ChatGPT and AI

Artificial Intelligence (AI) is all the rage. It is being said that AI will completely change the way we work and play—that it will pretty much change everything in our lives.

There is a large contingent that is very excited about the gains we have already made in business applications, and potential applications for enhancing medical care. There are concerns about safety and fears that it may one day turn against us humans. Like pretty much anything else, there are pros and cons.

But I am not going to discuss any of that.

My focus in this article will be on a much smaller scale, yet still significant.

Problems with automation

First, there seems to be a tremendous amount of hype surrounding all the new ‘auto’ software technologies that will ‘completely’ automate accounts payable, or accounts receivable functions, aside from the use of programs like ChatGPT. But even in cases where the new technologies work perfectly, what are we giving up in terms of controlling our own data, and maintaining a solid understanding of what our data really means from an analytical standpoint?

Secondly, how much do these programs actually accomplish that is truly better, faster, and easier than what we may traditionally do manually? To that end, I would like to share with you the results of a number of tests that I have performed, comparing ChatGPT output for Excel projects that I have done for clients with the solution that I provided. I think you will find it very interesting.

The experiments

Sample 1

A client needed a formula to calculate the following:

1. If the name in Column E is Willie, and the payment type in Column C is CC, and the name in Column D is either Lenny or Moshe, then Column J will show 100% of the amount in Column F

2. If the name in Column E is Willie, and the payment type in Column C is CC, and the name in Column D is NOT Lenny or Moshe, then Column J will show 40% of the amount in Column I plus the amounts for parts and tips

3. If the name in column E is Willie and the payment type is ‘cash’ and the name in Column D is not Lenny or Moshe, then the formula in Column J will show a negative 60% of the amount in Column I

4. If the name in Column E is not Willie and the name in Column D is either Lenny or Moshe, Column J will be 60% of the amount in Column I

The formula I created to do this was:

=IF(AND(E3<>"Willie",OR(D3="Lenny",D3="Moshe")),I3*0.6,IF(AND(E3="Willie",C3="Cash",AND(D3<>"Lenny",D3<>"Moshe")),I3*0.6*1,IF(AND(E3="Willie",C3="CC",OR(D3="Lenny",D3="Moshe")),F3,IF(AND(E3="Willie",C3="CC",AND(D3<>"Lenny",D3<>"Moshe")),I3*0.4+G3+H3,""))))

When I presented this project with all the variables to ChatGPT, it quickly resolved the issue and came back with basically the same formula. Perfect.

Sample 2

I asked ChatGPT to protect all the cells with formulas in a given worksheet. Done. No problem. It even suggested adding the option of using a password.

Sample 3

I instructed ChatGPT to get subtotals for each type of Acct Type from column C, also based on which branch opened which type of account. Surprisingly, the AI at least partially failed. The result it gave was:

The result should have been a simple Pivot Table as follows:

Sample 4

In the following example, ChatGPT was instructed to get subtotals for each location in Column B, while also providing a subtotal based on cards (from Column C) that have four digits. There were over 3,500 records.

The simple Pivot table solution that I used showed the following:

A bit surprisingly, ChatGPT did not come up with the correct result. The first attempt looked like this:

…and the second attempt was this:

Sample 5

 The requirements of the next sample were to calculate the percentages of answers to certain survey questions; create a table with the results, and also create graphs showing the results. (There were over 1,200 rows of data.)

The instructions are as follows:

The result that I came up with looks like this:

One of the ChatGPT results was this:

The table was not quite what it should be, and the pie chart was not made correctly.

On another attempt, the result was the following:

...also not correct or very desirable.

Results

There were a number of other projects I submitted that ChatGPT could not do at all.

For instance, ChatGPT was provided with a Master list of 3,700 configurations for garages or sheds, with categories for Series, Style, Type, and Price. There were eighteen separate sheets for the various ‘Series’, and each time there was a price change for any size (such as 8x10 or 12x16), someone would need to find from the master list the correct size, series, style, and type, and enter the updated pricing.

There were a number of elements within the various sheets, which made this more confusing, but in any case, I wrote a program with a ‘one-click’ solution that automatically updated each sheet with the correct pricing in about three seconds. ChatGPT was not even able to provide a partial solution.

In another instance, a very large condominium maintenance company had an ADP payroll report with up to 1,000 employees, 60 to 70 different departments, and multiple columns with various amounts. They needed to manually add sheet tabs for every department in the master list and name each new tab, bring over all of the employee names to the corresponding sheets based on the department in which they worked, bring over four columns of amounts, find what type of work the employee did and bring over a percentage to each individual sheet next to the original amount and multiply it to get a total, then look up a GL account number from another sheet and match it up to each of the types of amounts being billed out, and create a subtotal of the various GL amounts at the bottom of each sheet.

The solution I came up with was a program that accomplished all of these tasks with one keystroke. ChatGPT could not begin to do the project.

Takeaways

In conclusion, while there are certainly instances where ChatGPT can and will come up with a perfect solution, there are many times that the solution will be either incomplete, incorrect in some way or not aesthetically acceptable, and in many instances, will not be able to come up with a solution at all.

I believe there is an inherent danger that people may come to rely on new technologies and accept whatever results are given without even knowing if there are missing elements in the results or even mistakes. Without someone who is familiar with formulas and programming in a spreadsheet environment overseeing the results, or someone carefully following the processes and results in an accounting program, the risk of getting incorrect information runs high.

There is one other point that I believe warrants serious consideration. Even in the instances where AI can provide perfect results, and assuming that the technology will continue to improve, do we really want to rely on this level of automation without really understanding the data or results of a given project? It will always be critical to understand the data and be able to analyze data on a daily basis. To rely on technology or have employees who really do not understand how to interpret and analyze data may be a grave mistake.

For more information or details, I would be glad to review with you the various projects via Zoom sometime. Please feel free to reach out to me at (305) 323-1005, or acs@computersacs.com.