In this special episode (part of Datarails’ FP&A Con 2024) Glenn hosts an all-star panel to debate AI and FP&A: Hype vs reality.
Join Nicolas Boucher, Christian Martinez, and Gabriela Gutierrez, alongside CFO Glenn Hopper, host of FP&A Today, and an AI thought leader and author. Delve into real-world applications, separating fact from fiction, and gain valuable practical insights into the present and future of AI in financial planning and analysis.
The Guests:
- Nicolas Boucher, Keynote speaker on AI for Finance & FP&A
- Christian Martinez, Finance Analytics Manager, Kraft Heinz
- Gabriela Gutierrez, Financial Planning & Analysis Specialist
In this episode:
- How many FP&A pros are using AI in their work – how do you compare vs the results of our poll
- The main misconceptions + and what not to do in the finance and AI era
- The significance of ChatGPT 4o
- The biggest barriers to AI adoptions
- Practical uses of AI to use today , including forecasting and practical ways to save time
- FP&A AI use cases that have shocked the panel
- Which FP&A processes could most use AI help?
- How to get started practically
- AI capabilities in existing tools such as PowerBI
- Why 2024 is about Python as the tool to learn to break the barrier of AI
- Automation in finance’s relationship with other departments
- Creating your own chatbot services
- The importance of data maturity (before AI maturity)
- Download financial data (balance sheets, savings) from a company on Yahoo Finance and plug into Chat GPT 4o (Prompt: Act as a management consultant and analyze this data with an FP&A lens”) to get volume analysis, volatility analysis, moving averages.)
Full transcript
Glenn Hopper:
So, um, welcome to FP&A Con 2024. This is our first session on AI and FP&A hype versus reality. Today we’re gonna dive into real world applications of artificial intelligence and try to separate fact from fiction, um, fact, not all fact from fiction, just when it comes to the state of AI today. Uh, we’ll look at the past, uh, not the past, the present and the future. Monday morning. This is a tough start here. Um, we’ll look at the present and future of AI and financial planning and analysis. Um, and we are gonna record today’s session. So we’ll send a link to the recording after the event. And notice, like I said before, that we do have chat and Q&A enabled here. So please be sure to submit your questions and, uh, we’ll try to answer ’em as we go.
And at the end of the presentation with whatever time we have left, uh, this session will also appear as a special episode of FP&A today, the number one podcast for FP&A Pros. Um, I’m Glen Hopper, the host of FP&A today, and the head of AI research and Development at Eventus Advisory Group. Uh, I was a startup CFO for 15 years and came up through the FP&A ranks, so I genuinely love all this stuff. Um, I’m also an author and lecturer on AI and its applications and finance and accounting, and I’m proud today to host my esteemed panel of finance and AI experts. We’ve got Nicholas Boucher, Christian Martinez, and Gabriela Gutierrez. Gabriela is a finance leader, a keynote speaker, and an advocate for financial innovation with 10 years of industry experience. She focuses on strategic financial management, driving growth, and optimizing business performance.
She is currently transitioning from local CFO at Teads to founding her own startup called Tabs, which aims to solve some of the biggest problems in finance today using AI. Christian is the global finance analytic, global finance analytics manager for Kraft Heinz, not the CEO, despite our best efforts. Um, and he’s also an AI for finance instructor. He has over eight years of experience in multinational companies across international markets, including Australia, Mexico, and the Netherlands. He teaches Microsoft copilot Python and advanced chat GPT for finance. Christian was named among the 30 under 30 in accounting and finance industry in Australia in 2021. He also secured the International Rising Star in Finance Award in 2023 and the EMEA data Democratizer award in 2022. Outside of finance and ai, he’s a marathon finisher and a world traveler having explored over 70 countries. Nicholas is a recognized thought leader on finance topics like AI for finance, FP&A and controlling.
Nicholas is also the founder of the AI Finance Club and international companies, including Mercedes-Benz, Chanel, Hugo Boss, and KPMG Trust Nicholas to train their teams. He has more than 15 years of experience working in senior finance roles. He speaks three languages and has lived in five different countries. Um, okay, before we get started, um, and I’ll let you guys all <laugh>, I I think we’re gonna jump into the poll and while we’re, um, an answering the, the poll, um, the, the first poll question is, have you used AI in your work previously? And this question we added at the last minute, but we thought it was a good idea to get kind of a level set of where we are. So we’ll give everybody, um, a few, few minutes to, um, answer the poll questions here and then, and then we’ll dive in. Christian, you wanna take a guess on what, uh, what our percentage of, uh, people who’ve used AI and their work previously is gonna be?
Christian Martinez:
Yeah, definitely. I’ll say 30%, um, have used, that’s what I’ve been seeing in the, in the past.
Glenn Hopper:
Yeah. Have you seen it? I’ve, uh, and, and this is, uh, Nicholas and Gabriela for you too. I, I have seen this number come up as time goes on every month. It seems like there’s a slightly higher number. Um, Nicholas and Gabriela, you wanna take a guess where we’re gonna be on who’s used it?
Gabriela Gutierrez:
Maybe 40%. Let’s say 40 50. If people will say 50, I would be very surprised. But they, uh, but they, we use also AI for many different use cases that we don’t even notice. So yeah, if we would include them then 50.
Glenn Hopper:
Yeah,
Nicolas Boucher:
I, I think in, in my experience, I do a little game when I train, uh, people on ai, um, especially in person, and I put people on a wall. So like, imagine a line of 20 people and ask them to step, uh, one foot ahead each time. They will say yes. And the first question I ask is, did you ever use the GPT or copilot? And then you have already only 89, 80% of people who step, so you have already 20, 30% who didn’t move one step. So they never use it. And that’s interesting after to ask why. Then if you ask when, uh, you will use it for work, yes or no, then you will have like, half of the people will step. But then after, when you ask, are you using it every day or every week, then you’ll have 10 and 20% only of people in front. And that’s interesting to see after like physically the, the map of the people and to ask them, uh, why they are using it a lot or why they’re not using it.
Glenn Hopper:
Alright, let’s see, let’s see the results here, where we came out. Wow, 47%. Yes. So this, this number is coming up every month. It’s, uh, and also maybe people are self-selecting by joining this webinar too. If you’re joining the AI and fp and a, you might be, you know, kind of self-identifying they’re out of the gates, so, okay, well I’m only two minutes behind in the first seven minutes, so I guess we <laugh> better dive in quick. I think, you know, so we had the poll to see where, where everybody is, how many people have used it and are, are using it in their daily workflows. But because you guys are out there at the forefront, you know, working with teams, um, how are you guys seeing finance teams using AI and finance today? And then I guess as a follow up to that, how do you see ’em using it in the future? And I guess Nicholas, let’s, uh, let’s start with you on this one.
Nicolas Boucher:
<laugh>, sorry. So the thing is, you have a, like Gabriela says, a lot of people were using AI before the AI was hype. So it was like banking, it was like, um, big companies with a lot of fraud detection. Uh, but now that we all have since one and a half year, ChatGPT and co-pilot and Gemini, actually, it’s open to everybody. And the first thing what people do is they write emails and uh, but after they have like played around and wrote emails, they get frustrated because they don’t know how to use it and they use it like Google, like that’s what I call like the Google method. And if you use it like Google, then you will only get something super generic and you need to really change your, I will say your framework because you have not Google in front of you, but you have like a mini assistant or intelligent, uh, trainee, which has read all of the Wikipedia articles and, uh, which can process much more data than what you can think.
And you need to ask concrete questions, you need to get to be specific. And uh, an example I always do is when you want to write, uh, dunning letters, then if you just say Dunning letters, you get, oh, here are the 10 ways to recover your money. But if you ask, can you draft for me a donator? You get already a good draft. But then after, if you explain, my client never paid, so how can I really have something to move the needle? Then it will write really like a lawyer with exactly the legal actions. And like in really one minute, if you tweak your prompt, you get from having something super generic, and that’s where 90% of the people stop to having something super concrete, that even A CFO will have a hard time to write by himself or by themself.
Glenn Hopper:
Great. Um, how about you Gabriela?
Gabriela Gutierrez:
Um, I would totally agree with Nicola and, uh, for that I have seen most of the use cases of people using, uh, any language model to send emails, to reply to things. Uh, but I have also seen for some companies that they’re using language models to write policies. So, and even policies about using ai. So it’s a bit funny, like, uh, you write, you ask the language model, please write me a policy of for employees of how to use ai. But, uh, for them it was a really nice use case because they even, or the model came up with more edge cases that they wouldn’t have thought before, and then they were able to implement it. I think also what is important is to always have the human in the loop. So someone like reading about policy and making sure that it fits to the company and, and as well, like the feasibility to really prompt again and to ask for changes. It’s, uh, quite amazing. I mean, also another use case I would say like predictive analytics and so on, that you could just put your data and ask for questions. Obviously that comes, uh, a bit into the privacy topic, but I think we can discuss that in later details.
Glenn Hopper:
Yeah, yeah. And actually, Christian, this, uh, this moves over to you. Uh, let’s, uh, I want to introduce something that maybe a lot of people haven’t seen, and you might be a good one to introduce this. So thinking about the ways that a lot of departments, you know, marketing copy, marketing copy, marketing copy <laugh> rather is, uh, um, one thing, but when you’re actually, like Gabriela referenced using, using it for data analysis, that’s something completely different. So, um, I, I know you’ve done a lot including, you know, writing Python and using the data analysis tool. Um, can you talk specifically to use of of that tool for, uh, for fp and a?
Christian Martinez:
Yeah, definitely, because I, I would say that’s one of the more exciting, uh, use cases of artificial intelligence for me in FPNA, it’s actually these like machine learning predictive analytics models. So I think like 10 years ago when I started studying like a, a master’s in data science from finance, that’s where I wanted use to learn like, okay, how to use program languages like Python to start doing forecasting, let’s say, of, of sales. And then you need to really study a lot of time in order to make your first, um, forecast using machine learning. But nowadays, if, um, a team is leveraging in a correct way, like these, uh, lms, they can basically generate Python code and then implement these machine learning models, uh, in order to do forecasting of like sales revenue or even identify financial performance, uh, drivers. So I think that’s one of the barriers that some FP&A teams are breaking right now. And for me, yeah, it’s very, very exciting.
Glenn Hopper:
Yeah. Yeah. So, um, and I, I think also, you know, thinking about the data analysis tool and then just to sort of inherent abilities of an LLM, um, Gabriela tell me, so a lot of people have, I mean, there’s current perceptions and then some misconceptions on, um, AI and FP&A and one that I always lead with is LLMs aren’t inherently good at math. That’s not what they’re designed for to do math. Now obviously tools like data analysis tool, it’s writing Python under the hood and it, it is doing math that way. But, so to me, the biggest issue I have with people, uh, or that, that I want people to understand when I’m first introducing that, hey, you can use these LLMs in finance and accounting is, but don’t get the straight large language model to build you an amortization table because it’s not gonna actually do the math unless it’s actually writing the, you know, even, you know, simple things like, uh, calculating depreciation. If you’re just asking the language, large language model for results. I, I, that’s not what I’m trusting. So that’s the biggest thing I have is there are ways to do this, like Christian said, you know, writing code or, or using the data analysis tool, but the LLM itself, I’m not gonna trust when it’s just in line giving me a textual answer. But Gabriela, what do you think are, um, some perceptions and misconception misconceptions and things that people need to understand right now with, uh, using AI in finance?
Gabriela Gutierrez:
Um, I think like overall, that, uh, then, then the thinking that you could just put all your data and it will give you all the answers would be kind of a misconception, uh, that you could, and also if the first answer is not the correct one, um, because, uh, the way how language models work, it us, usually the number of tokens that you would have a token would be, it could be a word or it could be a mix of a couple of words. Um, they, they could be misinterpreted. So the straightforward, like you are able to request the prompt, the better would be the result. But also if you include a very large prompt, uh, it would be a bit more difficult for the model to even understand what is the task to do. So also, I, I wouldn’t, um, or I would encourage people as well to try more prompting and, and I think like, uh, like people right now, they are even specializing on prompting to really get like better results and to get the model, uh, to do even more kind of more customized or more precise, um, answers.
Um, I will say that would be one of the, uh, the misconceptions, like really giving all the data and having the best answer on the first try. So that’s not happening. Um, another misconception is that, um, that for example, on the data privacy part or like they could train to with my data and so on. So there are many different like companies, so with many different terms. So I would also encourage people to kind of, uh, like you have the free version and as well you have the paid version. So within the year, different terms, uh, so on, it could, you can have like also kind of your own cluster your own data. So like they wouldn’t share it unless, um, it is obviously written in their term, uh, um, conditions. Sorry. Um, um, and again, like on, on the third use case as well, it’s for example that on the predictive analytics or on more machine learning is that you could just pull all the data again and it could give you the perfect forecast. I think like it needs to be a bit of a trial and error that you play with many different models and then you choose what are like the best, uh, that they fit, uh, overall for, for the use cases because, uh, it is not a one model fit all, but rather it’s like going into various specifics and really trying and, and playing that. I I always call it trial and error until you find the one.
Glenn Hopper:
Yep. Yeah. Christian, what do you think?
Christian Martinez:
Yeah, I would say, um, two things. So like, one of the main misconceptions that I have been seeing people having is that LLMs, it’s just the only artificial intelligence, uh, application that they can use in A PNA. Um, so there are like so many more applications of, of ai, like, as Gabriel was mentioned, like machine learning models, uh, and techniques, for example, that are so much more powerful for finance and fp NA. So it’s not just like you will go to a charge G PT and then you ask questions about your financials and so on. Like, it’s more, again, like you can use charge GPT to generate Python code that create some, uh, financials, either visualization forecasts, uh, and so on in another environment. So it can be like Google call up, it can be Jupyter Notebooks, um, it can be something else, Azure from Microsoft. And then I think that’s the more powerful use case of let’s say like AI in, in FP&A.
Glenn Hopper:
Yeah, it’s funny ’cause you know, for years dorks like us have been <laugh>, uh, you know, using, uh, machine learning to, for classification and regression and fraud detection, you know, all the uses that, that there have been to it. But the, the barrier to entry on all that is, well, you have to know Python or you have to, you know, you have to be able to code and you that you have to have sort of the data science approach. So now with generative ai, this is a lot of people’s first experience with ai. So it’s, it, there’s an, a temptation to put all AI in the box of just what these generative models are. So, I mean, the good news is it’s, um, actually, you know, it’s kind of democratizing ai, it’s making it accessible to more people. But the danger in that is, um, that if you don’t understand how the model works or what to expect out of it, that you know, you’re a <laugh>. If, if you don’t know how you’re using it, you’ve got a very powerful tool that could, could give some, some bad results. Um, Nicholas, what are you, what are you saying? What do you, what do you think?
Nicolas Boucher:
So what’s interesting, because a lot of people say AI, and it was like, I think 20, 30 years ago when computer arrived and people are saying, oh, computer are going to do our work. And when you think about that is not the computer who is going to do the work, it’s maybe a special program is going to do a special task, uh, that before was doing by the human. And AI is the same. You cannot say AI is going replace, uh, US AI is going to take all over, uh, our job. It’s more like, okay, if, uh, a tool using some AI techniques, so OCR, NLP, machine learning has access to a part of your data and can on top, right, so create, then it can do a part of the work. And if it doesn’t have access to all of this, then you and AI is going to do a part of your work faster and better.
And so learning about what means AI, what, uh, Christian explained, and Gabriela and you as well as well what you explained. So learning what it can do, and then learning what you have, what is your current setup, and see how you can connect, uh, the more you’ll have, uh, connection. Like today, uh,Open AI just, uh, or last week they opened the connection with Microsoft and, uh, Google Drive, but is still, you have to drop manually and to decide yourself. But companies right now start to have integration within their own company where humans don’t have to connect things. They just say, I want to see, uh, the sales of last months, and then they get a report in front of them. But you first need that, somebody connect that, and people need to figure that out. That is not AI. Who is going to do all of this is a set of AI and a specific program using AI. And that is like a big misconception that you need to train and educate people. And once they understand that, then they know what is possible today. And they also have kind of an idea where they need to go, and also when they’re going to select tools, what are the criteria to select, uh, those tools.
Glenn Hopper:
Okay, Nicholas, you set me up for the perfect transition into our next poll question. And, uh, and what, so I’m gonna throw up the poll right now. And, uh, for those who are listening only, the poll question is, what is the biggest barrier to AI adoption in your organization’s fp and a processes? And our choices here are lack of understanding or awareness, data privacy and security concerns, inaccurate data, budget constraints, lack of skilled personnel or other.
And you know, while, while we’re waiting on responses to this poll, um, I think that Google’s latest gaffe from last week or the week before where, uh, you know, it was telling the, their, their AI search in enhanced search, whatever they’re calling it now, was telling people to eat one small rock a day to put, um, cheese on, or to put glue on pizza, to keep <laugh> cheese from sliding off and whatever the absurd things that the AI was kicking out.
And Nicholas, to your point, that that human in the loop is so important right now, and I understand there’s sort of this scientific idea or this science fiction rather idea of AI that’s gonna replace us, but as we’re seeing, um, these, these kinds of issues that human in the loop stays so important. But if you are, you know, if you type the wrong numbers into a calculator, it’s not gonna give you the right result. Or if you do, if you do your order of operations wrong, you know, so this is a tool that we have to know how to use. We as users have to understand how to use it and when to trust it and when not to. So I think, you know, a lot of this lack of trust right now is, well, we’re not ready to hand over the reigns to AI to run our fp NA department, but there, if you’re using FP&A or if you’re using AI and fp NA, it is giving you a leg up on, on people who aren’t.
So it’s, you know, it’s not AI versus people, it’s people using AI versus people who don’t. So, um, okay. I’ve rambled on long enough. Let’s see the, uh, the poll response here. So biggest one, yep. To, to the point there, lack of understanding or awareness. So people joining stuff like this, that’s step one, just to get kind of the broad overview. But I think what I’d like to do, um, before the show’s over is talk about, uh, you know, where some people can learn, uh, more about AI and, and how to use it. And I know, uh, everyone on this panel, uh, has a, a lot of great resources for that. Um, data privacy and security, very big. And, and Gabriela hit on this a little bit with, I always tell people don’t just don’t go to the publicly available, uh, interface, you know, chat GPT-4 oh free for everyone.
Now if it’s, if they’ve rolled it all the way out. But when it’s free for everyone, it’s kind of like Google, uh, you’re <laugh>, you know, when you’re not paying for something, we’re getting something from you. And that’s something from you is the data. So all of your data is being used to train the model. So there are, there’s chat PT for enterprise, there’s closed, uh, environments for all these. And that’s, uh, if anyone is gonna use proprietary information, that’s where to go with it. So lack of understanding data privacy and security inaccurate data, that’s, you know, that’s that trust era. And we can, uh, talk in a little bit about how we can trust more budget constraints. I get, um, a little bit, a lot of these now the pricing is, is, um, is coming down. Uh, another big one is lack of skilled personnel.
And that is, you know, we are people who are trying to implement AI today. We’re all kind of at that bleeding edge where we’re out over our skis a little bit with this. Um, so it’s, so that it’s up to the rest of us to, to kind of reskill and upskill to understand how to use these tools. So very interesting poll results there. So, um, let’s talk about though, for people who are, um, successfully using AI and FP&AToday, because I think everybody wants to know. Yeah, it sounds great. I get it. I, I got it to write a limerick for me. It was great at that, it was funny. Um, but how can I actually turn this into my professional workflow? So I do wanna talk a little bit about, um, what you’re, you guys experience are of successfully implementing AI and FP&A and I I know Christian, I think you’ve got a, a good, a good story on this. I’d love for you to, uh, lead us off if you can.
Christian Martinez:
Yeah, definitely. So what I have been seeing on, um, let’s say like teams that successfully implement AI is, as someone was commenting on the chat, that, um, training in a incorporate environment. So they really need to be trained on how to use AI properly. And then there are like so many use cases, for example, the very, I guess like, um, simple one, but powerful. It’s just for forecasting. So a lot of us in FP&A, we need to do forecast every month, sometimes every, every like three months as well, every year and so on. You can improve your forecast so much more if you use artificial intelligence either to do like a, let’s say, clustering method that is a machine learning technique in order to segment, let’s say for your different like business units, your customers and so on. And then implement other algorithm in order to actually predict the sales that we’re going to have, the revenue, identifying the top financial drivers and so on. Those are, I think, the main, um, use cases that people can have, let’s say in forecasting. Then the second one is just around saving time. So a lot of us, especially also in, in FP&A, we need to either to consolidate files, to merge files, some mappings, like we need to be like, uh, done manually and so on. You can use as well, um, AI tools in order to do that, uh, for you. So that part of like saving time is the, the second one.
Glenn Hopper:
Um, I, so I’ve been try, you know, like all you guys, I’ve been trying to push kind of the limits of what generative AI can do. And, uh, when 4.0 came out, I was pretty blown away. Uh, speaking of forecasting, I was doing scenario analysis just straight in 4.0. So I had my, you know, three years of quarterly financial information and, uh, in the past I’ve gotten, um, uh, the data analyst tool to, you know, do a simple regression or whatever to just carry the forecast out. Um, but with 4.0, I was actually able to get it to do some more complex forecasting it in one shot, did a rema forecast it de seasonalized det trended, did that full seasonal auto regressive, integrated moving average. I think I got that right. I, uh, Sima forecast, which was much more accurate than regression. And then I got it to do, you know, three different scenarios.
And then I got it to run Monte Carlo simulations all direct in line. So it wasn’t like I was getting it to write the Python, I was going in another application, it was doing this right there. So that was the 4.0 I think kind of quietly went under the radar for me. But when I saw it doing that all in one shot accurately, um, I I was like, okay, now we’re starting to see some more practical use cases for this. Uh, Nicholas Gabriela, do you, have you seen some, uh, uses that you’ve kind of been blown away with recently?
Gabriela Gutierrez:
Should I go first?
Glenn Hopper:
Yeah, go ahead. Go ahead. Yeah.
Gabriela Gutierrez:
Okay. Um, I, I think it’s very similar to what Christian mentioned and, um, three years ago actually, or maybe four, I was working for, um, for a different employer. And, uh, there we had tons amount of data, like, uh, our Excel sheets and spreadsheets will collapse because of the amount of data that we had. And obviously running that forecast or even creating a budget for four different products, it was quite, um, a challenge. Uh, and therefore, like we even tried to look for solutions in the market, but really we couldn’t find the one that it would fit for all of us. And that’s why when we change completely to running like our forecast into like using a call up, so it’s like a Jupyter Notebook, but on Google Drive. And then we, we were able to automate all of our kind of data ingestion into directly to our call up notebook run there, our forecast, and then we just did a small integration of the consolidated version one spreadsheet, so we would be able to see it okay, month by month. We would do then like, um, variance analysis and so on. So I would say data has been one of the most successful use cases as we, we were able to achieve like an accuracy of 99%, so for a 12 month period. So that was one of the biggest use cases so far.
Glenn Hopper:
How about you Nicholas?
Nicolas Boucher:
Uh, I had the chance actually to talk with the OpenAI team and, uh, their finance team. And what was interesting is, uh, when I talked with them, they explained me that they had to go from a business doing I think 20 to, I think 20 or 200 million, uh, annual revenue in 2022 to $2 billion in 2023. So if you see the graph is like, and uh, their team, they could not keep, uh, up with the, the pace of the growth. So they had to grow with the members, but there was too much work. What they did is they, uh, hired robots <laugh>, and they didn’t hire them. They actually created those robots. So two, two type of business cases for these robots. One is a bit more what Christian and, uh, Gabriela explained. So creating Python script that will do the work that normally humans do.
And even though they were, they are in open AI and they have access to the LLM, they will not create that inside the model because for three reasons. The first one is not scalable. ’cause you cannot, like, you have token limits and you cannot run a discussion or a chat for each of the operations. Second one, you cannot actually, uh, audit it. So, uh, if somebody comes after you and then, uh, look at the work, then it’s not possible to, to see what happened is the black box, even if you have the code written in the discussion. And the third one is still, uh, they didn’t want all of the data in the model and to be trained because they want, didn’t want to put it all of the models or like I would say, another company, even though it’s, uh, open ai, you don’t want maybe all of your data there.
And so for these three reasons, that’s why it’s not, uh, good to do all of your competitions there. And the second one is actually a chatbot for all of the questions they receive. So for all of the questions they get on, uh, so it’s more the accounting department, but on travel policies, uh, reimbursements, all of the thing or how to categorize expenses, they created a chatbot to transfer all of the questions. And like this, they could filter down to only a few questions per day, rather than a hundred of questions from all of the new, uh, a open AI employees.
Glenn Hopper:
Yeah, and I, creating these chat bots, I mean, they’re obviously you can customize and use the assistance on the backend, but I think, uh, and I, I don’t wanna be a homer for one, uh, one LLM over another, but the GPTs from OpenAI that you can build yourself, not GPT as in generative, uh, the gp, I don’t, they’re terrible at naming things at open ai, so they call ’em GPTs, but these are chatbots that you can build with no coding. And these are, uh, to your point there, these are great tools. If you have, you know, internal documentation, assuming you’re on the enterprise or team account where you know, you know that your DA documentation is secure, um, where you can build these GPTs with no coding, so you could make your customer service chatbot your FP and a q and a chat, bott or whatever. There’s, I mean, and these are the democratization of these tools, is that you don’t have to be able to write Python to do it. These are very automated, easy, uh, easy, uh, tools to make. So, um, definitely something worth. And if you’re,
Nicolas Boucher:
Yeah, if you’re using Google, if you have a contract with Google or Amazon or, uh, Microsoft, they already have that as well. So GPT is the famous one because we are so used to it, but I think copilot now, since last week, they have the same, uh, user interface where you just chat to create your, uh, your own mini assistant and all of the others, uh, big ones have that. So I, I will also advise for people who are saying, oh, but uh, my company doesn’t have a contract with, uh, OpenAI and until they have one, and maybe they will never have one. Where do I start first? Look actually, what is your company using? If you have a contract with Microsoft Azure, and I’m not, uh, uh, affiliated with them, but you can do so much more already just with that, or same with Amazon, uh, web services, same with Google. So just look at that first, and if you don’t have the money for it, there is a lot of, uh, small, uh, small alternatives that don’t, don’t cost much.
Glenn Hopper:
Yeah, and I think, I mean, that’s, if, if people, so the 50% roughly who have not used, uh, GPT or, or not using it in their work, I always think, don’t, you know, don’t start just uploading your company’s financial data straight into, into one of these LLMs. Start with just interact with it, ask it, you know, super benign questions, just understand sort of the capabilities, or I always say take some public company financial data that’s already out there and put that in and play around with, I think people, you know, that sort of remember the other poll where it was, you know, lack of understanding of what’s out there. I think the first thing people can do is just start interacting with it. Don’t worry about solving your work problems right now, just understand what’s working. And I think another part of that is you have to also kind of learn a little bit.
I’m not saying you’ve gotta become a, a, a developer or a data scientist, but you do need to, if you’re gonna be trusting this tool or interacting with it, you do need, do need to understand that it’s different than asking a magic eight ball a question and shaking it up and seeing what your answer is. You need to understand what is, you know, roughly how is the LLM generating this, this content? And I think it goes back to that what I was saying about the problems with Google and their, and their web search. You know, we forget because it’s, it’s easy to anthropomorphize these models, um, but it’s easy to forget. They don’t actually have any knowledge of their own. They’ve read, you know, the entire internet, but they’re spitting stuff out sort of in a statistical probability way. So that’s why a human in the loop, if you know how to prompt it and get the right information and know how to sift through and know, I don’t need to be eating rocks, or I don’t trust this amortization table, or whatever it is, it still can be a, a powerful, um, tool. And I think, you know, maybe that’s a good point to transition into. Let’s go to our, our next poll question. Um, and this is maybe where we’ll drill into some more specifics after this, but the next poll question is, which FP&A processes do you think you could use, um, AI’s help with? So we’ve got budgeting, reporting, and analysis are the three categories here, obviously a million more, but we’re gonna try to keep it into a, into a range here with what we as FP&A professionals do.
And really, uh, is there a select all option? Because <laugh>, why, how do you pick, uh, <laugh>, how do you pick just one of these? So, um, okay, let’s see. Let’s give everybody about 30 seconds more. Gabriela, what do you think, uh, what’s the, what’s the low hanging fruit out of these three, do you think?
Gabriela Gutierrez:
Reporting? Yeah, I would say on the data cleaning side on consolidation, there is a huge use case. Yeah.
Glenn Hopper:
Christian, what do you think?
Christian Martinez:
Like my favorite one to be part is the, the analysis part. So because it’s not just, let’s say describing what the data is having, like when you really report and so on, but AI really can help to improve the financial performance of your, your company, uh, via these more like prescriptive analytics, let’s say.
Glenn Hopper:
Yeah, yeah. Uh, Nicholas,
Nicolas Boucher:
Yeah, I will go with analysis because if you have the possibility to give data that you don’t have time or don’t have the volume or capacity to analyze, you can get, uh, the help of ai. But what Gabriela says, actually what people want is first reporting because they spend so much time on reporting that all of the questions we get is not, oh, how can I analyze better my da my data? The first thing, how can I just make a reporting and then like, I can work analyzing. And the thing is, because what I said at the beginning is your AI tools that you are using, so if it’s a LLM or if you’re in copilot for, for Excel, the thing is it’s not connected to anything. So reporting is getting your data or actually getting a lot of different data sources to go inside one framework, either your management reporting or your, um, uh, legal reporting. And this has a model that, and the data needs to feed inside it, and AI cannot do the integration for you because it doesn’t have access. Maybe like in the future it will be able to write the code to do all of the APIs integrations to put it inside, but it needs that somebody let that happen. And, uh, you need to have like a, a Python, uh, I would say like, orchestra behind, I will do all of these APIs until you have the final output inside, um, Excel or PDF file.
Glenn Hopper:
Yeah. Alright, well, let’s, uh, let’s take a look at our poll results here and see what everybody said. So analysis 77%, reporting 57, budgeting 53. And it’s, you know, I, I, I think that really, I would want to just check all three, but, um, you know, in selecting one, Nicholas, to your point on reporting, I think about all the ex, you know, Excel models over the years that I’ve had, where you’ve got, you know, you get your GL data at the end of the close cycle and you’ve got, you’ve got your index match or your v lookups or whatever you’re doing <laugh> to like pull it back in. And like, just the, the amount of time that just goes in, getting the data from raw data into a management report, into your legal reporting, whatever it is, like, you’re not adding value there. You’re just, you’re formatting and it’s, uh, it’s frustrating.
But if you, with ai, that sort of, the promise of it is if you can automate that task, um, and then you spend more time in the actual value add of providing some strategic insight to it, that’s where FP&A a can really, um, start to add add value. And I think that this probably is a great transition into the next section that we wanted to talk about, which is practical advice and best practices. And Nicholas, if you could, uh, maybe get us started with, um, for our listeners who are, you know, I think where most people are right now, yeah, this all sounds cool. I don’t know where to get started. So if, if I’m listening to this right now and I’m thinking, okay, yes, I wanna start bringing AI into my shop, what’s, what’s the first thing I do? What are the steps I follow to start being able to use AI in my, in my processes?
Nicolas Boucher:
Yeah. So I actually divide that based on what is the current state of your company because, uh, a lot of things that we can do with a small company you cannot do at a big company and vice versa. Um, so let’s start first, can you see my screen all of you? Yep.
So I’ll say that, um, if you are a small company, meaning zero to 1000, even one ton starts to be big, but, uh, i, I had to rank and to choose. So, um, so everything that you can do for a small one, so I will repeat myself for a small one. Normally you can do that for medium and large company as well, but if you’re a large company, you have much more means than small ones. So that’s why I segment it this way. So for a small one, I will start by making everybody educated on what is possible to do with the generic tools, because there is no need to go and buy something. There is no need to bring a consultant, you don’t have money, you want to have a result first. So you need to educate your own people on use, uh, using generative AI tools like ChatGPT, copilot, uh, Gemini.
Then you need to look at, for example, and that’s just an example on Power bi. They are already AI functions inside the tools you are using today. And in Power bi you can ask questions and get, uh, a report for you if you ask the questions the right way. Same in copilot for Excel, same in, uh, PowerPoint. So, uh, that’s an example. But this is the way to use AI in FP&A. Then the third way, we talked a lot today about Python, but for me it was really 2023 was, uh, ChatGPT, the tool to Learn and Gemini and the other ones. But LLMs. 2024 is Python because this is how you are going to break the barrier of leveraging AI for your business. ’cause you cannot count on only external tool. And Python is something you can use on your own environment with your own data and that you can do everything with it because you can automate, for example, you can consolidate 100 files in one file in just one line of code.
You can create a heat map analysis for cohort, which is really hard to do in Excel. In Python is also also just one line of code. You can do forecasting with machine. Really simple, uh, as well, like you explained, uh, also Glenn and those, all of these in your own environment without any problem of confidentiality of data and, um, also scalability problems. Then if you are a medium company, so you start to have enough volume, you start to have also, uh, money. So you can start investing in native tools, for example, in your accounting or even in FP&A to, uh, have those tools that already work with AI algorithm. Uh, so either OCR to digitalize your processes, either NLP, to ask questions on your data so you can invest on those tools. And on my LinkedIn profile, I made a, a map of the 100, uh, AI finance tools.
So you, you can have a look, um, this one, uh, I wanted to change that, but uh, still the same. But the third one on automation, you can bring automation where, for example, when you need to collaborate with a lot of departments, well, instead of you being the person who goes to all of the people analyzing the emails, combining all of the output, you can automate that with, uh, automation, automation tools. Same if you need to send, uh, reports to other people. Either you create, um, I’ll say dashboard that you can have, uh, on demand or you automate this. And then the third part is where for large companies you can invest money, you can bring external teams and build your own stuff. So you can start to have your own AI models, which are basically 99% GPTs plus your data and your fine tuning.
And that’s where, uh, I’ve seen companies, uh, starting to do that. Like, Hugo Boss and other companies, they already have their own ChatGPT, they didn’t start in finance yet, but they start to feed up, for example, with the, um, onboarding for HR or for, uh, it, uh, IT problems. You can ask your questions and you have all of the IT help guide inside your charge, GPT. Then the second one, you can create your own forecasting models. Because if you are that big, then you have a lot of historical data. You have also probably a lot of macro economic, uh, dependency because if you’re a big company, you are probably, uh, selling to a lot of the world. And the third one, like I explained with OpenAI, you can create your own chatbots. So those are segmenting like this, uh, it’s the way I will say to start small, even if you are, like, if you don’t have, even if you’re in a big company, but you want to start, like start like a small company and then go up the, the ladder. And for everybody, just quickly, if you don’t know where to start, we deal with Christian like, uh, AI maturity assessment. So I’ll put the link also inside the, the chat, but you can here where you are inside your, uh, AI readiness and, uh, use this questionnaire to, to do that.
Glenn Hopper:
Right. And I think, uh, a big part of AI maturity, it, you can’t even really talk AI maturity until you talk data maturity. And I’m sure you can suss that out in your, in, uh, your assessment as well. But it’s, you know, we’ve been talking about digital transformation for 30 years now, and if <laugh> and there are still companies who are kind of lagging on that, but it, you know, if there ever has been a time where it’s, you know, you’re banging on the drum of get your data house in order, it’s now because you’re not gonna be able to use any of this. If you’ve got garbage data, then it, AI won’t, won’t work for you. Um, I, you know, I’m the implementation of ai and I think, um, Gabriela and Christian, I want to ask this same question to you, but from two different, um, perspectives and Gabriela will start with you.
I, I wonder and, and thinking, you know, with what you’re doing with tabs and with what other, um, companies are doing out there, whether it’s it’s Datarails or Oracle or, or, or whoever, like for a lot of businesses, if you don’t, you know, going back to that initial survey question of um, you know, not having expertise in-house, I think for a lot of people, their first exposure with AI is gonna come from the software, from the SaaS tools that they’re using. It’s gonna be incorporated in that, and it’s not gonna be something that you build in-house. So if you could tell me about, you know, what users can expect to start seeing in these AI powered tools or maybe AI enhanced tools, like maybe it’s something they’re already using, but it’s getting new AI functionality. I mean, how do you think that will, uh, find its way into people’s workflows using, uh, third party software?
Gabriela Gutierrez:
Um, I think also like, or the most common use case, it’s like seeing more, uh, chat robots that you are able to chat with them and ask them to do like a report or take, like, I need to clean up this data. Can you remove the title? Can you add here and there? Um, but I would actually, like, or my way of how to really implement AI would be to start with a problem. Like for example, if we really want to, uh, automate something or improve something, let’s just start with a problem. And from there we can see if actually AI is the best way to do it. Because in many use cases, it could be that it’s a collaboration issue. So then we need to completely use a different tools. So for people out there, I would really like, like them to think about what are like the biggest pain points and then like if AI makes sense to take it, like, for example, build it in-house to build, like in-house AI, it requires a lot of like, IT work and software development work. So it would be a huge investment for a company. So usually that’s where you will outsource to different softwares. And then I would also like for finance, as we have this huge, um, knowledge gap of not knowing how to use Python or to code, then I think no code solutions with a chat interface that is easy to implement and to use, and especially that they care about the data quality, I think that’s kind of the way to go.
Glenn Hopper:
Yeah, makes, makes perfect sense. And Christian, I think you and I are, are kindred spirits here in, in this, and I think a lot of my, like, I never set out, I never wanted to be a, a coder. I didn’t wanna be a developer, but I wanted to get stuff done. And I knew back of the house, <laugh>, if I asked something from IT or the dev shop or whatever, we were always kicked to the back of the line. And so it’s like, well, if I wanted to get stuff done, I’m doing citizen coding, I’m running some bizzaro database, um, on my laptop instead of on a server, I’m doing this citizen coding. And I, not, not that I’m trying to advise anyone on the show to do this, but I think maybe I’m gonna throw this out on two levels. One is the, the warning of two companies, you need to have an AI policy in place and make it very clear what it is.
And I’m not saying it needs to be locked down draconian, but you need to understand what’s out there, how the people are gonna use it. Because I think what you’re seeing right now is for companies that don’t have a policy and don’t have a handle on this, then you know, they may tell their employees, nobody’s allowed to use AI for anything. Well, a big percent of your workforce is gonna become like secret cyborgs who are using it on the side, probably in dangerous ways, <laugh>. Um, but if you do have a policy and you’ve given them a playground, whether it’s an Azure environment or enterprise or whatever, you know, whatever the environment is, there are some things. ’cause I do believe that the way not gonna be rolled out, like dropping in a new ERP system, it’s gonna be rolled out. Um, you know, give the employees access, let them full democratization, let them find use cases for it. So if, if you’re in a company right now and you’re trying to figure out, uh, ways to use ai, what’s your approach to using these tools at this point?
Christian Martinez:
Yeah, definitely. So I’ll start first with the part of the, the getting stuff done that, that you mentioned that it’s very important. So how can, let’s say people start? Right , like, let me share my screen quickly and I prepare, uh, something for all of you. So I always advise to get started, um, as you mentioned, like not with your real data, but with data publicly available. So you can just go to Yahoo Finance, let’s say choose your, any company there, you can download already, like balance sheet, income savings and so on. But I just click here and then download this one and then go to ChatGPT4o oh, and just start with like, something very simple access management consultant and analytics data from F lens I uploaded already there where I downloaded from, uh, the public available domain and it will start analyzing.
So already been here, but um, for people to start seeing, it will start, uh, generating all of these different like analysis, like closing over price, volume over time, and so on and so on. So how do you take this with your real data after it does this analysis? You just need to ask them for the actual Python code and so on, explain it where you are using it, and then, um, then get the code. And then from there, then you go to actual Google code. That is kind of like the Google Docs of, um, programming. It’s very, you don’t, you don’t need to download anything and you just need to copy the code in there. And then right now I uploaded exactly same data, but the idea of this, you can upload in this environment, uh, your own data and then create exactly the same, uh, graphs and so on. So also like similar, like I’ll did it quite quickly in here, but for people if you want to follow along and so on, I’ll also drop down, uh, uh, quickly in the, in the chat where I basically give this type of example and all of these other, uh, examples to start with this like AI literacy in, in finance and fp and a.
Glenn Hopper:
That’s great. And I, you know, I, I’m thinking about had I not been, you know, neck deep and, and all this stuff, uh, for years, people can see that and it looks like this is magic. I don’t even understand what just happened. You know, he just <laugh> grabbed us a a spreadsheet and uploaded it and immediately we had all these charts and everything. And that’s why I say to people, like, just like you said, take public company data, plug it in, see what it does. Um, there was actually a paper, I’m sure you guys have all seen it, a paper that came out a couple weeks ago May, around May 20th, about, um, how <laugh>, uh, a these LLMs are, have proven in many instances better than humans in fp and a and that that paper, um, I think it was called using large language models for financial analysis.
It was something un unexciting like that, but it came with, uh, they had built their own custom GPT, and it’s actually a chain of thought prompt. And I’ve built a a million of these GPTs, and I, I know you guys have too, but it, this one actually did a really nice job prompting, and I’ve tested it with, uh, several public company financials. And I think, you know, whether you’re using a GPT or just the, the straight plugging it in there, I think if people haven’t had this moment already, that’s gonna be a real eyeopening kind of light bulb goes off, uh, moment, um, for everyone. Um, so I think to your point, get the public data, get out there, play around with it, and we’re, we’re already coming at the, uh, to the end of the show here. And I, I feel like, I mean, one hour is barely enough. We’ve, we’ve only scratched the surface and I know this group we could talk for another six, seven hours without, uh, taking a breath and still, uh, have a lot to go. Um, I do wanna see if there’s any questions that have, um, come through and that we haven’t answered. And while I’m looking through these, um, let’s maybe take a, take a minute here. Any closing thoughts, Nicholas? What, uh, anything you’d like to leave people with as they, as they walk away today?
Nicolas Boucher:
Yeah, so when you see, uh, the example you just said with this paper, when you see what ChatGPT 4o is doing with, uh, advanced analytics, uh, I’m surprised myself even though I was like looking at it, everything so closely, like how fast it is, uh, working now and how well with also, uh, financials or figures and competitions. And I think it’s because, um, we all need at, at the end that it’s, uh, when you do a calculation that it works. And at the beginning, all of the marketers were taking over GPTs because you could create and do creative content. So write ’cause GPT are excellent, uh, and LLMsto write, but the real problems are solved also with figures and not only with words. And because of all of the backlash that happened of all of this funny, uh, answers with numbers, I think now you can see that the teams, the engineers behind that, they took that seriously.
I think they are running like a mini step in between with where you have like validated codes that can run. So, and that’s how you should do it as well. If you build, build your own model, you should not let the user build your their own query. So imagine just SQL or Excel functions. You should have like a validated set of queries for those type of software and then you just map it. And I think that’s what they did now, so that’s why it works so well. Uh, but uh, yeah, like if you think the next step will be just APIs, when your LLMs will have the authorization to go in your Salesforce, in your SAP, in your, uh, QuickBooks, and we’ll just say, okay, based on the first six months, uh, here is your performance. Then we’ll reach really another level and, uh, I think, I don’t see any reason why we’ll not go there. The the ChatGPT 5or whatever they want call it is with APIs and it will be the next, uh, big change.
Glenn Hopper:
All right, we’re gonna try to stick to landing here, Gabriela. Some last, some final thoughts you’d like to, uh, leave people with as we go.
Gabriela Gutierrez:
Um, I will encourage everyone to go and try like the different tools. I think like it is just like we need to learn and everyone is learning, uh, right now. So I will just encourage everyone to go and try many different tools and yeah, we, anyway, all, all of us start from somewhere, so it is just a matter of time.
Glenn Hopper:
And Christian last thoughts?
Christian Martinez:
Yeah, I would say that for everyone, like don’t, um, like try to do new things with all of the different AI tools that are out there, like experiment, but also try to, uh, invest in learning. Like there are a lot of great, uh, content on there, great courses that if you do some, uh, course of let’s say advanced financial analysis with chat GPT or for Microsoft copilot and so on, for finance and for FP&A specifically, then you will really change the way you see all of these, uh, tools because they can help you a lot in FP&A.
Glenn Hopper:
Great, great. And one question, uh, that just came in, I do want to try to hit it before we go. Um, let’s see. Uh, any general guidance as to which AI to use for FP&A co-pilot? I think co-pilot’s getting there. To my mind, it’s not completely there yet. I really think the data analysis tool in ChatGPT is, is, is the best tool to use for as assuming again that you’re in that teams environment, the paid environment where your data’s not being used for training, not to go into the free environment. Does anybody in our LA in our last seconds here, anything other than data analysis through ChatGPT? Okay, I think that’s, I think that’s a great place to start. So, well, we, here we go and an hour has quickly gone by. So, uh, thank you everyone for joining. This is our, the first session for FP&A 2024. This was a session on AI and FP&A hype versus reality. Um, thank you to the audience and thank you to our panelists, Gabriela, Christian and Nicholas. Uh, thank you all for your great insights.
