Randy Johnston and Brian Tankersley welcome Stephen Edgington, Chief Product and Technology Officer at Dext, for a forward-looking conversation on AI, small language models, agents, and the future of accounting automation. The Accounting Tech Lab is an ongoing series that explores the intersection of public accounting and technology.
View the video below:
Or listen to the podcast using the below player:
Edgington shares how Dext has evolved from OCR-based document capture into a sophisticated AI platform processing over 353 million documents annually with 99% accuracy.
The discussion explores how Dext leverages specialized small language models rather than massive generalist AI models to ensure precision in document extraction and classification. Edgington introduces Dext Assist, an AI-driven evolution of supplier rules that learns from 12 months of bookkeeping edits to identify subconscious accounting patterns and automate complex client-specific logic.
The episode also dives deep into Model Context Protocol (MCP), AI agents, API strategy, and the emerging agentic world of connected systems. Edgington compares MCP to Bluetooth—powerful but requiring iteration—and emphasizes that the next evolution of accounting technology is about integration, iteration, and human oversight.
The central message: AI will not replace accountants—but accountants using AI will outperform those who do not.
==-==
Transcript (Note: There may be typos due to automated transcription errors.)
SPEAKERS:
Randy Johnston; Brian F. Tankersley, CPA.CITP, CGMA, and Stephen Edgington (Speaker 1)
Brian F. Tankersley, CPA.CITP, CGMA 00:00
Randy, welcome to the accounting Technology Lab. Brought to you by CPA practice advisor, with your host, Randy Johnston and Brian Tankersley.
Randy Johnston 00:09
Welcome to the accounting Technology Lab. I’m Randy Johnston with my co host, Brian Tankersley, and we’re very fortunate to have a guest today from Dex software, Steven Edgington. He’s joined DEXT about three years ago, and we’ve been around Dex to the product for the better part of 15 plus years. So have known the product for a long time. I use DEXT in my k2 Canada operation, and we have used that really, since it was available in Canada. So we have a fairly long history with the platform, but Stephen, not so long with you. Can you give a bit of an introduction and background to our listeners?
Speaker 1 00:47
Please? Sure, absolutely. So actually, I’ll give you a little bit more of a background, so it might help in explaining how I think about the product. So when I first started out, my kind of background is really computer science and software engineering, and I set up a business, and I founded that for 14 years, and that business was consulting with businesses all around the world, many in the US, working on manufacturing systems, yeah, going to companies and looking at how to transform them and really bring them to life, which gives you an incredible amount of kind of exposure to how companies work. And that business successfully sold to a company in the US called Epicor, who’s an ERP company. They’re quite large, and in effect, I then worked at Epicor for seven years as Chief Innovation Officer, helping lead that team and drive that so that gave me a lot of background into lots of different businesses and how they work. And then DEXT fairly recently. So you obviously worked with DEXT before I was there, but I’ve been there now for three years, and a lot has changed over that time. And really, coming into DEXT, I am the chief products and technical officer there. I’ve got about 200 people in my team and sit over product engineering support, AI and yeah, help shape the future, really, that’s what we do. And I’m really excited to sort of get your perspectives and also share where we’re going next in the world of accounting.
Randy Johnston 02:05
Well, that is super In fact, it’s nice to know about the Epicor manufacturing background too, because we had the good fortune of being epicor’s largest reseller in the United States back in the 80s. So do have some history with multiple companies, so I’m surprised we had not bumped into you while you were at Epicor. But you know, things happen for a reason, so the reason now must be that you’ve got control, and I guess I’ll say vision and insight as to where you’re trying to take this platform. I think you have a fairly substantial team that report to you at this point. Is that true?
Speaker 1 02:42
Yeah, so I said over about 200 people. And yeah, you can imagine, not only is AI transforming what we can do in the product, it’s also transforming how we build products, yeah, and it’s pretty amazing to be in business at this point in time, where you can leverage this technology to do just so much more. So super exciting. And, you know, there’s various different elements, you know, DEXT, we processed about 353 million documents last year. So that’s invoices and receipts, and in a single day, we’ll do over 1.5 million. And so, you know, there’s lots of companies that will claim, oh, yeah, we can do that. We’re using that AI to do this. But the big question is, how do you do it at scale? How do you make sure it’s right. How can you rely on it? And you know that requires you to have quite sophisticated sampling methods, you know, teams that check auditors, that check the auditors, and that’s exactly what we do at DEXT, and that’s why, you know, we’ve got a very good, I truly believe we’ve got a very good product in the market that’s loved by customers all around, and they depend on us, and we are there to support that process. You know, we small businesses, and I was one. I had a business of 14 years. I would have loved text if I could use that in my business. It just wasn’t around in the days we were using sage, yeah, and other tools, it’s very manual, and now being able to transform a business and help them, especially when there’s an accountant and bookkeeper involved in that relationship is just so powerful, because it lets the business owner focus on what they can do uniquely and in the world of AI now, you know, it kind of opens up another whole sort of level for people to work at, both in the accounting and bookkeeping world and in businesses. So it’s a super exciting time.
Randy Johnston 04:22
It is super exciting. In fact, our listeners know but you would not know that about me, as I actually have computer science background as well, and have worked on AI for a little over 50 years at this point, because I was writing ai 50 years ago in Lisp. So you know, have some unusual background that most accounting professionals don’t have now, on the other hand, on this, you know, I remember what a bugger it was for decks to work out the machine learning in the early days, because it was hard to get these bloody transactions classified correctly. But it sounds like you’ve got some new techniques. Technologies that may be a bit of a breakthrough for the platform. Can you give us some insight on that?
Speaker 1 05:07
Sure, absolutely. So, yeah. I mean, a lot of people think text technologies, OCR, yeah. They’re like, Oh yeah, it’s just OCR, Optical Character Recognition, which is still an element of, like, there’s a lot of intelligence that goes into OCR, but it’s, you know, it’s very much more complicated than that. You know to, you know, as you would know, to do real good machine learning and AI, you need data, and you need good quality data, and you need lots of it. And luckily, Dex has been around for a long time that we’ve got billions of documents, terabytes of data, that’s all well labeled. But it’s not just well labeled. It’s also been, you know, double checked by financially acute understanding professionals and then put into ledgers. Yeah, so you’ve got kind of, like, very high trust data. And you know what we do is, you know, the transformer architecture, which, again, many listeners probably won’t understand what that is, but that’s the kind of underpinning technology that is in chat GPT, that paper, was released by Google sometime in 2017 and that was the real sort of inflection point in the industry that really unlocks all the things that we’re doing today with AI. And in fact, Dex uses that similar technology, but not massive, huge models. We use what’s called Small language models that do this. So they’re very specialized, and they help us just focus on a particular element, and that’s what gives us the ability to get to that 99% accuracy, because it’s like a collection of small language models that helps us do our processing. And so that’s sort of that’s not the new thing we’ll be talking about, but that’s important to understand that it’s not just a case of taking a document and sending it to somebody else and getting back all the answers. Yeah, you’ve got to get the data out of the document that’s correct and use all of the dimensions for it. The currency might not be stated. You’ve got to take into account the context of one of the countries. You might be in Canada, but you’re in the US, yeah, because someone’s traveling and buying something, and the tax implications of that are different when it kind of comes you know, you might have a supplier, you know, is 711 but that’s not what you want to put into the accounting software, because you don’t want it to be populated with all these different suppliers. You want it to know that, no, we put that in as gas, or we put that in, there’s like insurance. So on the surface, it sort of seems like, oh, well, why is it so hard? In fact, the hardest field we have to predict is the supplier, because it’s the one that’s actually it’s not what’s on the document. It’s what you want it to be accounted for. And also, there’s nuances, like, if you see someone who’s bought a laptop, well, maybe their business is to sell laptops. It’s very different to if they are buying a laptop for a staff member, how they’ll categorize that. So there’s, like, all these nuances that you need to kind of take into account. But again, that’s all what we do all the time. And it’s in ill you know, it’s in every country. It’s in multiple languages. It’s in all these different currencies. It’s one page. It’s a picture of a CRT monitor for the daily sales, it’s a coffee stained receipt crumpled up. So we see everything, as you can imagine. And then what’s really good about AI is you can train on this, you can refine on this. It loves data, and these models now are getting really pretty powerful. So obviously, there’s a huge amount of investment going into these data centers, you know, 50 billion, well, 500 billion, you know, like Stargate open AI x ai. And these kind of factories, these token factories, that are just filled with high powered GPUs and compute are crunching data to produce tokens, yeah, and train new models that can help predict the next value, and these models now are getting to the point that we can do even more with them. And so as an example, as you’ll know, when you use text, a document comes in via various different methods. We might collect it from email. We might collect it from a supplier website. We might get it from somebody taking a picture on their mobile app, and we’ll process that, and then you’ll look at it most of the time. You might automatically publish it to your accounting software, but you may make a change. You might change the description. You might change a tracking category or a dimension in QuickBooks. You might decide to flag something or move something. You might decide that delivery notes or supply notes you want to put them into a different area and all those changes we’ve captured for the last 14 odd years.
Randy Johnston 09:28
I’m sorry to step over there, but you know, I’ve been reflecting on a couple of comments that I wanted to call out for our listeners. In prior episodes, we’ve discussed these language models, from the narrow to small to medium to large, and how they are fit for purpose. So I think it’s really fascinating that you’ve been using a small language model, or, sorry, a narrow language model, to do this matching. And there’s some real benefits to doing it in that style. Number one. Then the second thing, your comment. About billions of transactions that have been accurately coded by accountants into these systems. So in effect, you’ve got some of the best accounting transactional data set, possibly the best in the world. I don’t know that in fact, but billions of transactions properly classified over a decade plus of time is pretty darn stunning in today’s world. So that really sets you in position, as I would say, it Steven for a brilliant future where we can begin applying even more AI concepts to the data. So I think that’s where you’re headed. Sorry, that’s where I stepped over the top you but I think that’s where you’re headed.
Speaker 1 10:47
Yeah, let me explain that maybe a little bit more. So we’ve got so you know, the large labs open AI anthropic, like, if you wanted to build a large language model today, you need billions in kind of investment. In fact, Andre caparthi, who was one of the founders at OpenAI, who now actually has put together, there’s a great blog post. He’s an excellent sort of tutor on these things, but he recreated chat GPT two, yeah, and he kind of calculated that he compressed like seven years ago, they released chat GPT two that no one really hears about, but that was the early model that really gave the science of intelligence, and he recreated that now, taking all the learnings, and you had to spend like, 50 million before, and now you can do it with $50 right in seven years. That’s the compression of, like, time and technology and everything. And there’s a great stat as well. Like, you know, when you look at the speed of the rise of the Internet and then the speed of the rise of chat, GPT, it was kind of like, you know, one one AI year is seven internet years. Yeah, you know, to get to 800 users. So this technology we’re in, you know, a lot of people call the the exponential era. This technology is just getting better. But people, you know, when we talk about billions and these things, like, they lose kind of like, how do you bring this back to reality, right? And the reality is, there’s these large language there’s these large models that are really generalist. You know, if you’re on chat, GPT or anthropic clawed on your phone, and you ask it a question, it seems to be an expert in everything. It probably mostly is, but it will get things wrong, yeah, and that’s why there’s a warning at the bottom these things. These things will get things wrong. Can it extract data from the invoice? Sometimes, yeah, sometimes it gets it right, sometimes it gets it wrong. But these models aren’t specialists, and they don’t have all the checks and balances of a specialist, because they are very large generalists, yeah, and that makes them very powerful. But if you can distill that down to just like, focus on one thing, just like nature does, yeah, like, you start to specialize, right? And you don’t end up there’s not just one type of bird, there’s not one type of animal. There’s lots of specialists that you know, it’s all trade offs. And these language models, small and large have trade offs. When they’re small, they can be fast, they don’t drift. They can sort of get things. But if you can combine the two, you get something pretty amazing. And that’s what we’re kind of introducing. Is a thing called Dexter cyst, which you probably won’t have seen yet, but it’s the next evolution. And so DEXT originally had a very rudimental thing called supplier rules, which is, if it’s this supplier, set this category on the transaction. Works well for lots of things. Yeah, if it’s 711 and we know that they just always fill up with gas and, you know, it kind of gets coded to that transaction. Or if it’s like from this supply, we know it’s always the same. Works great. Completely falls apart when it’s Amazon, yeah, and people are buying more from Amazon, because sometimes it’s this, sometimes it’s that sure you can do line item extraction and get the breakdown of the items. But how do you what you know? How what we’re going to categorize things as? Likewise, some of your clients, when you’re working with them, they want to have, like, a bit more interesting treatment of documents. So we’ve got a good example. Is we have a one of our accountants has a client, and they provide higher cars, higher vehicles to Formula One teams that travel all around the world at different events. And you know, that sounds like great, so obviously, we get lots of invoices from them for this higher company. But what they want is they want them categorized to a dimension which represents the country that race is in, so that they can do analytics and see how much did we spend, and, you know, kind of like all of that for each of these different races, you know, for marketing analysis, etc. And so there’s somebody in that bookkeeping firm that has to look and go, Oh, Valencia, that’s Spain. Oh, Abu Dhabi. That’s this one, yeah. And so you can imagine, indexed, our AI gets all the information correct. But we didn’t know what they were trying to do with this, because sometimes the category doesn’t exist yet. Yeah, that country doesn’t yet exist in the in QuickBooks. So you’ve kind of got to do two things. You’ve got to go add the country. You’ve got to be able to connect the dots between the city and the fact you want to be the country. Look at this other. Context. And so what we’ve introduced with Dexter cyst is pretty magical. It looks at all the changes that you do over the last, sort of like 12 months of an account, every single field change you’ve made on a document. And then it basically leverages large language models to, in effect, try and establish what’s the subconscious rules that you’re applying for that client, and then put that into natural language guidance, okay? And then every time a document gets processed after it’s been through all the small language models and been validated and all the checks and it’s looked at everything else, and we know that it’s correct now, applied guidance and the guidance can do everything from flag a transaction that looks like it’s personal versus business, or it could be like, set the country dimension to be the value, you know, the country that this document’s in. Yeah, no, it’s like, you’d explain that to a bookkeeper, and they’d get it try and write that. You know, as a computer scientist, it’s impossible. How do you know what could like? It’s really hard, but these models can really do that well, especially when they’re shown lots of examples and they’ve got access to tools. So Dex desist is basically our evolution of supply rules. It takes it from like, from like a if this, then that to just describe what you want to do in natural language, and it can do things like move anything that looks like it’s not an invoice to the archive. Yeah. Flag items that we want to check are correct. Yeah. So anything you can do in the interface, the guidance can do. And what’s really powerful about our guidance is the guidance can be shared across your clients in a practice. So not only do we so we’re not saying, hey, come into this area and type your subconscious thoughts of how you do the bookkeeping, because no one knows what they do. You know good bookkeepers, good accountants. It’s just instinctive. You’ve done it so many times, you just know how to code things right. And so that’s what’s really powerful, is these models are really powerful at finding patterns and extracting kind of reason. And so we’re basically using the large language models, the expensive thinking models, to go how do we save you more time? How do we look at these edits? And so Dex desist. That’s our new capability. It’s a kind of not needed for all clients. It’s just needed for some of those clients, especially the ones that you spend the most time on. And yeah, we’d love your thoughts on how you think that might work in your your experience.
Randy Johnston 17:31
Yeah, understood. And you know, this multi layer model is a big deal. It’s been around indexed for some time, but not interactive, like you’re describing. And in fact, we considered last year 2025 the year of agents. And we considered this year the year of model, context protocol, MCPS, which we have covered in another accounting Technology Lab. And you know, in effect, these models that can interface to AI, with natural language, applying logic and so forth. We think that’s another big breakthrough. And I was reflecting on how this classification could work, because in our case, we do cross the border, we have some transactions in the US and some in Canada. And you know, if we’re getting a maple donut, let’s say in Canada. You know, that’s a pretty common purchase, many times at a convenience store, like a 711 so you know, how do you classify a donut purchase in the morning? So Well, Brian, I’m going to actually call you in at this point because you have been pretty quiet this time around, and I know that you’ve also used decks through the years. But there might be some other lines of reasoning and questions that you might have for Stephen and that you think would be helpful for our listeners, sure.
Brian F. Tankersley, CPA.CITP, CGMA 19:02
So Stephen, one of the challenges that I’ve seen with adoption to all kinds of tools that ingest information is that, you know, the dex does a great job of taking the data out of the unstructured formats of the images of the documents and converting it into digital data and putting the right context on it. However, getting the data from there into some of the more arcane systems that are out there, you know, you dealt with Epicor, I don’t know if Dex has an interface into epic or not, but I mean, there are a lot of there are a lot of systems out there, like, you know, acclivity, acclivity, or, you know, some of the others, some of the others in the space where there isn’t a direct interface. Do you see any innovation happening with the AI code generation to help people create bespoke import tools to get that information into those systems? Right?
Speaker 1 19:59
Chris. Question, yes. So first of all, Dex doesn’t work with Epicor yet. Obviously, I understand all the APIs of Epicor and how it could do that. We also, I mean, we our main focus initially, and our main 36 plus integrations are to the smaller online cloud based, you know, kind of general ledgers that are typically API first, and have got clean ways to work. Where things get more complicated with larger systems is where there’s customization comes in. So you post an invoice and you need certain custom fields now populated or, you know, and there’s ways to do that, but Dexter’s kind of like originally been really focused on that sort of super simple self serve. There’s no, you know, you need weeks to implement this. You can start a trial online. You can be up and running the same day, yeah, you can process transactions and just really start to see a return on investment, whereas with an ERP system, yeah, you’re kind of rolling that out over nine months now, to your point on AI tools and generation. So there’s a few dimensions we’re looking at next MCP. I’ll talk about that in a bit. But MCP is kind of one unlock. We weren’t or aren’t really thinking about MCP from a how do we leverage this to get data into other systems? But your point on code generation and in effect, and also, I think computer automation is super interesting, because there are now ways you could absolutely without knowing any code, process items into DEXT in an unintegrated way. Have Dex do all the work and then have an AI read the browser of DEXT and then automate manual software on your desktop and type that in into a format so
Brian F. Tankersley, CPA.CITP, CGMA 21:46
that’s achievable. You’re talking about robotic process automation there, right?
Speaker 1 21:50
Well, rpa, yeah, so blue prism and these old RPA systems are kind of like no longer really that relevant. You’ve now got the likes of like Claude co work as an example that will control browsers, your desktop software. It can do this now, yeah, there’s a big divide, or there’s a big sort of gap, between what these models can actually do and what a lot of the population understand they can do. So you could, you can do that without any knowledge. Yeah. I mean, it would require you to spend time and refine it. It will get it wrong, but you could build something that was reasonable, and then you can take that automation. It won’t be fast, it will be quite slow. Be using a computer like a human would, but anything that a person can do this day and age on a machine can pretty much be automated, especially if you can guarantee that the data is right, and you just doing an exercise of taking it from here, taking it to there, and having some kind of checks and balances in place. Now, for Dex, specifically, we’ve now got engineers, as you can imagine, using these tools too, we can point it to an application and ask an AI to generate us an interface and interaction to what we’re looking to do is just open up our APIs more. So we have an API already. We call it extraction as a service. So there are software companies that want to use the power of DEXT, like if core or others wanted to use the power of Dex and have that capability. We actually now sell extraction as a service, as an API offering. You don’t get the UI, you don’t get some of the bits, but you can get the kind of core AI capabilities. And we’ve already got our number of customers leveraging that. But the other API we’re looking to do is if it’s allowed people to integrate to us without us knowing, yeah, and that basically just gives you an API key, get a feed of the documents that have been processed, and that will really help. But what’s I’ll get on to the MCP part, because I think it’s interesting. It kind of relates, to some degree, to Dex assist. So, you know, we process documents in and DEX becomes your form for bookkeeping, your automation sort of platform in a firm, yeah, so your users can use it to save lots of time. Your clients, you’re getting the data in, you’ve got a that’s the foundation of of getting the data in, making sure it’s right, and unlocking the value we’ve also added in that time tracking now we’ll be releasing soon, and what that does is automatically track how long do your team members that do the bookkeeping in your firm spend doing the bookkeeping on the clients? Yeah. And what that lets you do is, if you can now go, Okay, we’ve got 600,000 small businesses that we help do bookkeeping on, we can now order them by Who are you spending the most time on? And then you can turn on Dex desist on those that need it, right? And that’s a real great you know, combined with our certification, you know, it’s sort of like a three pronged attack, yeah. Are your users certified? Do they understand all the tools? Indexed? Tick, yeah, which clients you spend the most time on? These enable dexes to perform that action right? So that’s the kind of like, that’s the first phase of optimization. MCP lets you then plug DEXT, and it’s like an AI connector. Plug text into. Your productivity tool. Maybe it’s chat, GPT in your firm. Maybe it’s Claude, maybe it’s Manus, maybe it’s any of the new ones that are just about to like pop up, because we see them every day. But the kind of MCP people like to think of it as USB. I know you probably covered this, but I don’t like to think of it as USB, because USB always works. When you plug a USB in it, your mouse, it works, keyboard, it works, right? MCP, though, to me, is a bit more like Bluetooth. Sometimes it works, but it works. Sometimes it doesn’t work, and you have to, like, try and pair it, yeah, like, it can be finicky. Sometimes it fails. And really, that’s how you should think about MCP. It’s not a problem in like, you know, how Dex is implemented, or how someone else is the problem is, when multiple things are connected, these models sometimes get confused. There’s only so many tools they can have, and so when you plug in more things, there’s, there are limits with USB as well. Yeah, I think it’s like 127 devices. Yeah, the 128 keyboard won’t work, yeah. But you know, again, these models, they are, in effect, really intelligent things that you can ask questions to, and then you can connect them to tools. And then they can see these tools and make requests. So just put into practical terms, you can have chat GPT in your firm, plug in a connector to next, and then you can do things like, Hey, I’ve just got a new client. Can you add them? And it will add it to text, because there’s a tool that says, Add client. And when it knows to call that, it knows I need the name, I need their email address, yeah, and maybe their phone number, it’ll ask you join me to send them an invite, yes. And so automatically, you’ve not had to use the interface. You’ve just added the client, set them up, they’ve got a message on their phone to download the app and get the documents in. You can also say, hey, how many clients have got items that we haven’t processed? Yeah. And so all the data that’s indexed now is available for these models to pull in, reason about and then come back to you. So then you can say, hey, which clients do we spend the most time on? Yeah, which ones aren’t set up properly? Who’ve got configuration that’s different? Yeah, where’s so it starts to unlock a huge potential. And then, when you’re using like index, we have more than just the bookkeeping. We also have a thing called Data health and insights. And what that does, that’s a module that reads data out of the QuickBooks and zeros, and it pulls it together. It looks for heuristics. Do we have transactions that are miscoded? Do we have categories that aren’t set properly? Do we have duplicate contacts because someone’s processed things incorrectly? So it’s it’s basically like a data that’s what it says on the tin, data, health and insights, right? Really powerful. The problem is, don’t know how to interpret it. But guess what? These models know how to interpret it. They just need the tools, right? But they don’t want to. You can’t connect the model to QuickBooks and Xero to get the data. Is it’s like not it needs to be processed first in some sensible way. You have tools over it and the foundation over it. And that’s where then, you know, you get to some interesting scenarios, because you get to the point where you say to these models, these assistants, who am I meeting with tomorrow? Oh, that client. Give me a summary of in effect everything I need to be thinking about whether the costs increase. Have they took on someone else? Yeah. And so previously, Dex only had the context of the costs, the sales, the bank transactions, but we also, you know, while I was there, added vault, if you’re familiar with vault, but vault is the general document store. It’s an intelligent document store for everything else. It’s the leases, the contracts, the employment contracts, the non disclosure agreements, everything. That’s not a cost item, right? But that is super important for advisory, right? Because you can’t give advice if you don’t have the full picture, right? The AI will tell you complete rubbish if it doesn’t have the full picture. But so would a person you know, if you go and ask them advice about tax planning or anything, they don’t know what you’re planning, right? So really, what we’re doing with Dex is building a foundation, a platform that kind of can bring all the context together intelligently for a client to enable you then to do the next level and MCP that all this, you know, AI connector, connecting your tools together with these generalist models like the clause and chat GPT to specialist systems. Yeah, they’re still going to always be specialist systems. That’s just the nature of how kind of evolution works, and these specialty systems, in effect, connected to the right productivity tools, unlock some really powerful new capabilities, and it will just save everyone so much time, and also allow the advice to be so much higher quality, more real time, but it still requires someone to be accountable, yeah, and someone to assure the stuff. So I don’t believe the profession is not going away. Yeah, there’s still going to be accounts and bookkeepers in the future, but the bar is being raised, and the ratio of clients to bookkeepers that will go up.
Brian F. Tankersley, CPA.CITP, CGMA 29:53
So, you know, it’s interesting you compare MCPS to, you know, in the data interfaces between a. Them, between the AI engines and the ERP and the systems of record, as you know, not as reliable as USB. Do you think that there will be a you know, what kind of improvement do you expect on those over the next three, 510, 15 years?
Speaker 1 30:19
MCP, typically, is a lot of people that are introducing MCPS, yeah.
Brian F. Tankersley, CPA.CITP, CGMA 30:25
I mean, in the last two weeks, we’ve seen MCPS. We’ve seen we’ve seen demos of MCPS from Acumatica, NetSuite, many other folks. So it seems like this is the year of the MCP. So again, talk to me a little bit about that.
Speaker 1 30:40
Well, the problem is, I mean, and again, this, I haven’t looked at others, but just to sort of like, as a broad explanation of it, we started with API’s, yeah, so an API is a application programming interface that allows one system to talk to another system. It’s a clear, sort of like contract that says, create the customer, create the invoice, read the values, and you have lots of different types of APIs. Some are well designed. Some are not well designed, yeah. Some have got 1000s of endpoints, and the problem is the more complicated the product. You need to know which order to call those API’s, yeah? Now MCPS are kind of like not designed for a software application to talk to another software application. An MCP is meant to be the ala carte menu that an AI can use. Yeah, it’s kind of like a human kind of interface, right? And the problem is, sometimes people generate an MCP over their existing API, and they don’t rethink what are the questions that people are going to be asking and, in effect, establishing what should the shape of the tools be? So I think what’s going to happen is loads of people are going to be announcing MCPS, launching MCPS, but just because they’ve got an MCP doesn’t mean it’s any good. Yeah, it just means that you could plug it in, right? Yeah, but it could be like, 1000 entries, too many. These models, like, you know, if you go over 50, get confused. Perfect example, the in the world of development, like, when we’re writing code now, it’s completely changed, and you can plug MCPS into open your browser and do things, yeah, and you can run MCPS to read other things. But these models get confused. They don’t know which one to read the logs from, and then they can, kind of like, waste a lot of tokens and thinking they get, you know. So what you actually want is, you really want to know, what are the questions I want to ask, and can I get that from the systems? Yeah, from these MCP. And so the question is really kind of like, not, do you have MCP or not, but what questions, what kind of solve with this? You know, what is it trying to solve? Is it just trying to create a surface area that lets you do anything you want, because you can kind of do that through the browser as well, or is it going to be more efficient? So don’t think of the kind of the fragility of MCP, like the Bluetooth analogy. It’s not because of the technology. It’s because of these models. Don’t know which one to call, yeah. And I think what’s going to happen is you kind of need to work with these. You need a number of iterations, just like anything, yeah. You need to be able to, like, get the first version out, get feedback, refine it, get the next one. So you want to look at technology companies that you’re working with and make sure that they’re continuing to innovate. As soon as you stop innovating, if they’re not making releases, they’re dead. Yeah, like, if you’re not seeing changes, it’s over the world. Now is like, there’s no excuse not to be making things better, and you’ve got to get that feedback loop. And in fact, you know, a kind of interesting so on the MCP, it’s like, what do we think? You know, accountants and bookkeepers need to be effective, you know? Can we sit with them, watch what they do and work out? Because when you ask people, What do you what questions you wish you could ask, chat GPT or thought about next? People often like, I don’t know or know what to ask. But when you show them how you could add clients like this, you can find deadlines. You can chase paperwork. You know, what if you could basically have a scheduled task that before the deadlines required for you doing the preparation? Yeah, it basically looks at the previous year of transactions. How many did they put through? Looks who it was, yeah, looks in the bank to see which ones are missing, associated evidence, or, you know, and chases the person that provided it last time. That’s a lot of work to do manually. You say, I can just do that automatically. And what does it need? Well, it needs a tool to look at transactions that haven’t got kind of like good reconciled transactions. It needs a tool to be able to search for previous transactions. Yeah, so, and when you stitch these together, you could create a tool that allows it to provide you with a summary of what needs to be chased. And then you’ve got another tool, which you send an email, and then you have a scheduled task that says, right now for all my clients that I need to chase, especially the offenders. Yeah. Can we automate that chasing process? Or, better yet, can we help them by setting up AI to go through their mail to find the invoices to forward it on? Yeah. So there’s just lots of opportunities to, sort of like, unlock this. So. I wouldn’t say, think whenever you see MCP, think about it with a grain of salt. It’s that’s just the first bar. Yeah. The question is, can you get the results out of it when you ask the questions, can these models understand the tools to use in the right order? Yeah, and there’s going to be a lot of learning in that, because these models are getting better at, you know, there’s actually, actual tests these models do now. There’s, like, evals that we call them, yeah. So, like, you know, there’s a eval for lots of different things, but there’s one for tool calling, yeah. And so there’s a lot of learning and adaptation that’s going into these models to work out, how do you best call them? And so, you know, don’t be put off if you connect your general tools to these MCPS, and they don’t work as you need to. You need to get the reps in to know how to ask the questions, just like any technology, to kind of get the most out of it. And it’s not just a give it a question, get back an answer. Typically, you’re kind of building skills. And there’s a big thing in the development world now where, like a skill is kind of like a text file. It’s not complicated. It’s a text file that describes how to use the tools. And so you can use skills that help give these models, sort of like the matrix. When you plug in a learn a new language or something, it helps them understand how to use that MCP. And, you know, Claude have actually just released what they call plugins that are collection of mcps and skills, yeah, that can sort of help the models understand, how to use spreadsheets, how to use, kind of like, you know, different things. So there’s going to be, you know, I think it is the world you said before. It is not necessarily just the world of MCP. It’s the world of all of this. So it’s the agentic world, it’s the agents and swarms world. It’s the MCP world. Ultimately, you know, this next year is going to be about the model. You know, we don’t need any more powerful models. We just need the existing models of their current capability being leveraged by more people and like connecting the dots between systems, and that’s interesting.
Brian F. Tankersley, CPA.CITP, CGMA 36:54
So what I hear you saying is that the real evolution that’s going to happen over the next few years is really through the iteration, through the doing the reps of things. So effectively, what you’re saying is that if, in the analogy of AI being a gym, the LLM is the gym equipment, the weights, the, you know, the exercise bike and so forth, and then the challenge for us to get fitter is for us to go do the work and try it, and to make mistakes and to figure out how exactly to structure things, to to make it work better for us in the future, exactly.
Speaker 1 37:29
And if you followed along on this kind of malt book and open claw, all this kind of like stuff that’s been going on, you know, with these like agents and assistants, you know, there will be a point where, well, how we interact with computers is changing, yeah, you can just talk to it now. It can perform actions for you. Yeah, it can see what you can see as world models. Can generate games without anyone programming anything. So we are at a point where lots of stuff is changing, yeah, you know, can you send out to your clients a kind of, you know, a voice memo or a video explaining their financials in terms of, understand, absolutely you can translate things. Now, it’s really powerful. And, yeah, it is a little bit like a gym. You have to, kind of, like, get the reps in. I think the challenge is, it’s like a gym that changes every month, yeah? So, like, one prime, the exercise bikes over there, and then it’s, like, moved over here, yeah. So you just gotta have a mindset, which is, like, you know, that beginner mindset, to make sure that you kind of, like, really leaning into this. It is a bubble, and it’s not a bubble. Yeah, same respect, like, loads of money’s going into it, but the technology is real. The change will be, will transform every industry. Jobs will be displaced, but there’ll still be jobs, yeah, is my view. And there’ll be more software. You know, there’s lots of scaremongering. It’s not going to replace everything. It still needs supervision, but it will be able to allow you to do more. And if you’re not using it, competing against someone that is using it, they’ll just be able to do more than you. And that’s going to be difficult, because then prices will come down, but there’ll be more businesses that need support. So I think it’s kind of like this real kind of like everything’s been thrown up in the air. And the best way of learning in these technologies is Be curious and lean in, I think, you know, with regards to like, DEXT, you know, as an example, and what we’re doing, index, assist, and how we’re thinking about this. You know, all of our teams, all of my teams. I mean, Dex is about 400 people, all of our business is reimagining, how do we do things. I mean, development is changing how we do things in support. You know, we have about 50% case deflection now in support. That was zero a year ago. So now, you know of the tickets that come in when we’re helping people, if they can get an answer immediately, that’s great, but not if the answer is rubbish, right? The answer has to be good, and they’ve got to be happy with it, right? And that requires so the team that’s doing that support now, they’re not being sacked, yeah, half the team. What they’re doing now is they’re providing better guidance to the AI, writing better articles, using AI to write better articles. And it becomes a now you go to like a really, you can get exactly the information you need. And then what we do is we move that to proactive support. So now you’ll find that if you trial decks. And maybe someone will reach out to you. Do you need any support? Yeah. Or if you’re stuck, or we see that you’re stuck, there’s a rage click, or there’s something that hasn’t worked, we’ll just pop into the chat. Do you need help? Yeah? And there’s a human there that can be escalated to, and we believe it’s that type of same thing. So you’ll have exactly the same thing in the accounting world with your clients, yeah. Maybe years ago it was kind of like, they drop off a bag of stuff, and then you complete the stuff, and you’re done. Then it’s kind of like quarterly, then it’s some of them might be a bit more real time. But what if you could say, Hey, have you noticed this trend? Or I’ve seen this happen with this other client? We can help you with that, and you can start to, you know, change the trajectory of their business, because you can spend more time and analyze it, yeah, and you’re in effect, pattern matching across all the clients that are solving this, you know, you’ve just got that different perspective. They’re in the weeds. Yeah, they’re solving it, but you can look at it from a different lens, and it’s that external view. So, you know, for us, Dex Assist is kind of like just the one next evolution. But it’s not, you know, there’s so much more to do. It’s like, how do you then apply this even further, you know, schedule jobs over all the work. How do you do more in the bank space? How do we do more involved around contracts? Could you, like, compare them? Can you do sort of, like, more work? And you know, a lot of the small businesses, they need a tool, you know, self you know, sole traders also need somewhere to keep things safe and secure. So, yeah, I think it’s a really exciting area. And what we’re looking to do is just leverage this technology to make everyone more successful. You know, our kind of vision is like, you know, making accounting effortless and making time for business, and when you take time and effort, yeah, well, let’s just reduce both of those, yeah, and use technology, but it doesn’t remove the human. We’re not going out saying this is a replacement to bookkeepers. It’s like, No, this is an assist. Yeah, you score more goals. Yeah, if you can, kind of like, just get those, get set up better, will make you be more successful. And I think that’s really how you should think about this kind of technology. It’s like, this is if you lean into it and you learn it, it’s the first technology we’ve ever had that can teach you how to use it by asking questions. Yeah, how do I you know? And it’s about so MCP, you’ve just heard that you want to know more. Ask it to explain to you. Someone said it works like USB. How’s the analogy? And it will explain it to you. You know, you can ask it to quiz you, right? Show me. Ask me multiple choice questions, and based on my response, you use that to work out if I know I’m if I understand it, and then ask me more tough, challenging questions or simpler, challenging questions. So this is like, it’s a kind of just a complete new world that is super exciting, and if you don’t use it, you’ll just look at the news and you’ll look at all the same things. And there’s many people trying to sell you things that don’t work, yeah, or that seem to work but don’t actually work like you’ve got to look for well, how many documents Did you transact? How do you validate this? Correct? Yeah, this is hard. You know, it’s hard to get one half million a day. It’s hard to do 350, 3 million a day a year. If it doesn’t work right, you’re not going to get good reviews across external websites if the product isn’t good. Yeah, you’re not going to be the biggest connected app in QuickBooks and Xero if your product isn’t good. So, you know, take with a pinch of salt. You know, just because it’s got AI in its name doesn’t mean it’s great, yeah, you know, test. You got to test. You got to check, but you’ve got to know what the bar is, so you need to have a good sense for it. I think you just have to be technical to some degree. But yeah, hopefully that gives you a
Randy Johnston 43:25
little bit of vertical it does. So we’ve had the pleasure of having Stephen Edgington as a guest today. He is product manager and visionary at DEXT. You know, DEXT has 36 plus integrations, I believe he said earlier, and that includes zero and sage and Zoho and Intuit QuickBooks and Bill and Zapier and reckon and Twinfield and many other products that we’ve covered in other episodes. So hopefully he’s given you a little bit of insight. And perhaps it’s time to make sure that you’ve got your personal trainer around to help you exercise at the gym a little bit more and get some of this right. So Steven, it’s a pleasure to work with you today, continued success. Thanks. Thank you for being here with us on the accounting Technology Lab. Good day.
Brian F. Tankersley, CPA.CITP, CGMA 44:16
Thank you for sharing your time with us. We’ll be back next Saturday with a new episode of the technology lab from CPA practice advisor. Have a great week.
= END =
Thanks for reading CPA Practice Advisor!
Subscribe Already registered? Log In
Need more information? Read the FAQs