Transcription
Hello everyone. Today's going to be a bit different and I'm going to share with you my experience of how AI is changing my role as a financial analyst. I've had about a month to play test using some of the stronger AI LLMs out there, mainly Claude, and I just wanted to share sort of my accomplishments, my struggles of going through this process because I think it will be pretty good context for everyone else out there that's also trying to use AI in their process.
A little bit of background for myself, I am an FP&A manager and I oversee majority of the FP&A function at the company. So, I had a lot of opportunities to test using AI in different FP&A functions and [music] I was also very lucky that the company allows me to use Claude. From what my friends are saying, bigger companies are not allowing their employees to use Claude or ChatGPT because of data sensitivity issues. And so, they're just stuck using Copilot and they've been saying that it hasn't been that useful. So, I think I find myself to be very lucky that I was actually able to test [music] using Claude in a lot of my work. But, that being said, um AI has its up and downs and that's sort of what I want to share with you today.
So, after using Claude for about a month, my first impression and the impression that I still have today is that it is very very stressful. Before, when I used to work building financial models, designing workflow processes, >> [music] >> assessing metrics that stakeholders want to see, all of this sort of just came naturally. Um I didn't really actually have to think that much. It would sort of just come like that. And I was always very confident in the quality of my work. But, now that AI has come into play, I feel like that standard has sort of gone up and now if my processes aren't just a click of a button or it's not updating [music] automatically on its own, it feels like I'm not really hitting the potential of AI. And everyone's still learning AI. No one's an expert at it. So, you're constantly trying to study it as you [music] work and because no one knows the actual ceiling of AI, it feels like you're always in an uphill battle.
So, before if someone came to me and asked me to generate this report, I would have no issues generating the template, the logic, the source data, and then think about how to bring in the source data and all of that. But, now [music] after I'm done that, I also have to think about, okay, how can I streamline this process further using AI? [music] And this requires a lot of brainpower because I actually don't know the answer to that. So, I'm spending time trying to research like what is AI capable of doing? Am I using AI correctly? How can I improve this process when I don't know how to? I hope you understand what I mean.
If I had to give like an easy example, it's like math where you know how to solve a type of problem and you're constantly receiving the same type of problem and you're able to solve all of them even if they're slightly different because you know how to solve those type of problems. And then let's say that someone discovers algebra and then it's like this whole new mathematical system and now you have to start solving these problems using algebra. It's just so much more frustrating than what you were doing before because first of all, you don't know how to do algebra, so you have to study it. And then you have to start thinking about solving the exact same problem but in a different way that you're not comfortable with. So, [music] you can just imagine how much more effort that requires and as a result of that, there's just more stress.
But, I will say that sometimes it has opened up time for me because when I'm doing something that would have taken me like 3 hours, I [music] will let AI look into it first. So, I'm just going to send it a light prompt and then while it's loading, I just kind of watch it do its thing. But, I understand that it's just part of a learning process and I sort of think of this as technical debt where you're incurring more time to try to solve a problem more efficiently for the future. So, I'm spending a lot of time currently trying to learn the best processes and how to streamline stuff and hopefully in the future, some things that would have taken me 2 to 3 hours will now take me 30 minutes. [music]
But, the 1 month that I spent has not gone to waste because there are things that I have implemented AI to improve my process and I think this is the part where it really gets interesting. So, the very first thing that I did with AI currently is I have automated the process of data extraction. Once you have new information, you have to go to the source data, download it, and then migrate it into your working model. How the process looked like for me before was that um let's say I wanted to get updated financials, I would go into our financial system, generate a custom report, filter the dates, download the report, open the CSV, copy the data, and then paste it into my working model. That's a pretty manual process depending on how often you update it. But, now with AI, what I was able to do was set up a power query using an advanced editor that directly connected to the report [music] within my financial system. So, now when I'm updating it, all I have to do is just press refresh in a power query. Now, this was definitely doable before AI even came out because power query is not even an AI function. But, what AI has enabled for me to do was generate an advanced editor code that I can directly use. So, it was able to guide me to generating an API code, the report ID, and then set up all the parameters so that [music] I pull the data exactly that I need to. And when we close financials, I actually update the financial reports pretty often because there are changes that occur between days. So, this has actually saved me some time and I've also definitely incorporated this into my other processes as well such as pulling customer data, headcount information, and etc. So, it has removed a lot of my manual processes in terms of pulling data from different sources. Claude also has a function called Co-work which is supposed to be able to go into your Windows and actually move around files and make changes for you. So, there's probably something that I could do there because some of my processes require taking data from one Excel file to the other and I'm still doing that manually, but maybe Co-work can do that for you. I'm not sure, but I'll have to look into that. And this is what I mean by never-ending stress because you know that there's always something that you have not done yet when it comes to AI.
The second way that I've implemented AI is analyzing financial performance. So, right now in a standard variance analysis for most companies, >> [music] >> what you will do is pull the financial transactions from your financial system and then compare that against budget. [music] Now, depending on how you analyze it, there is a lot of data that you work with. You have to look at spend by different divisions, vendor, GL account, and sometimes even look at descriptions to see if there are any ad hoc [music] expenses or some abnormal items. And investigating all that detail takes time. So, what I did was I have set up a variance analysis template that fully formulates and links to [music] both sources of actuals and budget. I've set up helper columns to clearly identify what the AI should look at and I was able to give it pretty specific instructions on which fields that I wanted it to look at to determine the variance in the performance. And for the most part, I think it does a good job in setting up your first draft of analysis. So, it will generate all the commentaries for you depending on the fields that you've specified and then I'll go in and review the commentaries, just validate it based on the financial results, and then [music] see if there's anything additional that I want to add or remove. It's helped save a lot of time because I don't have to start anything from scratch. I just have to validate most of the analysis that AI has done for me and then try to further expand from there. >> [music] >>
And one thing that I was able to realize and learn while using AI is that you don't just want to let AI do everything from scratch. You want to set up a template that the AI should populate for you because I feel like experienced analysts have [music] a very specific way of analyzing items and sometimes AI might give you very good work, [music] but it's just not your preference. I don't know how to describe it. You have to experience it yourself, but sometimes [music] you'll see amazing work, but it just it's not what you're looking for. So, it's better that you set up a template [music] first so that AI can follow your guidance and then analyze. And that's where technical skills still come in handy so that you can make sure that AI analyzes things exactly the way you need it to rather than trying to explain every step by step. It's almost like working with a junior. Like if a new junior joined your team and you just tell it to analyze something, they're just going to give you something that is so different than what you used to do. So, you want to give the junior like your old work so that they understand how you've done things before and they could just kind of continue doing that.
The third way that I've implemented AI in my work is [music] analyzing revenue and that is completely separate from just a standard variance analysis of the company's financials because revenue is some material that you have to go into all these details and actually analyze all the drivers that are impacting your revenue. So, your data is not just coming from your financial system, it's coming from the company's data warehouse, customer information, and etc. And there's just a whole new level of data that you have to look into, >> [music] >> and what I did here was I actually didn't change my working templates too much, but I've expanded on it so that AI can clearly identify what I need to look into. One example is that when I analyze current month results for revenue, I would generally compare it to budget and see how it differs between my assumptions, but then sometimes I also want to compare it to prior year. And when I normally compare results with prior year, I would sort of just look back into my data set and see what happened last year, but in order for AI to do its job efficiently, what I did was I set up three new columns, one with current year, one with budget, one with prior, so that it knows exactly that it needs to look into those three. I'm sure with good prompting the AI could figure that out by itself as well, but then you just want to make the AI's life easier. So make everything >> [music] >> as obvious and clear as possible. Another example is if you're looking at the sale of different products, and one of the product has some issues with it and you don't want AI to analyze that, don't put it in your prompt to not analyze a specific product. You want to create an extra field that says like exclude or something, and then mark the product that you want to exclude. So that when AI goes through your data, it can reference a specific field and know that if it says exclude, they don't want to include it in the analysis. It just makes [music] it much more obvious for them and it makes the results also more accurate. So I've learned to sort of steer my working files so that it's more friendly for AI to use by making things very obvious and [music] clear. And in addition to this, I've also began uploading a lot of qualitative data to my worksheets so that AI can directly reference those as well. For example, we also analyze all the customers that have churned or left the company, and we have another team that documents their exact reason [music] for leaving. So for my management commentary, I would look at all the customers that left and the reason why they left to sort of put together a story of why [music] these customers are leaving, but now for AI to properly analyze this for me, I would clearly indicate which customers have left and also tell AI to reference a specific tab which has qualitative information so that it's able to tie all of these things together. Before AI, there was absolutely no reason for me to do that because it never actually impacted how the quantitative numbers came out, but then if I wanted to get value out of AI to analyze it more efficiently for me, I realized that having everything that I look at in one file just made it so much easier.
And the last way that I have implemented AI in my process is [music] to analyze large data sets for ad hoc work. So I generally receive a lot of requests from different divisions and [music] there's always some new data set that I have to work with. And before without AI, I would look at all the different fields and then try to make some judgments based on my understanding of the operations and then conduct some preliminary analysis to see if there's any correlations, relationships, and etc. between these different fields. But now I can just tell AI to look into these fields, I provide it operational context, and then look into for me the core fields that I should be referencing and if there is any relationships that I need to be wary of. And I would say that it gets you about 80% of the way there, and you will notice that almost everything that I do, AI gets you 80% of the way there. You always have to review and validate your work, but it still saves you a lot of time.
I've also seen some people use AI to come up with a first draft of a template for their analysis, >> [music] >> and I personally don't like that. As I mentioned before, I have a very specific way of analyzing things and then I realized that the templates that the AI generates just don't match my style. Maybe I have to prompt it better, I'm not sure, but the AI that's currently built in Excel right now don't have memory capacity, meaning that if you tell something it today, it's not going to remember that you told it tomorrow. So I found it much easier to come up with a template myself first and then guide AI to populate [music] that. I've also seen some people use AI to generate formulas for them and error check formulas. I think it's pretty useful, but I've never had too many issues with formula writing, so I don't really use it that much, but I think it does help if you need it to do something pretty complex. And I've also tried asking AI to like review all the formulas in my worksheet to see if there's any errors and then it will always tell me that everything's good and then I'll always find a mistake later. And I think I'm partially at fault because all I tell it to do is review my worksheet and let me know if there's any inconsistency or any errors, and it tells me no probably because it doesn't find any [music] error in formulas, but then there's like some calculations that are just completely off. So I would probably take it with a grain of salt if you're using AI to error check your entire file.
Overall, um that is how I am using AI in my work. I think it's pretty useful and I still don't think I've even tapped into 20% of what AI is capable of doing yet, so I'm going to continue experimenting and if I find anything else that is groundbreaking or very useful in my work, I'll try to share it with everyone here. I'm also very interested to hear how you are using AI and see if we can all share tips to improve all our processes using AI together. If you have any questions, feel free to reach out always and I'll see you next time. Bye.