be this: “name of one of the team members: Max, Tom, Eric. Write ‘other’ if not found.” Similarly to writing a traditional formula, it is important to indicate what the function should do if there is no match.

Table 2: Extracting the name of a person from a text field

Example 3: Data matching

In the third example, let’s look at a situation where the list of references already exists, so we essentially need to find the matches of the “dirty” data names with the “clean” data names (preferably coming straight out of  the database). It is a very difficult task for the most advanced traditional tools to understand that “WLM” means “Walmart” and parse out the “noise” in the field—but a very easy one for an LLM.

Table 3: Standardizing customer names

The final example that I’d like to share is a common use case of reconciling different naming conventions for the same SKUs in two different database systems. This problem is generally referred to as a “naming convention inconsistency” or “nomenclature discrepancy.” More broadly, it falls under the umbrella of data harmonization issues, where different systems or sources use different terms to describe the same thing. While a supplier calls a certain product “Blue lemonade,” its distributor will call the same thing “Lemonade, blue.” Every company’s supply chain department typically manually maintains large tables where the relationship between these values is established. LLM-powered engines are getting us close to a point where we won’t need to maintain such tables manually anymore.

Conclusion

In this article I’ve given specific examples about how you can start leveraging LLM technology for the daily data cleanup supply chain tasks. All these examples are non-customer-facing, low-risk, and very simple to verify.

Incorporating the LLM functions in the daily work of the supply chain department will help them focus not on “how” but on “what.” The team will reduce the time spent on non-value adding activities and, instead, focus on analyzing the results, discussing  them with colleagues and coming up with recommendations. Everyone wants a meaningful job, and LLMs can definitely help us get there quicker.

About the author:

Marianna Vydrevich is manager of operations research & network optimization at GAF, North America’s largest roofing manufacturer. Vydrevich is a seasoned supply chain expert with a decade of global experience, specializing in supply chain network design and data science. She can be reached at LinkedIn or at [email protected]

SC
MR