You Can Build A Working Analytics Workflow Without Coding
Transcript
Welcome To Indicators
Mike ChungWelcome to AutoCare on Air, a candid podcast for a curious industry. I'm Mike Chung, Senior Director of Market Intelligence at the Auto Care Association, and this is Indicators, where we identify and explore data that will help you monitor and forecast industry performance. This includes global economic data, industry indicators, and new data sources. Hello and welcome to another edition of Autocare on Air Indicators. I'm Mike Chung, and I'm happy to introduce to you Arthur Lee. Arthur and I worked together when he was at Click, which is a data analytics provider. They developed dashboards for business intelligence data. And he recently founded an AI analytics advisory called Fine. Arthur, welcome to the program. Hey Michael, thank you. Thank you for having me. This is a super exciting time to talk about this. Great. So tell us a little bit about your background, Fine, and the type of uh AI, LLM, analytics work that you do. Sure. Um I won't go all the way back to the beginning, but my original background was my business. I was a finance guy. I got into data and analytics pretty early in my career, and that's been kind of the majority of my career. And I think it's been a great, great foundation for where we are today with this advent of AI and large language models. So what what's super exciting right now is that uh companies uh with this new advent of this technology are looking to see how they can get better productivity and help augment uh all their employees to be um more productive, essentially. Oh, that's terrific. Yeah, what we do have fine is we help companies kind of, hey, what is AI LLMs? We advise them some enablement, some training, maybe some prototypes as well. And I what's interesting, I think, is sort of the confluence between technology and business, kind of how it comes together and how they can think about taking that, using this technology set, taking that data, get really good, actionable uh information for their organizations. Terrific. So it sounds like uh you're probably working across a variety of industries. Is that fair to say? I do. I'm an I'm an independent advisory firm here, so I have a couple of clients. Uh but it's very, very interesting. I think what's interesting in my prior prior uh experiences at CLEG and other BI firms, uh companies, is that I've worked with a lot of different industries, verticals, departments within those verticals, and the technology set is pretty much applicable to all of them, essentially. So it's it's a really super exciting time. Still early, I would say. We could talk about that later, but it's a super exciting time right now.
Arthur’s Path Into AI Analytics
Mike ChungTerrific. And I think you had a couple of examples of recent work that you've uh highlighted. Um would you like to share that with our audience? Yeah, absolutely. So there's uh two in particular, and I'll give you a couple use cases, if we're one particular that I've been doing a lot of work for. So without go give into names and industries, um, they're a B2B software company, obviously in mid-market size. Um a couple use cases here is pretty interesting. Uh, one, which is which is pretty topical. As you probably imagine, when you have a conversation, there's usually most of these tools like like Zoom, Google Meet give you a transcript and a summary of the transcript. Now, this BDB company, they have all these sales calls uh at the top of that funnel, the business covers calls. So one of the the sales people get some information from the existing tooling, but the marketing department wanted to understand um what a cost what are these prospects saying? And how can we use that content to create content to use their language at the top of the funnel? So one is create content and then also map how can we map what we uh what LLMs, if they're asked what questions are that they could use potentially asking to the LLM, so we can map the content to the type of questions that maybe actually put it in that chatbot, for example, like ChatGPT. So super interesting use case. Um, I think that's something that um AIs in particular are particularly useful for. Another use case, the sales team, they, and this is pretty typical, you have a business discovery call and you create this uh document of follow-up. Here's what we discussed, here's what you said, here's the follow-up. Very particular format, it repeats all the time. And they were doing this manually. They would get the transcript. Now the AI-genated transcripts help, but they used to have their format. So we created a process using AI and LLMs to kind of automate the whole process. Get the transcript, reformat it into their sections, and also format it into a Google Doc that they have particularly. And that's fully automated. They don't have to look at it anymore. They used to do this all manually before. So super prick. They know the answer, they just have to create the content, let the AI do the content and generate the format. It streamlines the processes quite a bit then, it sounds like. Yeah, it does. I think on both cases. One is more interesting because it's kind of like top of the funnel experimentation. The other is something they had a process for, and you know, it's just soaking up time, right? Where they can spend more time calling, you know, calling, more prospecting, etc. Yeah, I'm sorry. And I think with the first example, what kind of struck me is making sure you're talking the right language that kind of shows that you're familiar with the vernacular, the business of speak, all the quote unquote jargon, right? And that really kind of gets you on the ground floor in meaningful discussions, I think. Exactly right, Mike. And I I think the sort of other perspective for your audience here is there's a lot of data you have that these LLMs can unlock in different ways.
Automating Sales Follow-Ups And Content
Mike ChungUh, that's sort of the power now. So all these use cases were they were always there, but they're always very hard to do or manual intensive to do or expensive to do. Now at LLMs, they're much they're very possible and much more affordable and cost effective, essentially. And with all the increases in computing horsepower, for lack of a better word, it makes it uh this type of analysis more accessible to people who are not necessarily as um highly trained in computer science, for example. Aaron Powell That's correct. There's still, I would say, a lot of work to do here, but LLMs and AI make the universe of possible capabilities or solutions to you or I, who are not programmers, quite possible. Uh to be clear, I never programmed in my entire life. I'm maybe a little more technical than your average business user. My background is finance. What I find by using AI LLM is I am building programs without even looking at the code, which is super exciting. Super scary for some people, but I think super exciting for me personally. Right, right. Well, um, I know you have a demo, and I think when you're talking about data and how we can kind of apply it to our uh whatever uh situation we're at, the type of company we're at, the company, the data that we have, or external data sets. Um, I'm really looking forward to the demo. So I think is now a good time to sort of talk about what you have in store for
Agent-Built NHTSA Dashboard Demo
Mike Chungus. Yeah, let me share my screen. And as you're as you're bringing that up for our audience, we're we're looking at NHTSA recall data, so National Highway Transportation Safety Administration. And it's a public uh data set, and Arthur's gonna kind of walk through it. And if you're um not watching this on YouTube, we encourage you to make sure you take this in on YouTube so you can see the live action demo. So um, Arthur, I can see your screen, so please take it away. Sure. So I'm uh fairly new to your uh your industry here. I mean it's very familiar as a car driver. Um so I uh I thought a good data set would be the NHT NHTSA data set. It's open, it's open, it's public. Um just to show you what is possible. Um and I'm gonna show you this uh in three different flavors. One, let's call it what you're seeing here. How I use, without going with all the mechanics, how I use agent to build this dashboard on their NHTSA data set, number one. Number two is how to use the existing tooling you might have seen already in Chat GPT, Claude, or Gemini, for example. How can you use that tooling with the data set you may get from this uh public data source as well? And how can you do it today? So I'm gonna walk you through sort of progressions, how to think about this progression to eventually you could build something like this. Now, one little twist, Mike, here is that uh I actually not only took a slice of this data, I took the whole universe of data. I've got a little bit ambitious here. So, what you're seeing here, and I'll kind of walk you through where you can get this data, just to show you kind of why it's interesting how these agents can help you. Great. And what we're looking at right now is the dashboard you built from the NHTSA database. Correct. This is the dashboard we uh I built. Watch you say um I told the agent to build. Okay. Uh I want to be clear on this point. This is the magic of these uh AI L agents. You're gonna hear this all, you probably your audience may be hearing this word a lot, agents. This is the year of the agents for AI. So this coding agent built this dashboard uh on the data set. Now, what's interesting about this is that I did not download one file. So I'm gonna walk you through this. I did not code one line of code, uh, I just described the design I wanted and it built this. And I iterated just talking or typing in my case, talking and typing, by the way, uh, to the agent. So it gives you a sense of the possibilities. If you if say you're a business analyst, you know, if you want to deliver something to your stakeholders, you can design something pretty nice. And it's very presentable, as you can tell here. And I did this probably less than an hour, essentially. Amazing. Yeah. So let me kind of walk you through this because I think it's interesting to see. So for the audience here, I'm going to the website where you can get the data. It's essentially a bunch of zip files here, right? So if you were doing this and say you had to do this yourself, Mike, what what would you have to do? You first have to download all the data, all these zip files, explore them, and then understand what is in the data set. Now, if you're in the industry, you might understand all of it, but you still have to see what the columns are, how they join, etc. Okay. So that's one. The other is the website itself, they gave you a dashboard, but I would say the dashboard's like okay, right? And you can't change it. It's not for you or me or for the audience. So, how can you do this? So when you look at this dashboard, what were the steps involved? Uh, you spin up the agent, you just type what you talk to, you can actually talk to it. Hey, I am doing this analysis, go to this website, download the data, and show me some show me some information about the data. Let's discuss how we want to show it in the dashboard. If you do that, and actually it will go off, and I will do something else for the next five or 10 minutes. It will do all that background work for me, or for you, or for your audience, whoever decides to use it. And it will construct uh what it did for me, kind of a skeleton of what I wanted. And then I describe, oh, because I was in the, as you mentioned, I used to work at Click. I know a lot about dashboards and BI. I say, oh, what I really want is this. I want bar charts, I want summaries. And you see here for this data set, this is the entire universe across complaints, investigations, and recall. It shows it uh over time, and then it shows a rolling six months. So by year over time for each of these uh metrics that the NH NHTSA provides, and then it shows a rolling six months as well. So if you are looking at this if in your industry, if you can need to see something particular, you can start like looking at the aggregate and then also drill down if you wanted to to get more information. What's interesting about the agents is once you get this, you can just talk to the agent to morph this data to something else. And I'll show you an example of that. Just kind of walking through the dashboard some more, it kind of shows you that the top uh things that are currently happening, like what's kind of bubbling to the top, essentially. In this case, the fuel proportion system in terms of the number of components of complaints they're getting, uh, which one are causing which of the impacts. You have complaints and the impacts of the complaint, in this case, crashes, injuries, etc. And then you can actually see like how things match up in terms of investigations and to recalls as well. This is just a sampling of the possibility you can do uh using these agents, essentially. So if I can just pause you there for one moment, Arthur. So you create you use the agent to s and you specified the metrics of interest and you specified the source of the data. And I would imagine if you're building this type of dashboard, you can also say, I want you to look in these other sources for news items to perhaps link those kind of trouble areas that were bubbling up. I think you mentioned uh, could you scroll down for me? I I I've forgotten the uh um so it could be fuel propulsion systems. Um, find me three articles where this is a headline or electrical systems. Similarly, look to these other external sources so you can marry them up and have multiple sources being sourced for this particular dashboard if you wanted to. And Mike, that's like a one of the key unlocks of the using this technology set. It's really combining the public data sets or multiple, it doesn't matter if they're public, but multiple data sets. And what the what's great about these LLMs, because it's data, I mean, generally speaking, they can find the relationships and kind of marry the data for you to find these relationships. It kind of does all the statistical hard stuff for you. Now, sure, admittedly, you need to know a little bit about it because you have to you still have to review the data. We we'll talk about the process of doing this a little bit later, I think. The things to look for, but you still have to look at the data to make sure it's right. Like I had to actually look at it, I actually told it to make some changes because it wasn't doing exactly what I thought was correct, essentially. Sure. It's like the trust but verify. It that's perfectly correct. You you got it, you got it. And this is again, this is nothing uh I always say this when people think about AI. You don't throw away uh the old trusted standby processes that you had before. Like that still applies, even though it's AI and LMs. It just unlocks more things you can think about doing, essentially. Now, this is a quick dashboard I built. I also built something, uh, if you see this tab here, I'm gonna tab on it here. It's called the rear visibility watch. I just focus on one sliver of recalls, essentially. And it's just looking at one particular thing. And you can imagine the system looking at any slice of data that you might be interested in, and it could re-download every week if you want. And the engine the agents can just run in the background. There's a lot of setup you have to do. It's not, I would say, for the the if you're just new to it, you can set this up right away. We're not quite there yet. But we're getting there. I would say we're getting there. Uh, but the possibilities of what you can do and marry other data sets to this is is is pretty interesting uh from what what you could do previously. Absolutely. One other note uh and Mike, what you mentioned before, one thing I I I started doing by uh I was getting super excited with the data, but uh because I'm a data guy. Uh I I I didn't I didn't quite uh get there. Is how can you look at complaints, investigator recall together and see what the relationships are? I didn't quite do that analysis, but it's there. All the data is there. You could easily kind of take the the very three distinct data sets from the this public data set to do that as well. Oh, that makes perfect sense because I've used some um AI agents or AI platforms to analyze a document, give me a summary, and then I'll ask follow-up questions. So what you're saying, I think, is in line with that. Now that you have this baseline data that's being reported out, you can dig in for more details, it sounds like. Correct, correct, absolutely. Now, this is like the uh, I would say, um, kind of more a little more advanced use case, but how can uh how can you get started?
Starting With ChatGPT And Small Data
Mike ChungAnd my and one of my biggest uh advice or strongest pieces of advice to give your audience is it if you haven't gotten started, you should get started, is my is my suggestion. So let me kind of show you the progression, how you can shrink the paths, how you can get to a dashboard eventually as you get more comfortable with uh the technology set. So how you can get started uh here, I'm now uh for the audience here, is I'm looking at Chat GPT. Uh it's just a chat interface where you it's a chat, you talk to it, it gives you an answer. I upload a subset of the data. One of the things that uh what's nice about the agent is the data sets. You can get you can use larger data sets, right? Because that NHTSA data is millions of rows essentially. In the chat interface, there are limits. So that's one thing to think through here, which is fine, I think. I think most data sets that most people using this will be okay with. As you use the first chat interface, I would think about this more sort of think about it as ad hoc. So the way to think about the process of how to automate using AI is you have a lot of ad ad hoc questions. As you do and you get more comfortable using ad hoc questions, some of these you do repeatedly, right? You want to automate. And we'll talk about that. Like the thing you want to automate, and then once from automation within the chat, then how you want to do a complete workflow. Like that B2B software company, they do a complete workflow essentially. Okay. That's how you think about the agent, is the whole workflow. All right, so hey, I'm here, I'm here in Chat GPT. I'm not gonna go through the progression here live because it just takes time. Uh, I uploaded this data set, only about 399 rows. I don't know why that number exactly. I ask it a question. In this case, please review this recall data set, give me a short business-friendly summary. It gives me sort of a summary. Um, what are the biggest themes? What's the most important recall campaigns? Makes vehicle groups. It just gives me a report based on the prompt that we gave it. Now, the prompt is just a question you ask based on the data set you have. What's really nice about this is that you can ask multiple questions, right? You can have a certain report. And by the way, there's all kinds of prompt tipping and optimization, which is a whole other, like a whole other like world essentially. You could do training on that specifically. But uh here I have it, uh, to keep it simple. Uh but then you saw there's no chart. So I asked it, hey, create an instant interesting chart, and it creates a bar chart for me, right? And you can ask for any kind of chart that you like. Okay. Sure. Now uh you can imagine if you do this over and over again, if you do the same thing repeatedly, you want to make this like a template. Like, hey, I have this data, I'll upload the data. I want the same look and report over and over again. So I do it every week, every month, whatever the time scale is. Uh before I show you that, does that make sense, Mike? Any other questions? Absolutely. And just to kind of go along with that, it it goes along um with your recommendation to if you're not already experimenting, now is the time because you get practice asking the question, refining it, and then for an agent to make it a repeatable business process, you sort of get to the point where you've kind of experimented with it, you've landed on a format that you like, and you have the appropriately phrased prompt so that you can just have it running for you almost like your employee that's just automatically doing this at a set cadence. That's that's absolutely right. One thing I want to build upon when you said about the the coworker or employee. One thing the way to think about working with these chatbots is think of it as the smartest intern that you've hired that doesn't know anything about you, your job, uh, or what you want. Right. So if if you imagine you're working with this intern, what would you have to give that intern to do a good job? Right. You s similarly, you have to do the same thing, generally speaking, with the chatbot. I found a lot of people say, Hey Arthur, I gave this question, what happened? I and then my response is what question did you ask? Did it know what you were trying to do? What was the context? What was the background? What you wanted? What was good? Was a good answer? What's a bad answer? Did you give it any of that? Uh no. Okay, well, that's part of the theory. It's not a mind reader, but yeah. Not yet. Hopefully, you know, with the neuralink, who knows? But uh what Elon. But uh anyway, um, yeah, so I think the context question is an important one. I do think think of it like an employee, sort of is sort of a good framework to think you're chatting with someone to get to help you get to your solution or your answer, essentially. Right.
Privacy, Security, And AI Policy Basics
Mike ChungAnd if I can add one. Another thing, you mentioned that the NHTSA or NHTA data set, very large. And I know that a lot of companies might be thinking about security of documents, right? So I think that's something to certainly do a little bit of legwork on to say, whatever platform I'm using, if I'm having it reach into my hard drive or in a shared drive, making sure you're in touch with your systems team to say, are these files secure here? What is our preferred AI platform? And then I could just simply point to the file over there and not have to worry about uploading a 100-megabyte access database or something like that. Yeah, Mike, you're pointing out something really uh important. Um to make if you're a if you're individual, is one question, but if you're within a company and you you're kind of straight could kind of creating or starting this journey and creating a strategy and policy, everything you mentioned is sort of should be uh top of mind and decided essentially. What what tools do we allow? Like, for example, is it open AI, is it cloud, AWS, it doesn't matter. Decide on the tooling, decide on the privacy and data controls, um, and be very clear, right? Like what can you use and not use? And then I think what I find is that users are more comfortable. I think a lot of companies start and then users have I I say this in some other cases where in their Slack chatter they're saying, can we do this? Are we allowed? Who can I ask? So having a good policy and the cut uh the employees understand the policy, the better off, and the more the more experimentation will happen within that organization. Absolutely, because I'm sure people that are listening to this uh episode are thinking, well, the data that I use is proprietary. I have to be very careful with it, not getting outside of these four walls, if you will. So, yeah, great point. Governance, making sure you're on board with the policy and that it's established is really a foundational building block. And I think the nice thing now, um, and again, this is still an early and emerging technology, a lot of organizations have figured this out. So, you know, as they say, you know, beg, borrow, and copy, you know, look at other patterns and and and just see if it fits your pattern or not. And find you're gonna find people kind of loose, people are very strict, and you you gotta pick which one you want and see, you know, beg and borrow and steal from those policies, essentially. Sure.
Projects That Standardize Weekly Reporting
Mike ChungOkay, so back to the demo here. So now that hey, I want I have a standard format I'm thinking about. Uh these products have a capability in Chat GPT, they're called custom GPTs, and it's something called a project. Okay. So here's a project. So what's a project? A project is some place where uh you can upload data, like pre like the approved data set, like you mentioned, Mike. Like, hey, I have approved data sets I can use, and you give it a preset instruction. Hey, I always want uh the report to look this way, for example. And then now when you ask questions to this project, it's called Gems and Gemini, and it called it is called Project and Claude. Um, it's gonna always follow those instructions. Okay. So for example, here in this uh in this chat, in this particular uh data set, which is the same data set, I gave it an instruction to give me the report in a certain format. So this is if I anytime I ask this question, I'm always gonna get the format of, and for the audience looking here, executive summary, map main patterns, charts. How does a chart look? You notice the little footnotes that I have one data set, if I have multiple data sets of multiple documents, it actually references the documents that it's actually pulling from, which is great from a traceability and confidence. This is something new that's kind of happened in the past year, actually. Um it just gives you sort of I feel there was always a question of hallucination. Like, where are you getting the sure? Right. Um now you can actually point to the data set or the document, which is quite nice. So now you have a repeated sort of way of doing it that is always repetitive. So now if I want to do this every week, I would just ask the question, I'll maybe upload a new data set, you could you can actually point to Google Drive or Dropbox, wherever your data set is, it could pull in new data and it could just you could ask the question and run the report. And what's nice about this, Mike, because I used to be a financial analyst, I don't know if this happened to people. I give the report to my CFO or my manager, and they might ask a question. Now I can ask questions, right? Oh, what happened from last week to this week? And I'll actually get an answer here as well, which is which is quite nice. And that's sort of the power here is not only you're getting sort of the artifact, but you can ask questions around the artifact from yourself and from others. And there's no new capabilities coming where you can share this chat to other people, it's that security around it. So it's a a lot of sort of the enterprise features, they're kind of rolling out and coming to uh uh come to all these platforms. Fantastic. One other thing to show here. So I have this. What if I wanted to download this as a PDF or doc? So here, as this comes up, I I put this in a canvas. A canvas is just a way where you can edit the doc yourself. So I want to put some more commentary to produce this document, but maybe I want to put I saw something in this. I'm gonna I can edit it. But what I want to show, you can actually download it here, and I'm gonna show, I'm just clicking on this uh uh sort of what they call canvas here in ChatGPT. Now you have options to download as a PDF document, as a Microsoft Word document, or a markdown document. Markdown is very specific, but uh most people will do a PDF or Word doc here, which is quite nice. Sure. Now hand it out. Not it doesn't be in chat anymore. I can just make it a Word doc, clean it up, and then distribute it any way I like, which is quite nice as well. Um any questions? Anything you you you sort of you you see here in terms of a project? I don't think so. I mean, thanks for showing this demo to us, and uh it's just remarkable how how how much easier it makes uh analyzing data, creating reports, asking questions, asking for recommendations so you can just be that much more prepared when you're going and sharing this with whatever audience you have. Well, yeah, and you know, one of the things I thought about too, Arthur, is um it must have been like an information technology class when I was in business school that I remember the professor talking about. And this was 20 years ago, right? So kind of that that dashboard that a CEO might want to have, and what systems is it drawing from to give say monthly performance for the company, so total and by business unit, monthly sales, cost of goods sold, profits, competitor activity. And it used to be a a little bit more um more work involved, and now it's it's like you said, you can just talk to your agent, describe what you want, and have your own dashboard wherever you sit in the company to report on the metrics that matter, and um just just systematize it to for streamlined operations. It's it's phenomenal. I I think you got a spot on, Mike. I think what makes it um what makes us sort of uh accelerate within the organization is ensuring you when you when you start using it, I would suggest if those are sort of very early in your journey, as you use it, as you share it, ensure that uh what you're sharing has trust involved, meaning you build it in a where in a way where it's easy to verify the information. Like any dashboard, it by the way, this
Building Trust Through Verification
Mike Chungis not AI specific. If you show a dashboard that's not correct, the trust in the entire dashboard goes down. It's the same thing with the AI systems. So if you see something that's not correct, um you know, no one's gonna trust it. But as you build trust, the the rate of information exchange, I would say, exponentially increases. Let me give you an example, Mike. When I was to be in finance, we used to have these, I used to be working retail, we used to have these weekly forecast meetings, and the questions would come up. And sometimes the managers would have answers, sometimes they would not. They would have to go, hey, we have to go back to the system, look at it, we'll tell you next week. Now, in theory, in the meeting, you can ask your chat interface, depending on how you set it up or your agent, you get the answer right there. Right? So having instead of having this delay, you can have, once you have high trust in the system, you can you can it make decisions a lot rap more rapidly, and you don't have this like delay, uh, which can impact the decision essentially. And one thing I'm thinking about too, Arthur, is that sort of subject matter expertise. So um if you've been at a company, if you've been in an industry for several years, you sort of presumably know the industry and know the company and what the uh nuances of any data set are, and what might not be uh captured in a data set. So, just as an example, we might see a monthly dashboard of sales and performance and profits, but what might not be captured there is uh sentiment of our clients because based on face-to-face interactions or macroeconomic conditions or geopolitical things that may create some headwinds. So I think that's where um, alongside the trust, um, I think that's a good thing for us human workers, right? In terms of yes, I have a dashboard. If we have the right data and we're asking the questions in the right questions in the right way, and we're getting uh an accurate analysis, I can also add to it based on my own um intelligence, experience, industry expertise. Is that fair to say? That is that was exactly my right, Mike. It's kind of a uh to loop back a little bit. Um you think about this AI agent as a very smart intern. It doesn't know what you know, it doesn't know what I know. Um it it needs to do a good to do a really good um process to get a really good outcome, that person needs to really, really be experts, a domain expert. It's not to say in the future 10, 20 years, some of these will be have very good domain expertise. Uh I think that we're a long way from there. Uh because the reality is, and we all know this, every business, every department is is is different. There's slight differences. Like it it at Click, no one had the same, every dashboard was different, right? In even the similar industries. So that there's these there there and along as it's humans, which you know let's hope for a very long time, I think I I think this this is always gonna be the case. Uh it's not to dismiss the technology, it's quite powerful. Uh the humans always gonna be in the, as they say, human in the loop. We're still in the loop. In a different way, perhaps, but we're gonna be in the loop. Right. And like you said, and certainly I hope I didn't come off as dismissing of any technology, um, because the technology, the capabilities are are utterly fascinating. And to harness um the technologies to um make sense, if you will, of data and to inform our decision processes operationally, strategically, it's so much faster. The feedback loop is just like you said, so much more quick than it used to be. And I think it it's it's hopeful for those who embrace it to your point of um experimenting. Get on, go go for it, you know, give it a try. And because it can just augment what you're already doing, and you could use it as as uh to your uh to your advantage. Um I mean middlely I'm biased here, um, but I uh I'm trying to be objective. I my background is finance again. Um I've helped I've helped a lot of accounting departments when I was in finance, and I've seen a lot of parts of finance, you tend to kind of touch a lot of parts of the business because you need to understand it. Um this technology can be used and and it's a technology, it's a tool. I think it's important to think of it that way. Um it it does impact, it's gonna help everyone if you wanted to. You don't have to. Uh, but I I do think uh it behoo it behooves you to use it. The one other thing I want to stress here, Mike, I think which is uh what I find, to be really good at this stresses your ability to think systematically. Because like I say, if I am talking to an intern or a person who's new to the company, I have to be very clear of exactly what I want. So, in a good way, if you're a good manager or good individual contributor, you do that, right? So it should help you. But the agent doesn't know what you know, so you you're still driving it, essentially. It's like a super you're driving to you used to drive a Toyota, now you've now we have a Ferrari, right? So if you don't know where you're going, be careful. It's gonna get you there faster, essentially. You just want sort of analogy to think of it. Yeah, all great advice. And I was thinking of I love the intern analogy and just you know, setting the context, providing the information and making things clear. Because, like you allude to, um, a good manager will train an employee well, a good user of AI technology will frame things appropriately, and it's something we can practice and improve at. So that's it's very encouraging. I mean, uh thinking about applying these technologies um beyond just whether it's automotive aftermarket, business and industry. Any tips in thinking about how we can use this beyond our workplaces? Yeah, definitely. I mean, I uh to be clear, I use this every day. Um but I'll give you two recent examples that are not business related. So uh my daughter, it's both related to my daughter. She's uh a junior, she's doing a semester abroad in Spain, in uh Madrid. Beautiful city. Uh so two things. One, um my daughter, uh, we're keeping track of her uh food expenses. It's all this 529 tracking thing. If you have a 529 account, tracking expenses, so it's properly accounted for. Again, I'm a finance guy. Uh how do we do this easily for her and me? Like I don't have to like in an expelled spreadsheet looking at the receipt and transcribing it and translating it because it's all in Spanish. So I built an agent where uh my daughter has to screenshot a receipt into Apple notes. Have the
Personal AI Agents For Real Life
Mike Chungagent read that every week, looks at the screenshot, looks at the receipt, looks at the date, looks at the description, translates it, decides if it's food or beverage or board related. And every week I get a report saying this is how much she spent, this is how much you should reimburse her. And then she had a she uh I had or behind her budget for the year. And I just it just happens every week now. And I bit that I built that later literally, it's some testing of all, but literally like in two or three hours. Incredible. So that's it. It almost makes expense reports fun. Well, it should be like you don't want it, you don't want it, you just want it to work. You know, you have to type it all. I think by the way, a lot of our expenses have been doing this, but it's it's just interesting how uh the universe of like these personal or sort of like long tail applications are possible because of AI. Uh another example in Madrid, my wife, we planned to stay to spend some time with our daughter, spend uh uh about two weeks there traveling, and there was an issue with the trains and weather. So we had some Airbnb lines up that we couldn't go to because the trains wouldn't weren't going there, and I couldn't get flights. They were too expensive, and there weren't any the time span. Um, so one I I used this case chat GPT. It worked you know, where else? Where should we go? That helped, of course. But the interesting thing here is I have a Chase card, credit card shout out to Chase. Uh they have this travel cancellation and travel interruption uh insurance if you use the credit card. So uh I mean I uploaded the whole like legal doc, the sort of agreements. Hey, what could I do here? So, oh what you can do is this is called a travel interruption, and you gotta submit this. And what it did was I had this agent just pull all the news reports about the weather and train, pull my email uh if you give it access, pull my emails related to the Airbnb cancellation. It did like it assembled all the paperwork for me, and it told me how to upload it to the Chase claim site. And it wrote all my letters, actually. So it literally saved me hours, and I got and we got the claim. Like the Airbnb got canceled, we got the claim back from Chase. So that was a couple hundred dollars. Thanks for sharing those stories. You're inspiring me to think of my own AI projects. So just one short example is this. And I was uh when I download photos from my phone, I guess it goes the Google photos, and I download them, and I'm I I relabel them so it's like 2026-03-25, and it might be Skiing 01 and then through 20. And it's a little bit cumbersome. So I would imagine before all of the AI um platforms became available, I could have written some code to do that for me, but I'm guessing it's a lot easier to make those types of mechanical relabeling of files a lot quite a bit faster. So you've inspired me, Arthur. Thank you. It's it's it's definitely possible. And uh I'm already thinking in my head. Well one last note here to think through here. Um not everything you uh my conditions always solve it with this. My wife jokes about this. Um uh, but there is this uh you have to be not thoughtful about is this a good use case? Not everything's a good use case. Uh and I think you mentioned it slightly. I think uh if the application has a bigger blast area, meaning more users uh internally, then you should you get obviously sort of you gotta think about uh if it's a good fit. And especially if you're touching external users, your external universe customers, then your level of scrutiny processing is it gets elevated, right? So again, not everything is I think it is, it will be, but I just you have to be careful within the organization, thinking that through. Well, I'm the biggest proponent of it when I work with customers or my clients, I I do say, hey, you know what, we probably shouldn't be thinking about it for now. You know, do it in a limited sense here, but not you know, not the entire workflow. Well, you bring up a good point because some what's the analogy? It's uh a hammer looking for a nail it that doesn't always apply. So thinking I am that hammer though. Understandably so unfortunately I like to be a good idea. This has been everything's a nail. Yeah, well, this has been so informative. Uh is there anything else you'd like to kind of share with our audience or any uh anything you'd like to touch back on that we've talked about up to now? No, I uh uh the one thing I want to reinforce is that um it's it's an exciting time. I would say, as we said before, experiment, use it. It's just a tool. The better you it's a tool you it's the tool is evolving, so it behooves you to try it frequently, see how it changes over time. It's just gonna make you I mean personal, as we talked about, and business. I'd think be a business life career, it's just gonna make you better overall. All right. Well, Arthur, it's been a pleasure having you. And um, just a couple of fun personal questions. I remember you're in the Boston area. That's correct. I am. I've been living here for almost 20, over 30 years now, yeah. Oh, so for a long time, are there other parts of the country that you've lived in or that you enjoy a lot? Yeah, well, okay, that's two different questions. Uh I was just really quick. I was born in born in uh I was born in Busan, Korea, lived most of my formative youth before college, Boston College, by the way, and New Jersey, northern New Jersey, near uh Hackensack. Those who know who know New Jersey, yeah. And then from there I went to Boston College. The typical story, met my wife there, moved up to she lives up here, okay. Up here, so I've been here ever since. In terms of places, I love uh I've been to Korea recently because I came to America quite young with my son last year, and I want to go back. Uh I love I just like going back to Asian countries. I find it fascinating. I love the food,
Favorites, Food, And Final Takeaways
Mike Chungwhich helps as well. Uh I was in the US. Uh, my favorite places are San Diego and Hawaii. Maybe because I live in Boston, it's cold here. Maybe that's why I like to as well. The winters uh make you hardier, that's for sure. Exactly, exactly. So, what about San Diego and Hawaii? Nice climates, interesting people, cuisines. I just I it maybe it was just you know sometimes the confluence of when I went, the age I went at, and what I was thinking of. I think Hawaii was just very laid back. The weather was fantastic. I just I mean, Hawaii's like every day is like 80, generally speaking, every day is like 85 and nice. So it's quite I mean it's expensive, but I didn't have to worry about cost. It's hard to believe though at the end of the vacation, you know. I w I went there on a honeymoon and uh I remember towards the end, I was about to take any job that anyone would give me. So yeah, you know, the funny you said that because uh we went uh we went on my uh our honeymoon was in Hawaii as well. My first trip to Hawaii with my family. I actually the sec the end of the first week or second week, I was actually looking in the newspaper for the wanted. I didn't I didn't fit find anything fitting my background at the time. Well, I'm sure Boston is glad to have you. And then one last fun question. Say you're having a dinner party, you're inviting some friends over. What kind of food would you serve? Well, I'm lucky in that my wife, she is an amazing cook. Um, so she I let her make the decision. I whatever she cooks is typically fantastic. I'm usually in charge of the drinks, so I usually bring a decent wine, some IPAs, which I'm a I'm a big Scotch drinker and whiskey drinker, so I typically have some interesting scotch and whiskies. Yeah. All right. Well, I will be the first to ring the doorbell when I'm in the Boston neighborhood. Always invite a mic. Well, Arthur, thank you for your your uh warm hospitality and thank you for joining us for this show. And to all of our listeners and viewers, thanks for joining this edition of indicators, and we hope you learned something and that you will be using AI platforms and just like me and everybody else continuing to improve and embrace the technology. So thank you very much, Arthur, and thank you to our audience. Thanks for tuning in to another episode of Auto Care on Air. Make sure to subscribe to our podcast so that you never miss an episode. Don't forget to leave us a rating and review. It helps others discover our show. AutoCare On Air is proud to be a production of the Auto Care Association, dedicated to advancing the autocare industry and supporting professionals like you. To learn more about the association and its initiatives, visit autocare.org.
Description
A dashboard used to be a project. Now it can be a conversation. Host Mike Chung sits down with Arthur Lee, founder of Fyne and a longtime data and business intelligence practitioner, to show what modern AI analytics looks like when large language models and AI agents move from hype to real workflows. Arthur shares how companies are using LLMs to boost productivity, automate repeatable reporting, and turn unstructured text like sales call transcripts into usable marketing insights and follow-up documents that no one has to write by hand.
Then they get concrete: Arthur walks through a demo mindset using public NHTSA recall data, explaining how an AI coding agent can pull large datasets, shape them into charts and summaries, and produce a clean, stakeholder-ready dashboard in under an hour. They also unpack the “progression path” from ad hoc ChatGPT questions on smaller files to standardized templates using projects or custom GPTs, complete with citations and exportable reports. If you’ve been wondering how to operationalize AI without becoming a programmer, this is the bridge.
They also get honest about the hard parts: governance, privacy, security, and the trust problem. AI is powerful, but it is not a mind reader, and it does not arrive with your domain knowledge. Their best advice is simple and practical: treat AI like the smartest intern you’ve ever hired, give it context, and always verify outputs before you scale.
If this sparked ideas for your team or your own life, subscribe, share the episode with a colleague, and leave a rating and review so more listeners can find the show.