It seems everyone is using AI to help with their work – primarily coding but apparently also other things. Given that I generate a lot more Powerpoint slides than lines of code today, it has been a bit challenging to find any application of LLMs that might actually be helpful. However, the CoPilot function that Microsoft has added to Office might be promising. Here is my recent experience using the tools.
Excel: Formula Creation
I have found CoPilot for Office to be seriously useful when doing Excel work, especially for building spreadsheet formulas. It is handy to express what I need done in natural language and get a set of Excel formulas back. This is essentially coding, and that it works well is pretty logical given the success of LLMs on coding tasks.
For example, I had to write a formula to take the text from one cell and copy it into another cell, but removing any trailing “M1” string. I got multiple proposals from CoPilot:
The picture is an example of a pattern I have seen reported from our engineers using LLMs to code: bloated expressions. In this case, the Excel TEXTBEFORE function is robust enough that you can simply use it directly with no wrapping IF. If the text does not contain “M1” it just returns the text as-is. This does assume that there is no “M1” except at the end, but I happen to know that that is the case given the nature of the data being processed.
I had another example where I had to compute out a reasonable complex formula. CoPilot took a look at my sheet, found the relevant cells, and produced a formula that worked. The result was not particularly pretty or minimal, but it was produced from a rather imprecise prompt and informed by actually reading the contents of cells in the sheet. Quite impressive, actually. Probably saved me half an hour of working out the math.
Excel: Unexpected Assistant
However, going beyond creating specific formulas to insert into specific cells, things get interesting. I have had several cases where CoPilot creates entirely new spreadsheets as answers to prompts instead of adding computations to my existing sheets.
Here is a typical case. I had two separate spreadsheets. One contained a list of sales with product codes and amounts. Another contained a list of product codes of interest. I wanted to retrieve all the lines from the first sheet where the product code matched one of the product codes in the second sheet.
I expected to get some formulas or maybe a pivot table or something like that – a dynamic system that could be re-applied to the data when either sheet was updated (that can be built in Excel with lookup functions). As mentioned above, what I got was a new spreadsheet. The assistant spent several minutes processing, to the point that I wondered if it was actually doing anything useful at all.
Any change to the inputs would require another full pass through the LLMs. Obviously extremely inefficient. It is also interesting to note that the spreadsheet was a new file that I had to download, not a new sheet inside the existing file. It also indicates how CoPilot operates in the Cloud and not in any way locally.
This behavior of CoPilot is like asking someone to do the job for you and only caring about the end result, not the mechanisms by which it was achieved or how to automate the process. I guess “assistant” is a pretty good description for what it does.
PowerPoint: Incapable Assistant
PowerPoint has been an interesting case. For some reason, CoPilot is capable of generating new content, but not of editing content. This is both annoying and surprising, since editing to me seems like a good match for the fuzziness of LLMs.
For example, I had to do a major recoloring campaign on a large set of existing slides last year as they moved from one template to another. I am very picky about colors, and I want all boxes marked “application” or “operating system” to have the same color across all slides. PowerPoint themes and color palettes are not capable of doing what I need.
Thus, I asked CoPilot to go through the slide set and apply a specific color to any rectangular box containing a certain text string. It refused. But it did offer to produce a visual basic for applications (VBA) script to do the job. Given that I wanted to do this for a whole set of different standard components, it was just too cumbersome a solution. It was faster to just progressively recolor slides as they got used.
Could it be that CoPilot cannot interpret PowerPoint files and make changes to them? They are probably more complex than Excel or Word files as they are less structured. Just guessing.
What did impress me was the ability of CoPilot to compare presentation files and give not just a list of differences, but also a summary of the overall intention and audience of the files. This actually kind of contradicts the idea that CoPilot cannot parse PowerPoint files. Or maybe this is done by having PowerPoint save each slide as a picture and then applying CoPilot to that? No idea.
Word: Slop Generator
Word is an unmitigated CoPilot disaster. Using generative AI to generate text is simply not a very good idea for a word processor.
For example, a while ago I ask CoPilot to help write a blog post about a conference I visited. Not sure what I expected as a result, but the output was ridiculous. The silly thing spit out three pages of text based on a single sentence of input! Full of platitudes and generalities, but obviously also entirely unconnected to any real knowledge or information about the event or what I saw.
Classic AI slop generation. It sounds coherent but signifies nothing.
Rant warning. I do not understand how anyone can use an AI to generate piles of text just to have text… if there is no specific intention or new information behind the text, what is the point? Why would anyone read it? The goal of any writing exercise must be to produce something to be proud of, a document with lasting value that you can come back to ten or twenty years later and say, “hey, that was pretty good”.
AI slop is the text equivalent of empty calories, consuming reader time with no nutritional value. A real text takes time to write. Drafts, reviews, and rewrites over days or weeks that consider the message, the audience, and the information to get across.
I am rather shocked by the number of people talking about this kind of slop generation as a boon to marketeers. “Writing a blog post has never been easier”… and that blog post has never been more devoid of attention as it is just a machine regurgitating previously written test. To be interesting, everything needs to be grounded in something personal, new, interesting, and if not unique, at least rare. That is what you really need. You might get numbers up with AI slop, but nothing substantive will come of it long-term.