Each month, our team has a presentation on US manufacturing, and we review a few key questions. This month, we asked “would you let AI make the final decision” in manufacturing? I was a hearty no, but we’ll get to that later.
What do you think? Does AI have enough experience, enough data to make that decision for you? If it was, indeed, the final decision, would you be willing to send an order out to a customer without a human reviewing it?
At its very essence, artificial intelligence relies on data to work. The more data you feed it, the more reliable it will be. (This same principal works for your shop team – read on for that). The broader the sample, the better the answer it feeds you. The danger of AI these days, as everyone tries to get in and use it, is that sample sizes, depending on what you ask it, could be limited. Let’s explore that today.
the visibility of AI
If you ask Claude, ChatGPT or any one of the LLMs a question, your sample size is large. It should be. The basic tenet of artificial intelligence is not that it’s powerful or fast – it’s that it can sample well beyond what knowledge you have. That’s powerful. Something I find most interesting that many people are completely unaware of – LLMs do not understand what you are asking. LLMs at their basic level take words and convert them into patterns represented by a mathematical number. The number that represents a word is unique – no other word can have the same. It then links words that are often seen with yours as part of the answer. So, AI can be tricked.
If you had a team of people entering erroneous information into web searches, you could shift the model. Let’s say you’re searching on manufacturing. Our team has spent the last 2 weeks (we would never actually do this) entering “purple manufacturing” into the web – searches, articles, notations, comments. Eventually, AI will pick up purple manufacturing and all of a sudden you see it everywhere. Make it one step scarier – get the AI to do the work for you and you can skip the team and just start reprogramming.
Is the same thing true in your shop? If your team is telling you that you are missing something, over-ordering or having problems, do you believe them? Maybe not the first time you hear it, but by the time 3 of your staff have started saying the same something, you start to pay attention.
That’s not data, folks. That’s a string of coincidences. You trust your gut to let you know when the things your people are saying hold validity and when they don’t. AI can’t do that. AI has to have visibility into true data – reliable information that you can believe in. It cannot Let’s go back to our question: would you let AI make the final decision?
why I say no
I say no to the LLM having decision-making authority based on the mathematical model theory above. Not because someone could game it – for an entirely different reason. In your shop, information is critical. Your team is getting work done, completing orders, moving and shipping product. Let’s say your team is moving a pallet of goods and the forklift (or the driver) has an unfortunate event. That product is scrapped. If – and that is a strong if – your AI could connect scrap with that product, it would think that your production did something wrong and thus you have a lower throughput rate to finished goods than you really do. A simple forklift accident can be the difference between real and accurate information for you and your team. Should the LLM be put in charge of how much product to make based on throughput, you are now over-producing because it didn’t have the “brain power” to know the difference between an honest forklift accident and a production problem. Information in your shop. More is more.
You might also like: Shadow AI - Why Clean Production Data Matters First