Transcription:
Transcripts are generated using a combination of speech recognition software and human transcribers, and may contain errors. Please check the corresponding audio for the authoritative record.
Chris Gaetano (00:03):
Hi there. I’m Chris Gaetano, tech editor with Accounting Today and you’re listening to On The Air. The introduction of AI coding models have upended the once familiar dynamics of build versus buy for CPAs. What was once primarily a domain of large firms with teams of professional coders now accessible to even local practitioners with no engineering experience at all. Instead of spending hundreds of thousands per year on a vendor solution, people can just tell the AI what they need and in literal minutes get a custom application fitted to their exact needs in theory. Here to discuss some of the practical realities of Vibecoding and its uses in accounting firms is Ellen Joyn, founder and CEO of AI Focus Accounting Consultancy Edgefield Group, and before that, co-founder of Accounting Practice Management Solutions provider Aywin. It’s great to have you with us today, Ellen.
Ellen Choi (00:55):
Thanks, Chris. Really excited to be here.
Chris Gaetano (00:58):
All right. So yeah, first of all, for the benefit of our readers, what exactly is vibe coding? How do people do it? What do people use it for generally?
Ellen Choi (01:10):
Yeah, great question. So let me take each of those two words, vibe and coding. So coding everybody I think is familiar with. That’s the realm of what used to be for software engineers. Software engineers that can write these complex code like programming language code and develop web apps, mobile apps, software scripts. So that’s a coding piece of it. The vibe piece that now we have in this vibe coding, that’s what’s really transformational, which is that now this idea that you can just talk to AI and do the same work that only software engineers used to be able to do in writing code that is deployed and actually usable and productive for people. And the software that is building, so that is deterministic in the sense that the output is actually same as what we have seen for many, many years and decades. It’s like software engineering.
(02:10):
But what’s truly exciting about this vibe coding is that now people like you and me and other people who don’t know anything about Python or any scripts or any technical languages can just speak to ChatGPT, speak to Quad, speak to Copilot, and it translates essentially like a language and it generates this software code for them. In terms of how do people do it? You literally just talk to AI and ask it to build you something specific and then it generates a code automatically. And people use it for all sorts of many, many different use cases from what I have seen.
Chris Gaetano (02:48):
Okay. So turning to CPAs in particular now, what exactly are these firms vibe coding? Look, are we talking a single auditor making their own note taking app? Or are we looking at a whole firm that’s replacing its entire general ledger system with some sort of vibe coded alternative or I don’t know, something in between or maybe it’s all over the map. Look, you are the expert here. What have you been seeing?
Ellen Choi (03:16):
I would say that I have not seen yet to date somebody that had five coded on entire general leger system that they’re able to replace. So we’re not there yet today. We’re still in the whole crock, crawl, walk, run, sprint. We’re very much so in the crawl phase of people just getting comfortable with just idea of vibe coding and just generating something out into the world. So a lot of the examples that I have seen so far are much simpler, which is probably good for our profession at this point. It might be a partner even just writing better scripts that go into Power Automate. That’s actually a pretty common use case in the sense that they’re setting up internal apps that build spreadsheet automations. Really instead of having to use a OCR app, build vibe code, something internally that can take scans and put it into the exact format that they want for trial balance formation, data extraction, all of that.
(04:21):
And also just a lot of little simple broken fix it type of things that could be done with extremely simple apps. Seeing some examples of that I’ve also seen examples of people using this for internal productivity purposes, internal collaboration. So one top 50 firm, I’ve seen them, they were using Replit to build a prompt exchange sort of platform. So it’s like they vibe code it essentially. And that’s such a great use case because there’s no proprietary client data. It is actually vibe coding to foster more AI knowledge education. And it was a very professional looking ap and having a lot of engagement with gamification features, having fun with it, with emojis. So it’s a way actually for internally with accountants, I’m seeing teams of people use it for that type of lower stakes, but still interesting, helpful use cases and also very pointed solutions to automate things that just was difficult to automate because there wasn’t enough capacity on the IT team side to automate a hundred different things with Power Automate scripts.
Chris Gaetano (05:44):
So yeah, we’re not talking the firm’s full tax engine. We’re talking their full ERP system. We’re not talking about their CRM or anything like that. We’re talking more relatively single use, narrow use case kind of stuff, if I’m getting you right.
Ellen Choi (05:58):
Yes, it’s almost like the edges of those tools and glue all the things that they’re doing manually today to really use those systems together. A lot of vibe coding to help automate a lot of that. Just very tedious error pro manual work. And then as I said, some kind of fun creative work like the prompt exchange platform.
Chris Gaetano (06:23):
Okay. So why then are these firms vibe coding? What problems does vibe coding solve, AI coding solve that your traditional vendor offered solutions might not?
Ellen Choi (06:38):
Yeah, so actually I have seen this now. I think every firm has some pocket of practitioners that are just really frustrated at how existing vendors fall short of what they want. And actually it’s not necessarily 100% vendor’s fault because a vendor needs to build a system that scales and they can’t customize every single little variant detail to the practitioner’s liking. So first of all, there’s that frustration and desire of I really wish I had tools to fix this, but I don’t and I’m at someone else’s mercy. And then you also have the subset of these people who are just really what I call AI pilled. So it’s like the whole matrix red pill, blue pill. They just love AI, they love tinkering, they love exploring new things. So it’s not really like the firms are not saying, “Hey, let’s do this and let’s try by coding this.
(07:48):
” It’s actually people that are trying stuff on the edges and then they come and it’s like, “Oh look, look at what I built and this works within the Copilot system, whatever that may be sanctioned. Or if it’s a tool, a platform that’s outside, it’s completely information that’s not sensitive or whatever it may be. So it’s really this profile, the persona within firms that are starting to vibe code things within the firm. And many times that is either happening before or after the conversation’s happening at the firm-wide level. So it could be in other cases, but it does spur on more conversations. So how do we make this more available for other accountants and other practitioners?
Chris Gaetano (08:37):
Now, one thing that I didn’t necessarily hear, and honestly in my conversations also with practitioners about this, is that it’s not necessarily a cost to play. There can be some cost savings for this, but it seems to be less about saving money and more about getting just the kind of software that they want. I was wondering if this was something that you also had observed.
Ellen Choi (09:02):
100%. And I can tell you from actually my team’s personal experience five coding our product that we just launched for public, our AI learning hub, my conclusion in looking at vendors and five coding it ourselves, I don’t actually think it was any cheaper.That’s my conclusion. So when practitioners are looking at this and they’re looking at, can I solve for this with the exact set of different vendors or upgraded features versus five coding? I don’t necessarily know if the cost is really what they’re thinking about, nor is that an actual outcome of the trade-off of going down the five-coded route. What I will say is that the output that you get just so much more exact and specific. And so many times I have heard from practitioners and accounting firms, the thing that was vibe coded has 70% of the features that we could get from a vendor.
(10:06):
So it’s actually less powerful. It’s actually less functional, but we didn’t really need those other 30% anyway. And this exact set of 70 is a specific set that we built and customized specifically. So we actually prefer this even though we may have, it’s not necessarily cost-effective and you don’t get as much functionality for it, but we like it better. So I’m starting to see firms really make that calculus and come to the conclusion. And the other thing is from a cost perspective, a lot of these little problems that firms are solving, and I think many times a vibe coding product, it does start from this productivity.That’s a unit of measure. So it’s like time savings. And when you translate time into billable hours, then they can justify the cost in their head. Vendors aren’t necessarily building these very specific little niche solutions because the market’s not there for it.
(11:08):
It’s like a market of one, a market of five. But if I’m in the market, if I’m in the market of five, I care really greatly about having this little vibe coded product or this web app built for me. So if I can do it myself and the cost is net neutral, then it’s actually a win for me because I don’t have a good alternative anyway. So I’m definitely seeing some of that play out as well.
Chris Gaetano (11:34):
So it occurs to me as I’m speaking, so the offering that you had just talked about, the AI learning platform, it being vibe coded, I’m wondering just for illustration for our readers, what was the process like for that? What did vibe coding an actual platform involve?
Ellen Choi (11:53):
Yes, absolutely. Great question. So before I started vibe coding this with my team, I asked lots of questions to Claude about exactly how I can do best vibe coded, make a best vibe product platform. One of the things that came up was it’s really important to get the database structures and the backend schemas right. That’s the hardest thing to fix if it’s not done well in the beginning. So from our vibe coded product perspective, we spent a lot of time talking to AI through CloudCode about the backend schema, exactly what we need, writing really robust product requirement documents. So really putting on our product major hats on and testing it. And we also brought on multiple engineer advisors because we’re like, okay, we don’t need a full-time engineer to do this manually, but we do want somebody that’s a human that’s actually looking at this five-quoted product on a periodic basis.
(12:59):
So then in the beginning we spent a lot of time focused on the architecture and how the actual data schemas are set up and the backend, all of that. And then we were talking to AI and we used Lovable. I think Lovable and Replit are the leaders right now, really great tools. Many, many, many firms are using them right now. And we just literally were talking into Lovable, into prompting, to what the interface and all of that should be, all the wiring, all of that should be. So spending a lot of time upfront gave us the ability to actually generate things that were relatively quick by the time we got to the point of actually generating the code. So for us, that was a process. And this is a part where I come back to in articles you were rightly pointing out or you and I, I think we were talking about this idea of anyone can just generate code and have a web app that works, but can you scale that?
(14:01):
Does it fit within the governance? All of those points. So having a lot more upfront planning, and this is where I think my and my team’s training as product managers came in really handy, becoming excellent product managers and just being able to dump our thoughts and generate code like that as opposed to just having something that’s really vague generating code that’s also not constructed well.That’s something that’s going to make a difference in whether the web app or the software is built for the long haul or if it’s something that has to be thrown out and it’s more of a prototype.
Chris Gaetano (14:43):
So yeah, let’s actually, this is a good way to go into the next sort of question here. Let’s flip this around. We were just talking about the advantages here, but I mean if it were perfect, I’m sure everyone would be doing it. But what problems does vibe coding introduce that a traditional vendor offer solution does not?
Ellen Choi (15:01):
Yeah, so much. And by the way, vendors are vibe coding on their side as well. It’s just that they have a lot more controls and systems and human in the loop reviews that are actual expert engineers that are constantly correcting the vibe coded product, which is a thing that’s not happening with a practitioner because they’re not classically trained engineers, so they wouldn’t know how to do that. So on that side, because vendors have all these guardrails built in and they are actually much more hands-on in terms of what a good software product looks like that’s scalable and deployable, it’s already built like that. It’s out of the box when a firm adopts it as opposed to something that’s just vibe coded by a practitioner or accountant. They don’t know what they don’t know in terms of building highly secure, scalable, all of that. They can probably ask Claude and put some stuff in and I’m sure the folks that are V-coding on the accounting firm side are doing so, but they don’t have the lived experience.
(16:13):
So the product is just not going to be as good the moment you try to adapt it for a different service line or the moment you try to scale it up or have it work beyond that one person who Vivequered the product and it works exactly for that one workflow, but the moment you change a little bit, then the whole thing breaks. So it’s a lot more brittle. Whatever is V-coded, unless it’s gone through this very rigorous thinking and review and almost like I come back to, we talk a lot about human in the loop. The human in the loop is very applicable here too in terms of engineer figure actually human that is overseeing and ensuring that there’s good judgment baked into how things are written and structured because that doesn’t exist for five coded products. It’s just a lot more brittle. It will break more quickly.
(17:06):
It’s not scalable. It’s not necessarily deployable within the security governance framework of the firm as opposed to just making it work for one person. So there’s all these hoops that five coded products have to go through before becoming a true enterprise grade secure platform product like many of the vendors.
Chris Gaetano (17:29):
So from what you’ve observed then in all of these firms that are doing this, what are the highs and lows? So when a firm is doing this particularly well, what generally are they doing? How are they approaching it? And then conversely, when a firm is not doing this so well, what generally are you seeing?
Ellen Choi (17:49):
Yeah, so I think the firms that are doing this well are all, they have it within a system that has already been established, communicated with employees. So there’s already guardrails or there’s specific sanctioned, maybe Replet has been set up or there’s specific licensees that work better for the firm environment. So it’s more transparent and it’s empowering practitioners who want to go down that route and actually create these types of products. So that’s the best case scenario. And in these cases, usually there is a IT person or AI team, somebody from the AI team that is involved from the beginning that’s really fostering the ideation and the creativity and the vibe coding on the practitioner’s part, but supporting them from day one to ensure really almost that human in the loop support that I’m talking about, providing that from the beginning. So by the time that the vibe coded app is done, it is in a shape that is much easier for it to propagate and scale and be used by other people.
(18:59):
So that’s a done well case. And then the opposite end of that is definitely folks that are doing it on the side and the firm may not have a overarching firm-wide policy on this, but it’s not clear. It’s in the gray area, it’s unclear. People are doing it on the side and the firm leadership doesn’t know about it. The IT team doesn’t necessarily know about it. That’s obviously the worst case scenario because you always want more transparent open conversations around this and you want to know what they’re building. And accountants by nature understand they like to be letter of the law, which is great. So no one’s usually doing anything that’s flagrantly in violation at the same time because there’s no guidance. It’s not clear what is good, what is bad, what is ugly, what is not allowed, so all of that. So in those cases, it’s actually like there might be some people that are vibe coding within a firm, but it’s going to be a lot less people because there is no environment that’s fostering or nurturing this type of work and this type of AI enablement and empowerment.
Chris Gaetano (20:13):
So I guess it’s not just a matter of human in the loop, it’s which human?
Ellen Choi (20:18):
That’s right. Do you
Chris Gaetano (20:18):
Have the right human?
Ellen Choi (20:19):
Yes, absolutely. Which human matters a ton? Yes.
Chris Gaetano (20:24):
We’ll be right back with more accounting tech talk after this. And we’re back with Ellen Choi here to talk about vibe coding CPAs. So the vendors that I talked to about this, they seemed generally pretty sanguine about this. Sure they’ve lost some business, but seems they’ve also gained some from firms that tried and failed the DIY aproach and now need a professional like them. Overall, what I’ve observed is that there seems to be a sense that vibe coding apps, vibe coded apps can never really replace dedicated enterprise grade solutions, at least not right now. Are vendors right to be unconcerned? Overall, how do you view this affecting the vendor landscape?
Ellen Choi (21:16):
Yes, I think that’s absolutely true today, but let’s just look at how fast AI is moving. It’s just at this unprecedented rate in terms of adoption and the actual AI capabilities. So today it is hard to find code on app that is going to be fully secure, fully productionized, deployable, things like that, but it’s just getting a little bit easier every single day. So I come back to that. And the other thing that I think a lot about is not all vendors are equally exposed. So it really depends on the vendor category and what is their core value offering? For example, anyone that’s hanging their shingles on OCRs and document classifications, that’s just commodity through AI, through vibe coding, not even vibe coding, literally just through really intelligent prompting on Copilot, Claude, building agents. So it’s just looking at this universe of what do practitioners have directly available in their hands as alternative to vendors, which includes five coding and other types of tooling as well.
(22:38):
So there’s certain types of features which it’s going to be very difficult for vendors to sell that as a separate standalone offering. Things that are very regulated deep workflow specific, those things are safer because it’s just harder for off-the-shelf generalized solution or a practitioner to vibe code their way into something that’s going to meet regulatory requirements. So that’s just more complex. So I think those types of tooling are safer for now. And in general, from a vendor perspective, if they’re not going for becoming the system of record and the source of truth and really on the data side of things, I find that hard to believe that the AI capabilities won’t get so good in the future that practitioners can build little scripts or code or things like that that could work with those very secure systems of records. So that’s just another though on the vendor.
(23:44):
So on that, it depends on who you ask is what I would say. And I think their reaction in terms of actually vibe coding is good for us because they try vibe coding and then it fails and then they come back to us. So yes, I think that’s true for today, but then I wonder there’s probably a different business model here in terms of there’s a whole class of probably vibe coded app support, technology support, technology migration. I think that’s probably a rising market. So that’s a new actually class of competitors that vendors haven’t really encountered that they are now probably going to see more and more of. And that actually empowers accounting firms and practitioners to not just buy vendors, but be able to vibe code things in a way that’s secure, scalable and be able to maintain them over time with that technology expertise at their hands.
Chris Gaetano (24:47):
Okay. So going into the CPA realm again, we’ve been saying in the future for now, things like that. Where do you see all this going in the future? As the models improve, are we going to see more firms picking up vibe coding, more non-engineers developing apps? Or do you see this as maybe a short trend that’ll peter out once the drawbacks become more aparent? Could be a reassertion of expertise. Maybe the models will be only handled by software professionals. Maybe it’ll be more of a tool for them, something else entirely. I have no idea. That’s why I’ve got you on the podcast.
Ellen Choi (25:23):
Thank you. Okay, so this is my take. If you told me a year ago that I was going to be teaching accountants about markdown files and about GitHub code, I would’ve told you you’re crazy. Guess what? The SharePoint agents in the Microsoft ecosystem, now they have these skills repositories and those skills repositories files are in markdown formats. Md. So now in terms of teaching them that through the Copilot ecosystem, through the cloud ecosystem, we are actually teaching accountants how to be engineers even though they don’t know it. And there is this insatiable demand actually for that. And they don’t necessarily think of it as engineering. It’s more thinking of it as how do I best use AI and how to best use AI use these. MD files? So what I’m trying to say is this market’s moving really fast. And so in 12 months, am I going to be teaching engineers how to write Python scripts?
(26:38):
Maybe. And if you asked me that six months ago, I would’ve said, wow, that seems really far-fetched as a mainstream model. But I do really think going back to your question, all of us in society, including accountants, are becoming a little bit more of an engineer even though we might not think of it in those terms, but everyone’s becoming more technical. This is not a short-term trend. This is a foundational structural AI supposed to get better every day kind of trend. And even if AI masks some of those specific engineering complexities, there’s a need for practitioners to actually understand the fundamentals and the foundations because that’s going to help them understand how the technology also plays out into the future. So I do actually think that the future is a lot more engineers, not engineers, accountants that are wearing more engineering hats and vibe coding things on their own.
(27:40):
And it’s going to become a lot less scary, a lot more accessible, a lot more secure. But I think the really big win here is that in the future, when accountants and practitioners are able to put on this hat of builder, they actually get to be hands-on builders, it makes them better product managers. So it’s almost like we’re teaching them how to be mediocre engineers because I’m teaching them how to be excellent product managers in which they can hand off exact precise requirements to engineers that actually have much more expertise and experience that can build fully deployable, scalable software solutions for them. So I absolutely think that’s also a rising trend. That’s not going to go away. And that’s super exciting for me because it’s really democratizing software development for accountants in a way that was never available before. It’s enabling them to solve their own problems that may have been too small or the market was too small for any technology vendor to come in.
(28:49):
So I do think that is absolutely the rising future.
Chris Gaetano (28:53):
Okay. So we’ve probably got a number of CPA firm leaders who are listening to this very podcast and they’re going, wow, vibe coding, hey? Now how do I get into this myself? What advice would you have for accounting leaders who want to vibe code, who want to do it safely and effectively and in a way that benefits their firms?
Ellen Choi (29:15):
Yeah. So the first thing I would say is crawl, walk, run. Let’s look at where you are on the spectrum. And this year I have gotten so many more accountants asking me, I want to learn how to build an AI agent. And when I ask them, what does that mean to you? Everyone has a different vague answer, which tells me and I mentioned this in the context of similar to vibe coding. It sounds really cool and it sounds amazing and it is, but you need to have these layers of knowledge and training and baseline understanding of how AI works well… Role in order to take the most advantage of five coding. So in terms of firm leaders and accountants that are listening to this, I’ll just say make sure you are systematically moving your entire firm through AI literacy and education that comes with, do you even have a really strong written AI governance policy?
(30:23):
You can’t layer on a five coding project on top of not having one. Do you have no draw? Yeah, I hope not. Exactly. But you know what? That’s the thing, you never know. Sometimes I think firms take different steps and it’s jumping all over the place. So let’s just do this well systematically in the order that it should happen. So really think about that. Don’t turn people loose without specific guardrails and empowering them to do exactly what they can do within a very defined fence. There’s also a lot that people can do before they go to vibe coding in terms of these general AI tools, whether that is Copilot. Now with Copilot co-work, obviously cloud is a lot very buzzy right now, but you can actually employees can max a lot of the underlying baseline LLMs before even having to vibe code something. So I think just really having those layers of education and enablement and teaching people what AI is capable of at different levels that’s going to help somebody become a better vibe coder.
(31:30):
And I think in terms of foster innovation and creativity from vibe coding as well. So when you have all of that, can you run maybe mini hackathons or can you run little innovation days and challenges and how do you actually make this fun for people? So vibe coding is not something that sounds scary for some of the folks and that can really be translated across the entire organization. How do you have those AI pilled partners and accountants teach the rest in terms of how vibe coding works? So I think that’s one other thought. And yeah, going back to the right human in the loop, which is such a great way to look at it. How do you get into this well and safely? Just make sure you have the right team, the right humans in the loop that’s going to help make this a real investment that generates returns as opposed to just one person just builds it as a prototype.
(32:28):
How can you actually have a system or the right humans reviewing and building this out in a way that you could scale across the firm and just get real benefits across a much larger group? So really starting to think about that from a process and structure and systems perspective.
Chris Gaetano (32:46):
All right. And so with that, yeah, that is all the questions that I came pre-prepared with, but is there anything else you would like to go over about anything at all?
Ellen Choi (32:56):
No, I think this is a rising trend and I think in general there’s so many exciting potential for AI, vibe coding, building and customizing own solutions. We’re still in our infancy and I’m just super excited to see how all of this plays out for us.
Chris Gaetano (33:14):
All right. So that is all the time that we have for today, but I want to thank our guest Ellen for her time as well as you, the audience who make it all possible. This episode of On the Air was produced by Accounting Today with audio production by Admin Khan. Rate or review us on your favorite podcast platform and see the rest of our content on accountingtoday.com. Once more with feeling, I’m Chris Gaitano, technology editor with Accounting Today, and I will see you next time.
Credit: Source link
