Understanding How AI Listens

By David Jobe, Bosma Enterprises Salesforce Manager

When I was younger, my engineering father told me a great many jokes, only a few of which are safe for work. One of those in the latter category, I recently heard, but instead of an engineer being the focus of the joke, it had evolved into a software programmer. Seems the apple didn't fall that far at all. Here is the joke in question. 

A software programmer and his wife are working together to get everything planned for the upcoming Thanksgiving dinner. The wife asks her husband, “Can you go to the store for me?”
The Husband says that he can.
The Wife says, "Go and get a gallon of milk, and if they have eggs, get six."
Off the Husband goes, returning half an hour later with six gallons of milk.
Perplexed, the Wife looks at the six gallons of milk. "Why on earth did you get 6 gallons of milk?" 
He replies, "They had eggs."

    Now, unless you are an engineer or a software programmer, that joke may not make sense to you. If you are a software engineer, that joke may have some real-life equivalent that may not have been a joking matter at the time. If my Chemistry teacher had seen this joke, he would have laughed and added, "This is what you get when you don't label your values!" You see, the joke is that an engineer would see this as an if-then statement. The trigger is that they have eggs, and if this value is true, the number of milk gallons required increases. It is because the wife didn't label the value of six, and the husband is of an engineering mind; he reads that as a true/false value that changes the value of the other number. See the simple code below.

GET gallon of milk
IF they have eggs THEN
GET six gallons of milk
ELSE
GET one gallon of milk
ENDIF

    Neither of the people in that joke is wrong, but they illustrate a valuable lesson we, as software designers and engineers, need to be conscious of as we enter the AI age. Are we adding data for the AI to parse without instructing it on what that data is? Even the best-built AI can't give you proper results if it doesn't know what the meaning of those numbers is. This is a lesson that others discover as they begin their foray into AI. They, of course, knew that the data needed to be clean and that the response should be verified against "hallucinations"; however, they are now discovering that they need to revisit and retrain the AI on what their own datasets actually track.

    Here is a common example that has been cited by several of the companies we contacted. You have a product, so you have a label, a cost to create, and a sales cost. Pretty straightforward. What if your product is a box of items (Bill of Materials (BOM))? Do you have a field that indicates whether the item is produced by each, by box, or even by crate? Do you know how many of each item are within those boxes or crates? Do the raw items come in different quantities? Unless you are tracking all of that, the AI will assume a like-for-like comparison and may get the results wrong. If one supplier gives you 150 raw items at a cost cheaper than you get 200 of the same items from a different supplier, but the 200 equals less per item, the AI might still choose the 150 vendors simply because it thinks the values are the same otherwise. That choice would ultimately cost your company more simply because one value wasn't included in the equation. Now imagine this happening all the way down the line because some data isn't being tracked or isn't being included. Or, is there, but not documented well enough for the system to understand what that value represents. 

    If you are currently working to prepare your data for communication with a form of AI, take it from those who have already started and encountered the first stumbling block. Ensure that all your data is clean, and also verify that all valuable data points are being tracked and included. Then, ensure that all your metadata is properly labeled and thoroughly documented. It may take some time to update your systems with that data, but it will be more beneficial and cost-effective now than later, after you discover you are paying twice as much for your widgets as you could, simply because one value wasn't included in your AI's calculations. 

    If you have more questions about AI or how Bosma is using Salesforce to keep its business running smoothly, please feel free to reach out to bosmaconsulting@bosma.org. We would love to help.