There’s no doubt that Generative AI has the potential to significantly transform many aspects of business processes, ranging from automation all the way to strategic insights and model building. The more creative we get with technologies, the more opportunities we will create to leverage these technologies to help transform our businesses. However, combining the hype of these technologies, like generative AI, with the sole objective of showing some form of cosmetic innovation will lead to prioritizing Generative AI into use cases where it does not add as much value as some other use cases.

I was reading a white paper this morning highlighting the application of Generative AI in network design and optimization. The gist of that recommendation was that by leveraging a copilot, people can query optimization recommendations in everyday language and receive explanations they can easily understand. This use case exemplifies the force-fitting aspect I highlighted in the previous paragraph.

A copilot in analytics applications is always a great idea. It takes the domain of augmented analytics to a whole new level. However, one thing that you need to differentiate is that not every analytics application can be combined in a single bucket. There are strategic analytics applications that need a different strategy vs tactical or operational analytics applications. Network optimization is a good example of strategic analytics applications.

Assume that you are using a BI tool that has been augmented with the copilot. Last year your transportation cost for specific people was high. Using a copilot, you can ask a question along the lines of what the primary reasons behind this increased cost were. The copilot can then pull historical costs, do the difference, and identify what were the the key areas that led to this increased cost.

It can keep peeling the layers of the onion into these areas as long as it has the data to do so. For example, the fuel cost increased, and the data captures the fuel cost per asset, which is the fuel cost by every truck, then it can highlight the trucks that led to this increase in fuel cost. You will still have to chase insights that are not captured in the data, but the gist is that copilots work very well in this form of operational and tactical descriptive and diagnostic analytics.

In areas like strategic prescriptive analytics, which most network optimization tools are, the role of the copilot gets limited versus the tool’s capability. You have to remember that most leading off-the-shelf tools on the market today have already made it very hands-on and easy to use for users who are not technically highly skilled. But let us visit the scenario to see what kind of value a copilot may add in the context highlighted in that white paper.

The white paper indicated that through simple to use interface employees can query optimization recommendations in everyday language and receive explanations. If you have used a network optimization tool, you would know that getting this information will probably take few less seconds by quickly going into visualizations provided in these tools, than typing a query in a copilot. But let us say you decide to do that anyways.

While business intelligence and some other analytics tools need to be used by almost everyone working within the processes in a business organization, for organizations to become truly data-driven. However, the same is not the case for a network modeling tool. These are strategic tools that are leveraged by a select few sporadically.

The maximum you can push these tools out is into supply chain planners’ hands. And these planners can easily decipher whatever the copilot will be able to tell them, probably in less time than they would take to use the copilot to find the answers. You may argue that senior executives can use these tools as well if we empower these tools with co-pilots. Yes, but how many times a year will a VP or SVP of supply chain use this tool?

You have a model designed to find which warehouses you can consolidate. The only way a copilot can add additional value versus the visualizations and scenario insights already included in many tools is to list the top reasons that led to the model deciding to close DC X and consolidate with DC Y, which goes beyond the simple math of addition and subtraction, more like a strategic summary beyond numbers.

Typically, that means a significant post-modeling work, working in tandem with many other departments. But to be unrealistic, let us assume we want to train an LLM to learn all that.

To be effective, the LLM must learn the specific nuances of every model and scenario. That is a significant amount of knowledge that LLM needs to be trained on, and some aspects can not be easily converted into a training data format. So, even if it may be done, the critical question is what I highlighted earlier. Will it add a significant value? If you were listing essential impact areas, should it feature on that?

In prescriptive analytics tools like network optimization tools, the recommendations are not difficult to understand if you are looking to compare only in terms of numbers. The real challenge lies beyond these numbers in trying to extrapolate these recommendations to real-world nuances of these moves and factors that cannot be incorporated into a network model. The fact is that for network optimization tools to benefit from generative AI, the tool itself needs to evolve from the way it is designed right now and then needs to be coupled with Generative AI.

To effectively use Generative AI, we need to be innovative and tactical about it. If you follow me, you know that I keep on emphasizing that technology is not a stand-alone enabler. Hence, it needs to be planned from many different perspectives. Use cases like this one essentially cater to organizations’ FOMO aspects when it comes to Generative AI.

Feeding on this FOMO aspect of organizations in a predatory way will bring businesses to these companies, working on building a copilot that will probably not be capable of going beyond the numbers in the model. However, the organization implementing these crappy use cases suffers. Implementing Generative AI in ways that are not helpful means missed opportunities for the organizations. They can use those resources to leverage a large language model in the use case where it would have made a genuinely transformative impact.

In my upcoming report, “Generative AI in the Supply Chain”, I will suggest a framework that you can use to evaluate if leveraging generative AI for a specific use case, even if it can be used for that use case, is worth it or not. Because we do not need to chase every shiny technology blindly. The shiny technology may have the transforming power of the sun, but the sun’s shine needs to fall in the right spots for life to thrive in full bloom.