Track 4: How to Be Ready for The Ai-pocalypse

Transcription:

Jacob Sperry (00:09):

Good Afternoon, everybody. Welcome back from lunch. My name is Jacob Sperry. I'm the Vice President of Customer Experience at Trulia. We're the sponsor of the session today. Today I'm excited to introduce our esteemed speaker and topic. The remainder of this introduction was written using ChatGPT-4o. Welcome to our session on the transformative impact of artificial intelligence and accounting. AI is rapidly, rapidly reshaping industries with accounting at the forefront. Recent research shows that accountants are highly exposed to AI emphasizing the need to understand its implications. Our speaker, Hitendra Patil, is President of Global Finance and Accounting Outsourcing Services at Datamatics Business Solutions, Inc. With over 20 years in the accounting industry, he leads DBSI Global Services. Hitendra is the author of Rise of the AI Accountants and the bestselling, The Definitive Success Guide to Client Accounting Services, CAS. He's been named one of accounting Today's Top 100 most influential people for seven years running and is a top social media influencer in accounting. He shares insights on CAS/CAAS practice growth and emerging technologies like AI, machine learning, blockchain, and cybersecurity. Today, he'll provide us with strategies to leverage AI in our practices. Please welcome Hitendra Patil.



Hitendra Patil (01:40):

Thank you for that kind introduction. And are you ready for the AI pocalypse? I think it's not arrived yet because if it did, nobody would be in this room. Right. Okay. Alright, so hopefully this works. Okay, there it is. And if you notice something here, it says Advantage Accountants, and I'll go through that real quick very soon. So if the year was 1984, this is the person who've been giving this presentation, and if the year was 2044, this is the person who would've been giving this presentation. It looks like me, but it's a deep fake because it's an ethical deep fake you can actually make out from the eyes. But the year is 2024. I'm here in blood and flesh. If you can come on stage and pinch me, I mean on my arm, then I would say, ouch. That proves that I'm not a Deep fake.



(02:36):

And by the way, if the Deep fake was here, he would say exactly the same thing, and I was supposed to spend about 20 minutes on this slide, but they wouldn't allow me and he gave me a great introduction. So I'll just skip through this real quick. Okay, so in the AI world, each one of you will either be a centaur or a cyborg. So what do I mean by that? So centaurs switch between AI and human tasks at the task level. So essentially you gather method information from AI and then you go and do your work or you generate some text and come to AI to refine it. So you're switching between tasks. So that's a centaur and a cyborg intertwines the effort with AI at the sub-task level. So essentially you go to ChatGPT, for example, and assign a persona and say, now you generate something and it generates something, you push back. So this is not good. Can you shorten it? Can you elaborate it? Whatever that is. So you're interacting. That's a subtask level intertwining, and this is according to a Harvard Business School study when they studied, how are people using AI?



(03:47):

You are competing with AI. How many of you really believe this show? Raise of hands? You are competing with AI or is AI competing with you? Okay, embracing, sorry, embracing. Embracing good. Okay. Right. So if you understood this as AI, you are competing against AI versus you are competing using AI. That's what I think you're saying, embracing. So whichever way you look at it, AI is omnipresent now. You got to be taking notice of it. You got to be looking at how you're going to implement this in whatever work that you're doing, right? So how to be ready for this pocalypse. Let's go through this. Before we jump into the real topic, I want everyone to be on the same page as to what exactly is AI. I'm going to define AI for this session. For the next 30 minutes, 40 minutes, whatever we talk, we will refer to AI as one single thing that we understand.



(04:53):

Then based on my own research, continuing research, what are the likely impacts of AI on the accounting profession? What are some of the specific ways in which AI can transform the work of accountants? Why are there certain AI related concerns? And what are those and how do you develop new skills and competencies essential for your success in the AI world? So that's how we are going to go through this. 30, 40 minutes of session, I conducted one little research on asking how will AI, your AI impact your practice? So this was meant for accountants, and this is the result that I got. This is pretty recent, and about two in three accountants said they are not sure as to how AI is going to impact their practice, their particular practice. So which means there is a lot of confusion about AI, there's a lot of hype about AI, there's too much overload of AI information coming at you, so it's hard to make sense at times.



(05:56):

So what exactly is AI? Human beings are blessed with five senses. The sense of sight, the sense of speech, sense of hearing, sense of touch, the smell and test. And of course the super organism called the brain, which makes sense of gazillions, of bits of information being thrown at us at the rapid speed and we still make sense of it. So in the AI world, these functions, I cannot see anything here. These functions are replaced by machine learning that takes care of the brain function. The site is replaced by computer vision. The natural language processing takes care of speech and hearing, and robotics takes care of the touch aspect. What about smell and taste? I don't think AI is going there as yet, and hopefully in our lifetimes we can see a computer actually sensing the smell, right? Okay, this monitor is not working. Sorry, AI.



(07:03):

Okay. Alright, now it's working. Alright. Okay. So if you remove the computer jargon from this, so essentially you're looking at learning vision, language processing, robotics. So if you take away these things that are computer related, so what you're looking at is what you have been doing as accountants. You've been learning, you've been of course language processing numbers, trusting. You have the vision for the future, not only of your practice, but also of your clients as businesses and you're always deep learning. So essentially AI is kind of mimicking you in certain sense. So I'm going to make three statements and if you feel that the first statement resonates with you, raise your finger index finger. I mean if you like the second one, do this. And if you like the third one, do that hard sign.



(08:03):

So accountants intelligence is greater than artificial intelligence. Who here believes this? Anyone raise your finger. I can't even see. Okay. AI accountants. That is accountants augmented by artificial intelligence that is greater and better than artificial intelligence show. We see hands going up. Okay, and AI accountants are better than accountants with no AI. Do you believe this? This is a sign. Okay, perfect. So any presentation on AI in today's world will not be complete without the mention of large language models of AI, just like ChatGPTs. So the basic concept here is essentially ChatGPT is just adding one word at a time and it does it so rapidly, so contextually correctly that it feels very natural. So what it does behind the scenes, it keeps asking over and over again given the text so far, what should the next word be? And that's how it keeps generating that.



(09:10):

So ChatGPT is basically trained on worldwide web, three years of data, four years of data, very large sample of human created text sourced from the web. Now therein lies the first hint that on the web not everything is accurate, correct? Even real even fake for that matter. So if that is the training mechanism that goes inside AI model, what can you expect from it? But at the same time, because the volume of this data is so large, more or less, it throws at you correct set of information, and then there is a trained neural net inside it that can generate text like this. This is the one that the model has been trained on, and neural net is essentially small bits of computing at several nodes, just like how brains neural networks. So it's not a central processor. So it distributes the work, gets the work done, and brings back the result.



(10:08):

So it starts with a human prompt. First time when I heard ChatGPT I went logged in and looked at the screen, nothing happened. Then I realized, oh, I need to ask it something. That's the prompt, right? And then from it there it continues with AI generated text. That's like what AI has been trained with. So it looks at what you asked and then goes back, finds out and tells you, okay, here is what makes the most sense given what you asked. So let's say we were to complete this sentence. The best thing about AI is its ability to. So what will a large language model do? So it has been trained on billions of pages of human written text and it finds all these instances of this particular text and then sees what word comes next after these so many words. And then it not only looks at how many words, it looks at certain sense, whether it's matching the meaning in terms of what you're asking and the end result behind the scenes, not what you see is a ranked list of words with their probabilities.



(11:11):

So here is what it looks like. So essentially each word is tokenized and I don't know if you can see below each colored box there is a little token number. So then it finds the next tokens. So learn is at 4.5%. So most likely AI will complete this particular sentence with the best thing about AI is its ability to learn. So that's the response that you'll get. Now let's complete this sentence. The best thing about accountants is their ability to, so when I post this query or a prompt into a large language model behind the scenes, it generates these numbers tokens. Now you've got to detokenize it and find the words and bring it back onto your screen. So it says the best thing about accountants is their ability to meticulously analyze financial data, providing invaluable insights and ensuring accuracy in financial reporting. Now, is this correct, more or less? Yes.



(12:14):

But is this comprehensive? Is this real? Now that's where the question comes. Is this the best definition of what accountants do? You can explain in many more words, but not necessarily AI has stopped here. And that's one more thing to notice that if AI has been trained on whatever, the data that went into training is going to generate things from that. So which means you might get answers that not necessarily reflect everything very accurately or comprehensively. So where is AI going to go from here? So we have seen large language models, we have seen image generation models, we have seen video generation models. So all of those are generative things. So there is data that goes in training and based on that it generates something. So from that angle, from that step, where is it going to go forward in terms of development? So it is going to follow the human evolution path.



(13:11):

So if you look at human history, the ability to use language or create language and speak and express came in pretty late in the human evolution cycle. The same thing happened with the computerized world computer's, ability to generate language that makes sense. Came in pretty recently and I used to try out an IBM Watson thing in 2015 ish and I installed it at my home and tried to use something. I didn't realize what it was doing and then I realized, oh wait, I need some data to go into it and then I can ask question in any which way. Doesn't matter the language in terms of how I'm constructing a sentence, it'll still fetch the response. So that's when I realized, oh, without data, there is no AI. Then I kind of predicted at that point in time, if I can just input text, why can't I input audio?



(14:04):

Why can't I input video and pictures and all of that? So that's what I mean by this interdisciplinary confluence, which is actually happening. I just saw some announcement by ChatGPT a few days ago that now you can throw things and it will still make sense, right? Edge computing plus AI. Now if AI has to be at a particular data center, if you will, and your data is going there, AI gets applied to it and you get it back. So your data is going somewhere, but not necessarily every single data piece can go there. For example, if Pentagon has to analyze something using AI, it cannot expose that data to any private AI servers. So which means AI has to go there. So that's what I mean by edge computing where AI comes down there. It could be useful in medical science also. And then the safety and ethics standards of AI are still evolving.



(15:00):

Currently, AI is all commercial entities creating it, not the government entities really, and which means they can use AI for anything. All your data that goes in and recite somewhere, even if you opt out of the model training the data based on what you uploaded, you can opt out. Even then the models will keep your data for 30 days because they have to make sure that nobody's using it for let's say creating a nuclear bomb for that matter. So which means that data is lying there somewhere. So not necessarily within the AI solutions, but somewhere, and that is where the risk comes in and then AI will become more autonomous. Today you are entering prompts, and if I have to give an example on accounting professions AI, maybe let's say a financial statement is generated, AI has been trained with several different ways in which you interpret financial statements.



(15:59):

In that case, based on the triggers, AI can actually generate some insights without you even asking it. So that's the autonomous AI that I'm looking at. So how is AI getting implemented in the accounting profession now? So there was again another research by an educational institute. So they looked at knowledge workers and how AI will impact knowledge workers and accountants or office of course knowledge workers. So they found that 17% of the top half skills saw an improvement by using AI. But when they looked at the bottom half skills, which is essentially data management data and into task, they saw 43% performance improvement in such tasks. So which is essentially a hint towards, okay, where can you implement AR right now to get the maximum productivity and efficiency gain? So those are data intensive things like reasoning, judgment, correlating external information in the minds of human beings to the data that you're looking, that's still not going in AI way as of now so much, right?



(17:06):

So it's all about structured data related, which means you've seen automation happening, bank feeds coming in, categorization or classifications happening automatically. You can even reconcile automatically. It's all about structured data. There are tables and tables and we've been doing accounting for ages in those softwares. So software has that data and hence they can figure out what are the most common things that generally happen in such a transaction. So that can be used as structured data, it can match invoices to payments. Then in audit world patterns and trends, it can identify more quickly than human beings and which means it can help in detection of anomalies, not just after they happen, even before they happen. If programming is done correctly, which means you can actually stop frauds or identify frauds much more in advance than what you can do Now for advisory purposes, there are lots of trends really you look at in a given business in the industry, industry and the business relationships in all of that.



(18:07):

So you're making sense of it in a human way, but this can help do that much faster, much more efficiently. So I looked at a lot of softwares and how they're trying to implement ai. So I found the software not here in the us so it sits behind or underneath an accounting software and it learns from the bookkeeper. As the bookkeeper is entering transaction classifying transactions, there's technically a bot that is assigned to that bookkeeper, it learns. So next time that information comes in either through bank feeds or scanned invoices and all that, this software helps bookkeeping to be done using what it learned from. So now what happens if there are five bookkeepers in a firm? So each of one of them has a bot and each of those bots learn and those bots can learn from each other. So which means your entire firm's knowledge is now democratized. So next time those transactions come in, you don't really need a very highly trained bookkeeper or a new bookkeeper waiting to be trained. They can get productive very quickly.



(19:14):

Then there is this workflow and client collaboration software, pretty pretty popular one and obviously they have seen millions of workflows being created in that software. So now that data is used to train up the AI model and now that can generate based on a prompt, let's say, how do I do a month hand close So it can generate a very detailed checklist so that even a new person can immediately follow those steps. Here is what we can do to do that work. So again, relatively inexperienced staff can get productive very quickly. Then there was a website that came on in some April, 2023. So there was a chat bot on the website that said, Hey, describe your tax situation and we will give you an AI tax answer. So I actually entered an SCORP information, not like what the company is, but a tax situation and it came back with more or less a correct answer and obviously that chat bot was trained using IRS is published data on the IRS website on applicable rules and all that.



(20:22):

But what happened next was a very curious thing. It prompted me that was an AI generated answer. If you want to connect with a human tax expert, we will connect you with that expert for $5. So that was just a connection fee, not a human experts advice to me. So they're using it not just for knowing the tax situation of a prospect, but they're using it for marketing and lead generation. So by the time if I entered, yes, I paid $5, this company would know here is Hitendra. He had this situation, he's asking for this, he's looking for this advice. So all your discovery call and all that your marketing and salespeople do just goes away.



(21:04):

Then a data regard, client communication. Today, if you find something when you're doing work for your clients, let's say you find a situation where you think that, oh, this seems to be out of scope work. So let me know down end of the day I will write to my client, Hey, we need to charge you more and all that stuff. You don't need to do that anymore. You can then in there say, Hey, now can you write an email that this is an out of scope work and it'll generate a nice little draft. You can correct it and you can send it. So essentially you're not leaving too much money on the table because you're taking action then and there and that work of writing an email becomes like a split second work. Then a sales tax software, very popular one. So they have this something called sales tax AI interactive.



(21:50):

So as user can ask the platform to calculate and research sales tax rates based on the user's location itself, and it generates pretty much a good response. I'm not sure whether it's perfectly 100% AI generated or is it just data matching, but it just does that interactive communication through words. Here is what it means. So that surely is a ChatGPT integration in that software. So what happens here is at the firm level, obviously you are helping your clients make faster sales tax decisions and more and better compliance. So all these things are going, this is just a dims of how the software vendors in the profession are implementing AI or integrating AI into their softwares. So where is it going to go from here onwards? What's the future?



(22:40):

Now it'll become unstructured data related, which means you can literally just bring in anything, throw it into AI, it can still make sense, which kind of happens even now. So if you take a financial statement and upload it to ChatGPT, hopefully not giving the client's name or anything, any PII information, it kind of knows what it is. But now let's say you are on an audit, physical audit and you're recording inventory and you are speaking as you record, you want to make notes. All of that can go in, whether it's happening now or not, it's a different story, but that is possible given that ChatGPT just released what it released last week to take all that audio, video, text inputs. More likely than not this will happen. So again, it's going towards mimicking human cognition skills. As a human auditor, when you go, you're absorbing so much information in different ways.



(23:32):

The same thing can happen in AI, audio, video, text, image. That's what I talked about. The confluence natural language processing, which is already making progress leaps and bounds will progress even further intelligent communications to help stakeholders interpret. I've seen something where you send a financial statement to your client and the client can actually ask a question to that financial statement while it's on the screen. Hey, how come my profit went down because I thought I did better sales? So it can find information and bring back response that is very well drafted. So that's again generative language, but that is possible, that interaction. Then it'll go towards customizing advisory services in an individual manner. So let's say you have 300 business clients. You know each of those clients well. You possibly know their risk appetite, their tolerance, their way of looking at opportunities, their current business situations, and maybe in the future the softwares will allow you to capture this information very specific to a given client.



(24:41):

Now what happens because of that? Next, when you generate a financial statement for client A, it'll be based on these customized information. You generate it for client C, it'll be based on that client's preferences, the same output but customized without too much of an effort. That initial effort of describing and defining the client persona might still be there, but you're able to do this industry specific AI ChatGPT is like a public AI. Anybody can use it. We still don't have it in this industry, but there is likely that an industry specific AI will come in. The way the data that is used is only from the accounting industry. So all the machine learning mechanisms are trained by accounting, tax, payroll, audit, that kind of data, private AI, the big fours are creating their own AI tools trained on their own massive data sets from last 30, 40 years.



(25:45):

So their own IPR now gets converted into an AI tool that any one of their thousands of staff members can use it internally, and that's not available to anyone outside of that's a private AI. Then you have customized AI. Let's say you're a full service firm and you think, okay, AI becomes too expensive to create ourselves. So where can we use AI? So let's say we think let's do it in CFO advisory services. That brings in little better profitability. So let's use that. So you can train that model only on those CFO advisory information within your firm. So that's the customized AI. Then there is this embedded AI power. There have been some companies that are trying to embed AI right at the source of the information, like for example in bank systems. So as the transactions are hitting the bank accounts, they're getting checked for tax applicability and those notes are inside those categorization classifications are inside.



(26:48):

So when you download a bank statement where the bank has this kind of embedded technology, you actually get tax ready transaction treated right? Which means you don't have really looked through all those transactions and this is already in place. A company in Israel has been working on this for two and a half years now, and this software will never be seen by accountant. This adds a bank level. So as I said, recording of transactions then and there in tax compliant ways, which means if this descends down onto your traditional accounting software, you should be able to generate tax saving and tax planning strategies then and there instead of looking at it periodically. You can do it possibly on a daily basis if that makes sense for some bigger clients of yours. For auditors, it might be continuous audit, 100% audit. It's quite likely there's already some software in audit world that identifies about 40% more transactions that are likely to be fraudulent.



(27:53):

If the AI technology was not there, human auditors would not be able to find those. So that's the continuous audit or very quick audit kind of stuff that AI can do it for you. Then where is AI headed in the profession? So I looked at some of the biggest softwares and their published AI roadmaps, pretty complex, always changing. But then I studied that and I figured out, okay, let me summarize this. So what's their AI roadmap? The fundamentals are these. So they are utilizing their own extensive data for training their own AI models. So now you can easily guess who are these companies, obviously the ones that have the largest data sets in the world. So what they're saying is they're making this to help accountants make improved decision making through this data intelligence from this entire world's data that they have in their systems.



(28:51):

They're also saying that they're focusing on augmentation of professionals not replacement. So technically you see advertisements, okay, you can do this, you don't need an accountant, whatever that is. But behind the scenes, whatever this AI stuff is getting done, they're saying they're going to use it only from the perspective of augmenting professionals, accountants, and then they're also doing the integration of the likes of ChatGPT into their models, which means whatever work has already been done by billions of dollars of investments. So while leave that work and create your own large language models, they're not doing that. So you'll possibly see best of both worlds, right? So this is from a positive aspect, but it still leaves some questions unanswered about the future. So how will your clients adapt to the AI driven software? Will they think that, okay, wait, I can do this on this software.



(29:51):

I can actually chat with the software. I get the same answer that my accountant gives me. So what will happen because of that? Will clients expect from you lesser data insights and more of your expertise to be applied to their business situations or financial situations? You have to really figure out the answer on that. How will clients' expectations change? Will AI in fusion address the talent shortage? We're all talking about the epic talent shortage in the profession and despite all the automation that has happened over so many years, we still have this talent shortage. But will AI make any change in this or will it be that now that data insights are all in AI's domain, we need more human interactions between accountants and clients, hence we will need more people. What's the answer? Hard to predict right now. And last but not the least, how will the regulators, let's say the IRS accept an AI generated tax return.



(30:51):

Will it still be certified by or signed by an accountant enrolled agent CPA? Or can we say, okay, this is generated perfectly, I just look at it, I'm an EA. Let me just upload it to IRS. Will that happen? What would be the rules around that? I think this will all keep evolving right now, the impact of AI on accountants work, again, this has been a very deep and vast subject to figure out how exactly is AI working and hence what is it likely to do in terms of how it'll impact accountants. So let's see some of those things. So there was a research by University of Pennsylvania along with open AI, and they said accountants and auditors are among the occupations with the highest exposure to ai, and that percentage of exposure was at 100%. So what does it really mean? Does it mean AI will replace accountants jobs or does it mean it'll replace some of the data intensive tasks that accountants are doing manually?



(31:54):

Now, one thing is for sure that it is going to redefine what you call work. So if you are used to starting your day with something, going through the process, ending your day, you, you're done some work, that definition of work is very likely to change very rapidly. So then you are going to be adapting skill sets to take on more analytical, more strategic and advisory roles. And I think you've been seeing this already. In fact, most of the software companies, the thought leaders, the consultants have all been encouraging you to go towards advisory. So not just because of automation, but because we see that this particular pocalypse is going to descend down onto the accounting profession and hence you need to adapt to it real quick. Then you'll be talking things like strategies and trends and risk managements and all that. And one more thing is you are going to be reviewing the work that AI does.



(32:57):

Now, this is very interesting. I'm sure you must have heard of the story of this particular lawyer attorney who cited six cases generated by ChatGPT and submitted them in the court in defense of his client. In court obviously has to verify those, and none of those cases were actually, they didn't happen. How did AI create that? It's called AI hallucination. And there's a huge explanation, technological explanation, very complex on why AI hallucinates, which means it's a necessity that whatever AI generates a human expert needs to really look at it and say, oh, this looks correct, or okay, maybe I should check this, maybe I should check that. So that is where we are at right now, and this is in 2015. I had actually published a LinkedIn post saying that AI robots will report to accountants. I didn't know that after nine years this is going to come true.



(33:49):

So how can AI transform your work? So let's take some little glimpse into this. I'm not going to go too much into this, but more automation research, quicker research. This is really, really useful. Anything that you want to research. Now, let's say you want to go on Google and find out a tax law or a regulation or a subregion. It takes bit of a time, but a tax trained AI model can bring up that information in split seconds so that now I can reduce my research effort by maybe 90%. That's a very useful thing as far as current AI installations are concerned, advanced analytics, risk managements, risk assessments, fraud detection. I am personally very much enamored by this. How is it going to figure out how frauds happen? How will the AI learn from the frauds data from millions of transactions, maybe 0.1% fraudulent transactions? How do they find patterns in that? And some companies are actually already doing that. Client interactions enhancement. It's already happening. Some firms are already putting chat bots that are armed with AI to make sure that the interaction seems more natural, more human. That's already happening. So if you went on a company's website and said, for customer support, go to this chat and it wouldn't understand. Gone are those days. Now these chat bots seem very, very intelligent.



(35:18):

I think if I had to summarize this, then I would say in the future, if you are more towards data in your day-to-day work, you're going to be in trouble. Versus if you go towards helping your clients make better decisions based on whatever tools you have, based on your expertise experience, then you'll be more in business, more successful. So current busyness that you have, maybe because of talent shortage, maybe because of too much of data overload. So if you divide that by AI plus OPTs and OPRs, that is other people's time and other people's resources, I think that's where data mad comes in and combine this, your business success will be much higher. So essentially you're using tools to optimize yourself and release your expertise more onto client situations or on your firm's situations to make sure that you're leveraging whatever you have to the best possible ability that you have.



(36:19):

Yet not everything is on keyori. There are data security and privacy concerns. There have been cases of data breaches even in AI models. So essentially the same story. Your data goes up there, your credit card numbers, your social security numbers, your registering, that data exists somewhere, not necessarily inside the AI tool, but somewhere. And that's a standard thing. No AI company is immune to this. So what are those concerns? Again, this is the largest one, privacy and data security concerns. Because at the AI company which has invested millions of dollars, there's always this pressure of monetizing what they have. So what stops anyone from unintentionally misusing that data? What happens if data gets hacked there? And plus the hackers will know. Now this AI tool has become so powerful, so popular, so they should have more financial data or maybe identifiable data. So let's go and try and hack them.



(37:23):

But first because of the companies will take care whichever way possible to make sure it does not happen. But that does really release that security concern in the minds of people accountability and transparency. Now, this is a concept in the way the AI works right now technologically, how does one explain what answer or what creation the AI tool did? Now, Google CEO is on record that they don't know how Google's bar generated certain thing, and that is because AI just generates something based on the data that it has been trained on. It's a very, very complex explanation as to why you cannot explain what AI did. And that is where for accountants, it becomes very critical that whatever you use AI for, you're looking at whatever it generated and you're making sure that it makes sense. That review thing doesn't go away. Biases and fairness.



(38:22):

Now, this was a pretty big surprise to me when I first read about it, started shredding about it. What do you mean by bias? So again, the data that goes in is what is the base for AI to generate whatever it generates. Now if that data itself is biased, it is going to generate bias responses. I'll give you an example. So I asked a large language model, tell me a joke about accountants. Say, what do you call an accountant who is seen talking to someone? The answer was popular or a social account. Now what does it mean? Does it mean there is a bias in the data itself from where did it get this information or how did it generate? This? Very recently tax season ended and I thought, okay, let me ask AI to generate something. This is the picture that it generated when I asked. Show me a firm, small firm just after the tax season, okay, all relieved people slightly happy, they're looking tired. But now look at the bias. Does it include all the races? Does it include all the ages?



(39:35):

So why did it generate like this? So which means all the data that's gone in accountants are predominantly X and Y and Z is what it is getting perpetuated. So unless you go very, very specific, Hey, create a picture of accountants, average age should be seven years. It'll still generate a picture, but that is what I mean by biases. It can get perpetuated, which means in some industries this can become really a problematic thing. For example, bank loans and if bank loans data in the past, loans approved shows that 43% of people from certain classes of the society, they did not get the loans. When such people come to get loans, AI would probably reject them. That's the perpetuation of bias. But then there are opportunities for accountants in the AI powered future. So what could be those? So I tried to summarize this rather than going into too much of details.



(40:37):

So essentially this AI tool created based on large volumes of data. It's like a black hole. You don't see anything in that. You just believe that that tool has been created well, and it'll generate proper information, which means any system that's built for AI must have a very comprehensive data set. So if that is not there, obviously it'll cause some wrong instances. But then your opportunity lies beyond this black hole or the black box. So that's in your clients is new data and this new data does not exist yet. It's in their heads, they're making their decisions and that is giving rise to this new data. And that is where your impact is, right? You are talking to your clients here, you want to increase your sales or okay, you want to control your costs, you want to bring down your inventory. Those are discussions happening between your mind and the client's mind based on certain existing data.



(41:31):

But the future data is still not in AI. It should go in and hence in the future, maybe based on the latest data it will generate. So again, another thing is your AI tool that you're using, getting trained continuously or it stopped getting trained. When did it stop getting trained? And it's not easy, it's not cheap to continuously train AI tools on massive amounts of data. So there's always going to be this trouble as to how reliable the tool can be. But beyond this, greenfield is your portfolio field, your customers of your clients, the vendors of your clients, they are making their business decisions. You're obviously telling them, okay, when negotiate the price for this particular goods with your vendor because the market price has gone down, whatever that is. So you still know beyond your clients' own world. So that still is not in the form of the data.



(42:25):

So that is another thing. And then beyond that is this gray world of business environment, macroeconomic factors, geopolitical factors, changing regulations, all of that is a very complex world and it has an effect on every single layer. So you are going to be the custodians of these external layers beyond the data. That's where your opportunities will be. Now look at it this way. So you are generating some kind of a tax research report based on what is existing and you see a customer's or client's transaction in a particular tax applicable way, the AI tool generates it. Now tomorrow the tax law changes about that transaction. Now what happens? That data is not there. Inside ai. This is a brand new thing. Nobody can go to AI to generate responses because AI will not have enough data to train itself on millions of transaction with the application of this new tax law.



(43:25):

So there's always this flux of situations that require human beings all the time, and hopefully the government keeps changing more laws. So that should help you. So how should you stay ahead of the AI curve? So it's not about the tools really. So if I was to tell you just one sentence in this particular presentation, it would be this, learn how AI itself is created. There's a way the mechanism, how AI tools are created, irrespective of which AI tool. There's a large language tool or a video tool or technology tool inside accounting software. Then learn how AI learns by itself. So there's unsupervised learning, there is supervised learning, two different things. One, supervises all the data, keeps going into the accounting, the AI software. But then is the correct data going in? Who's going to look at it? Is it fully unsupervised, semi? You need to know what your technology partner is doing, training the AI, and then you kind of try to learn how this AI tool will work within your forum.



(44:32):

So obviously if you ask the technology vendors whose software you use, they'll be more than happy to explain to you. There would be webinars, information documents and what not. But the point is, you must learn how this whole thing is going to work so that you can anticipate and expect what can you expect from this AI tool? So you don't want to start with solutions or tools of AI, but you want to start with what will you use AI for? Is it just for the heck of it? The world is using AI, so let's use AI. That's not where you want to go. So you want to first ideally look at the outcomes your clients are expecting in the AI world. If they're expecting different outcomes and you're not able to produce them or you're not able to produce them cost effectively, you want to then look for tools.



(45:19):

Can I do this better using an AI tool, right? So that's where you start with ideally. The second thing is of course for from internal operations, you want to look at your productivity, you want to look at your utilization, process, flows, steps, duplication, whatever that is. So there's a lot of opportunity for your firm to increase efficiency, productivity, and profitability. So these are overarching goals for you to plan your AI journey. You can use AI to enhance your existing services. You could use even ChatGPT today to express your outcomes and insights a little way better rather than scratching your heads as to what should I write anyone? Creating a new service capabilities. So let's say you offer 10 different types of services and your competitors are offering 12 different types of services. So can you learn those two different types of services quickly? More likely than not, your research will be completed within minutes if you use AI and ChatGPTs of the world.



(46:21):

So that gives you the baseline so that you can apply your mind, so you can actually create and launch your services much faster now. And the path to go there is standard, just like any other software, you try it on a couple of clients' data, not real data, but just copy, paste, whatever that is, try it, test it, implement for a handful of clients. If it works all good, then you expand to all your clients, to all your prospects and then just repeat, rinse and repeat. That's the cycle that you do, right? As such, you don't just buy any software, you actually buy what you can do with it. So same way for ai, what truly matters is how you will harness AI to deliver a better impact to your clients. And there is only one word in this slide that is very critically important.



(47:09):

That's impact. How do you measure your impact on your clients' business or financial lives? It's not the process, it's not just a software, it's not the tool, it's not the output. Nothing. Even if you deliver a good outcome in time, does it impact the client positively? And if you look at impact as the driver for your AI implementation journey, I think you'll be more successful. If that brings me to an end of this presentation, I invite all of you to become an AI accountant. My booth is right here in Datamatics. If you have any questions, you can always come and talk to me on this. If there are any questions, we still have four minutes to go. I would love to take those questions. Anyone? Anyone?



Audience Member 1 (48:03):

Oh, there we go. Hey,



Hitendra Patil (48:07):

Don't use ChatGPT to generate questions, please.



Audience Member 1 (48:14):

I run, but my head.



Audience Member 2 (48:16):

Okay. That's okay. First of all, thank you, Hitendra. What a wonderful, informative presentation. Thank you. And I have not used ChatGPT, and my question will probably expose that, but how does the information get into the AI tool? Do I have to specifically go to ChatGPT or one of the other tools to try and utilize it? I think we had a presentation yesterday from a firm called Spark. If I went there and I started to enter data, is that where my entry would be? And would it be only the data that I entered in that one specific interaction or if I search on Google or are there other ways that the data can get into the model?



Hitendra Patil (49:05):

Yeah, good question. So ChatGPT for that matter, there are plenty of modules behind the scenes. It did showcase its ability to calculate taxes and all that. Open AI CEO actually went into YouTube video and said that. So essentially there are tracks inside, we don't see them. It's all transparent to us. We just can't know what that is. But when something is uploaded, it kind of is trained on knowing what it is. So you generate the financial statement, upload, you generate a tax statement, a payroll report. It knows that this is a payroll report or a tax return, and then it goes and puts that data inside whichever way it tokenizes. So essentially nothing goes straightforward. Everything is converted into a number and gets stored with the probabilities and all that stuff, and it just fetches that very quickly, lightning speed to figure out what it is and then puts it in the right places.



(49:57):

So that's how, so you don't really have to worry about training it. But then behind the scenes, obviously they have worked maybe for a few years, billions of dollars on creating those learning abilities. So that's what machine learning is all about. So all those tracks and components are already in place. You really don't have to do anything. And for that matter, each one of us has gone into some websites and tried to log in and you see the pictures, identify the rabbit or whatever that is, and you click on that. You've been training that AI model for these many years without getting paid for it. Right? Okay. Any other question?



Audience Member 3 (50:41):

I was scratching, but I do have a question. Yeah, so our focus is on a specific niche hosts, Airbnb hosts, things like that. I would love to show everyone how they're doing compared to all our other clients. So I want to aggregate all of our client's data into this sort of AI. How could something like that be done?



Hitendra Patil (51:00):

It definitely can be done. That is the trends and patterns thing. If you ask a large language model a question as to what is the average profitability of a business of excise in this industry, it'll give you an answer with a disclaimer that, oh wait, my data is only up to 20, 21, 22, whatever that is. But it can make out those comparisons and patterns and trends. And that's a big promise of AI in the accounting profession itself for that matter. I know I worked for a software company, accounting software. We pulled in data from different industries, their financial ratios and whatnot, industry-wide ratios. And whenever a particular company's report was getting generated, it would compare that with the industry ratios and say, okay, here is how you compare with the industry. But that was pure data matching nothing about AI. But AI has that ability to absorb this information very quickly. You can throw tons of pieces and ChatGPT is a model by itself. If it can make sense of worldwide web data, which is absolutely unstructured. There obviously have been mechanisms in terms of how they train their machine learning algorithms to learn from it. And that is what I mentioned a little bit about supervised and unsupervised learnings. There are paths towards that. It's pretty complex to explain anybody who's technologically oriented. It's a fascinating field of study, how the whole AI thing works



(52:36):

And more



Audience Member 4 (52:40):

So I think a lot of us really expected AI to be a great tool to help us deal with new changes in the law. But you sort of said that that's really not what it's for, that it's not really a rules-based system and that it would take a long time to train it to react to new changes.



Hitendra Patil (53:03):

So as of now, the large language models are worldwide web data. And that's why I said industry specific. When that industry specific AI comes in, not created by just one company, not owned by one company industry-wide. If that happens, somebody takes that lead and trains up the AI tool, let's say, on all the tax laws, all the tax returns, all the tax cases, planning cases, IRS cases and whatnot. If that happens, AI can definitely do that. But who's going to do it? Who's going to invest in it? How are they going to monetize it? See, everything is commercial, nobody's doing it. Government is not doing it. So somebody has to do it. So more likely than not, let's say an accounting software company, they will rather focus on their own software and say, look, if you use my accounting software, you have the AI power with you so that you can do accounting faster, better, cheaper, whatever that is.



(54:01):

Anyone else? That's it. We are out of time. Yeah. Do we have one time for one more question? Yeah.



Audience Member 5 (54:11):

Thank you. So you mentioned that about the talent shortage and how these tools could or could not address it. Do you think that firms that do implement changes using AI might actually attract a younger generation of accountants, CPAs, and tax professionals because they're addressing parts of the job that aren't the most glamorous and maybe it'll bring them in.



Hitendra Patil (54:44):

As long as you express that very clearly as a firm, that people out there know that you are technologically advanced, modern AI driven firm, you are likely to attract that new talent better than a firm which does not do that. How are you going to express that? That's the question, right? And obviously, yes, people don't want to do this mundane work. I was at Toronto in a conference and there was this top 40 firms of Canada in that room. And the partners mostly experienced partners, they were in the room and they're complaining about the work ethic of the new generation. They said they don't want to get the shit done. I said, why do you expect them to do that when tools are there? So how are they going to learn? So why do they need to learn if AI has already learned it? So why can't you just tell them, here's the basics, and then take it forward from there?



(55:36):

Do they need to do that? So it's a time to really question every single thing that we do at our firm. Why the heck are we doing it? Is there a better way to do it? Is there technology that's taking care of it? Do we need to be more careful? Do we need to do more review? All of those answers are specific to your firm. But from talent point of view, yes. In fact, in India when we hire, we have 1200 people in this business now. And when we hire, obviously there are questions from the incoming talent. What's your system? How do you work? What happens with the standard booking? Do you use bank fees? The candidates are asking these questions, they don't want to come and start entering those things. Right? So those things obviously matter. Absolutely. Yes. Alright. Okay. So we have over shut the time. Thank you very much for listening to me. I hope it was useful. Thank you so much.