Bringing AI to the table

It’s hard to believe that Chat GPT was launched just over 18 months ago. So much has changed in that time. Looking back at the early days of large language models, it was like an alien intelligence had taken centre stage and started shaking the foundations of our knowledge, making us question what it means to be human. Even I dug deep into my memories of seeing Nick Cave in Chile to make sense of it all…

Things have certainly changed. We are still debating what it means to be human which I expect will continue long after large language models are replaced by another technological leap. The difference is that we are beginning to understand how best to use these models and discover what they are good at. Perhaps the most lucid guide into this alien mind is Ethan Mollick who lists 4 rules for getting the most out of AI in his book Co-Intelligence – Living and working with AI:

1. Always bring AI to the table
2. Be the human in the loop
3. Treat artificial intelligence as a person, but tell it what kind of person it is
4. Assume this is the worst AI you will ever use

Perhaps the most actionable rule is #1 which is about trying things and discovering what works. At Knowledge Industries, we’ve had Microsoft Copilot for a few months already and have benefited from summaries of Teams calls and intelligent searches of our files. We’ve also used it for PowerPoint presentations and to summarise documents and reports. These are all useful and have provided a good return on the investment in Copilot, but I haven’t felt the excitement of trying something new, of truly bringing AI to the table until last week.

I had to produce a spreadsheet with a list of clients and their software renewal dates to send to a partner. My source data had a column with the dates for last year and another column with the name of the organisation and some description of the software and notes in the cell. Easy enough to do by hand, but this is the kind of thing I like to do with an Excel formula and refresh my skills in the world’s most popular programming language

I have done this many times before, but I can’t quite remember how. I know I am looking for a text function that can remove all text from the word “software” until the end of the text (lucky the source data had this pattern). Following my normal workflow, I open google and type in “remove everything after string in Excel” and after several pointless ads and cookie pages I get to the formula I need which I then copy into Excel and start modifying for my specific purposes. At this point I remember Ethan Mollick’s rule #1 and decide to apply it. Let me just ask the AI to do this for me. I open the copilot prompt and write:

"Create a new column with the data of Description with all the text starting from the word software removed"

After waiting about 30 seconds for the processing to be done, Copilot came up with a suggestion:

I like how it explains what it’s doing. Its interpretation of the task is wrong. I am not trying to provide cleaner and more concise product information; I am trying to remove the product information altogether. Perhaps in other scenarios this type of analysis may be useful. The valuable help is below that, where it suggests a formula and shows a preview of what the transformation would look like. It’s exactly what I wanted. As a bonus, I’ve learned about a much better Excel formula for these cases of text manipulation. The question is whether I will ever need to use it!

I apply the transformation and it all works as expected.

I now move to my second requirement, which is to update the subscription dates to this year. I start by asking the following:

"Change all the dates that are in 2023 to 2024 and all the dates that are in 2024 to 2025"

I thought it would be interesting to see if it knows which column has the dates in it. This is the way you would ask another human. There is a lot of assumed knowledge expected.

Again, after around 20 seconds it came out with the results.

Now all the dates end in 2025 which is not right! I wanted the accounts that renewed in 2023 to have their renewal dates moved to the same date in 2024 and to do the same with the ones that renewed in 2024. Mmmm. What happened? I think Copilot made the changes I requested in order. It first changed all the dates that were in 2023 to 2024 and then changed all dates in 2024 to 2025 which meant that it changed them all. Very interesting! Thankfully, Copilot has the feature to undo this change. Understanding that Copilot can be very literal, I prompt it differently this time; start with the newest dates first and then the older ones:

"Change all the dates that are in 2024 to 2025 and all the dates that are in 2023 to 2024"

This does the trick:

Instead of having to think about the order in which Copilot carries out each part of the prompt maybe there is a better way to ask the question. What about this way?

"Increase the dates by 12 months"

Almost there. This time the analysis is more accurate. I want to understand when the next renewal is due. Just need to add a clarification about the formatting which came out in US format despite my Excel software being configured for UK defaults.

"increase the dates by 12 months. Format in dd/mm/yyyy"

I know there are probably quicker ways of doing this without AI. That’s probably right, but if you don’t start bringing AI to the table, how will you know what it’s good for?

Clearly a lot to be gained by following rule #1 of co-intelligence. What about the others?

2. Be the human in the loop

In this exercise, being the human in the loop was essential. Copilot got a few of my instructions wrong and by fine tuning my requests, I got closer to what I needed. What I love about this rule is that it highlights the way I ask for things, the levels of ambiguity I work with and how I need to rethink the way I express myself for this intelligence to be effective.

3. Treat artificial intelligence as a person, but tell it what kind of person it is

This didn’t quite apply to this exercise, but there are many cases where providing the AI with some background and context on its role and expertise will make all the difference. This summary table from a great paper on AI prompting is a good example.

4. Assume this is the worst AI you will ever use

This is the scary one! Copilot in Excel is on preview and it’s fair to assume once it goes into general release it will be even better. With its current level of ability, I can see myself changing my workflow and never looking for Excel formulas again in Google and just working directly with Copilot.

As good as it is to improve my workflow and be more efficient at a task I need to do quite often, perhaps the most exciting aspect of this journey is finding a use case where this technology may be useful. This is not reading about a cool application of AI and adopting it, but genuinely just bringing AI to the table and seeing what it can do. Can’t wait for the next discovery!

If you have any use cases you have found, please comment or send us an email on info@knowledgeindustries.co.uk. We are keen to collect as many of these as possible and start developing a playbook.