A Revolution: Organized data allows AI to achieve its potential within ERP systems

In November 2022, OpenAI publicly released ChatGPT 3.5, a tool that could craft humanlike responses to text-based inputs. The ability to “converse” with a machine was widely considered to be a breakthrough in generative artificial intelligence (AI) and thus began the so-called AI revolution.

 

The general consensus is that AI will be a transformative technology among all industries, with metal service centers being no exception. However, despite the prognostications of various technology vendors, most of what is described as AI today is primarily regular software algorithms repackaged with marketing hype as AI. In the few instances where true large language models (LLMs) have been integrated into software applications, the results have been error prone, unreliable and more akin to a minor uprising as opposed to the promised revolution.

 

 

Goldman Sachs recently released a 31-page report, “Gen AI: Too Much Spend, Too Little Benefit?” It casts doubt on the touted productivity benefits and potential financial returns while also noting the negative environmental impact from the enormous energy and freshwater demands to power the AI models.

 

Does this mean the promise of AI will never materialize and the $1 trillion of investment about to be poured in will fail to result in any quantifiable productivity gains? Quite the contrary. The investments being made today will be foundational to growth and improved performance in LLMs, but expectations must be recalibrated and AI marketing must be viewed skeptically while the technology is in its infancy.

 

In computer science, there is a commonly used term: garbage in, garbage out (GIGO).

 

This expression describes the concept that poorly defined inputs will result in an output of equally poor quality. LLMs are trained on large data sets for deep learning, a machine learning (ML) technique that uses neural networks to recognize patterns and make associations. Much of the AI interaction today is the result of LLMs being fed inputs from disorganized, unsubstantiated sources such as social media pages and satirical websites, thus resulting in inaccurate and even bizarre outputs.

 

Fortunately, in the world of ERP systems, a structured database provides the foundational element to create true value-add AI features. For example, the Invex database operates as a rules-based engine that captures over 50,000 predefined data fields and intelligently organizes and stores them for analysis and retrieval. The result is a data set of excellent quality that can be fed into a potential ML model, thus circumventing the GIGO pitfall.

 

 

AI Building true productive AI functionality begins with focusing on applications that would benefit the most from assisted decisionmaking. Examples of two such areas are inventory planning and sales forecasting. In both these areas, the Invera data science team is using structured data from the Invex database and machine learning to develop models for time series and regression analysis.

 

As an example, Invera is developing and testing models using ML to offer inventory management predictions and forecasting in the Index material resource planning (MRP) application. On the sales side, ML models are being used to analyze Invex sales history data to offer sales forecasting and predictive analysis.

 

The productivity gains from AI will take several years to manifest themselves, particularly when factoring in the resource-intensive nature and economics of deploying AI models.

 

Today, any claims of true AI functionality should be met with a healthy dose of skepticism; however, the progress of the technology is remarkable. Within a few years, the impact could be transformative.

 

In order to truly benefit from this potential, having clean, organized data sets to feed the AI models is critical. This will allow metal service centers to use AI technology to make better informed decisions using data points not previously considered and position themselves for the future and avoid having to put out the (data) garbage for collection.