Randy Johnston and Brian Tankersley welcome Kashif Ali, founder of TaxGPT, for a discussion on why general-purpose AI tools such as ChatGPT, Claude, and Gemini are not sufficient for professional tax research and advisory work. Kashif shares his unconventional journey from journalism to software development and entrepreneurship, ultimately leading to the creation of TaxGPT after experiencing firsthand the difficulty of finding reliable tax information.
The Accounting Tech Lab is an ongoing series that explores the intersection of public accounting and technology.
The conversation explores the evolution of AI in tax research, beginning with source-cited answers and progressing toward autonomous agent-based workflows. Kashif explains how TaxGPT differs from consumer AI tools by focusing exclusively on tax and accounting use cases, implementing hallucination controls, maintaining vetted tax knowledge bases, and emphasizing security and professional trust.
The discussion also covers agent orchestration, AI operating systems, workflow automation, and the future of accounting firms. Kashif argues that AI should eliminate repetitive compliance work while elevating the value of professional judgment. The panel examines the growing productivity gap between professionals who effectively leverage AI and those who do not. Looking ahead, Kashif predicts firms will increasingly deploy specialized AI agents, reduce reliance on outsourcing, shift toward advisory services, and potentially move away from billable hours toward outcome-based pricing.
Transcript (Note: Some errors may appear due to automated transcription.)
SPEAKERS
Brian F. Tankersley, CPA.CITP, CGMA, Randy Johnston, Kashif Ali
Brian F. Tankersley, CPA.CITP, CGMA 00:00
Welcome to the Accounting Technology Lab, brought to you by CPA Practice Advisor, with your host Randy Johnston and Brian Tankersley.
Randy Johnston 00:10
Welcome to the Accounting Technology Lab. I’m Randy Johnston, with co-host Brian Tankersley, and we are so lucky today to have a guest, the founder of Tax GPT cash, Ali, and you know, cash. Appreciate you taking time with us today. Would you like to give our listeners a little bit of your background, please?
Kashif Ali 00:30
Yes, sure. Thank you so much, Renny and Brian, for having me. I really appreciate that. A little bit background about me, I pre.. I studied accounting in college, but I graduated with a journalism degree, and I worked as a journalist for six years. Then I did another pivot, a career pivot, and learned to program. And then I ended up working for Adobe for three years. I started two other companies that didn’t go anywhere, and when I was trying to shut those companies down. I had a lot of tax. I was looking up a lot of tax information. I was not able to find correct information, so in a very short.. that’s how I ended up starting Tax CPD. Well,
Randy Johnston 01:15
you know, that’s kind of an interesting background. I knew about your Adobe background, but I didn’t know about the shutdown story, so I appreciate knowing that. Now, friends that are used to being with us, we wanted to talk with Cash for a bit, because the application of AI in tax research is pretty broad, and as you know, the Bigs have this with Thomson Reuters Checkpoint Co Council and Answer Connect from Walters Kluwer and the BNA Bloomberg, all are AI powered and authoritative, but I believe that products and platforms like Tax GPT or Blue Jay or Accordance or more, and there are about eight of those products that we’re tracking right now, are just easier to use for much of your team now. That said, we really wanted to just get a little bit more insight on the application of AI. Now, we’ve talked in other podcast episodes in other labs about the way this AI is affecting the accounting profession in such a big way. We have the general tools, including the announcements of this week with CCH announcing their new relationship with Chat GPT, and you know we’ll continue to follow all those different things, but you have the large language models, the productivity parts, you have the AI built into platforms that augment the platforms, I think. Cash, that’s where I put Tax CPT, and then you have the agents and the MCPS, so kind of three different levels of strategies that you need to be able to address on all fronts. And we’ve encouraged you to set up the proper policies and to get the right governance in play, and so forth. So, cash, you’re in the thick of all this bloody stuff, and we’re, we’re thinking we’re headed into a token economy, in effect, and you know, so just give us a little bit of your insights on the application of AI and tax research, or AI in general, across the profession, please.
Kashif Ali 03:18
Yeah, so when we started, and it was very early on, the old way of looking up the information was sold, and we all used to do that. You go on Google, you do a keyword search, you read 50 different articles, summarize it, synthesize it, form an opinion, write an email, tell back to your client, like very manual process, hours upon hours used to go in there. So, and talking about my own frustration, that was my own frustration about looking up the information, and like I need, I can read the laws, but I’m not a professional. But how can I make a judgment call on this? Right, I didn’t have enough money to hire a tax lawyer or an accountant, so I was like, let me try to make a tool, and, and that’s how Tax CPD was born. Very early on, we went viral, and we had 1000s of people start using, and it was more consumer lens that we were building the product, and eventually, right after the tax season, we noticed that people still coming, still using Tax CPD, and turns out those were accountants, lawyers, enrolled agents, professionals in the industry advisors. So we did close to 300 discovery calls, and this is I’m talking about early 2023 to mid December 23 so the whole year with 300 back and forth calls with customers to understand, really understand, truly understand their pain points, workflows. I went into a lot of firms in their meeting rooms, I saw them work. And how they find information, so the early, the first one to building trust, the easiest thing we were, the my apologies, the first thing that we did, we started giving the sources, because at that time, I mean, ready, you have seen the conversation, how much it has grown in last three years. At that time, 2023 people were like, there is no way that AI can do what I do, and I don’t even trust AI. So, you know,
Randy Johnston 05:36
you’ll appreciate that that time the hallucinations were so bad. Oh,
Kashif Ali 05:40
yeah, yeah, our friendly face
Randy Johnston 05:41
later was declared dead in a meeting that we were in AI, and do remember our conversation, because we’ve known each other a few years at this point, that you’d done all these discovery calls to try to get to this conclusion on the platform, but this trust level, just to go back, people didn’t trust, they didn’t think that AI would get there that fast to be trustworthy, if you will, and you know, for Brian and I, that is one of the key things on AI, is how do we build trustworthy models and get accurate device, so this citing of sources that you were doing that was really critical in my mind.
Kashif Ali 06:25
Yeah, yeah, I mean, I
Brian F. Tankersley, CPA.CITP, CGMA 06:27
mean, it’s a, as a, as a, as a partner, you don’t trust your, your associate, your two year associates without somebody going through it in detail. So why would you trust AI at, you know, more than you trust the human being that’s been through, that’s been through, you know, master’s degree program, and you know, pass the exam, and all that.
Kashif Ali 06:46
Yeah, and 100% agree, like neither should people trust, trust but verify, like checking, right. So we started giving sources. There is a lot of cool tech that we developed under the hood. We made the tax code consumable to AI because tax code is written for humans to read, it’s start from somewhere, right? And without going into much technical detail, we made it consumable to AI and make it make sense. And then we created hallucination control mechanism, so it does not make up stuff. More than a million tax sources, documents, 1000s of trusted websites, where we look up the information on a regular basis in real time, and vetted by humans, CPAs, tax lawyers that are part of our team, that’s how we create the first version of the product, where we were giving people sources and walking them through in into the productivity, that was early 2024 Since then, we’ve been building on top of this foundation, and now conversation has changed, going looking for information to actually AI doing the work in a workflow, so we can chat more about it, but if you want to focus, yeah,
Randy Johnston 08:16
but you know, as it turns out, you know, part of the reason I obviously we’ve recorded a podcast on Tax GPT in the past here for the lab, but I believe it was March of this year that you released your autonomous tax workflow agent, and you know our listeners have had us or heard us talk about agents and why they’re important in this differential of, you know, are the agents separate and developed with agents and model context protocol, or are they part of the product? In your case, you’ve got this agent that I think you’ve appended, you helped me get the right words, but you’ve wrapped your product with this, so we don’t
Kashif Ali 09:00
have an agent, we have agents. So, what we created is, we created a model orchestration layer, because we have the best tax knowledge available, and as compared to, you know, we don’t have to talk about, as compared to horizontal model, like why their information is faulty, everyone knows that. Why people should not upload their client information into Chat CPT or Claude, because they are not secure. So all of that, with all of that information, as we are developing this product, we created this memory layer right after our research agent, where users can put all the firms are creating all of their client information, so imagine we create the best intelligence brain, this is our tax research product, and then we created the memory layer, that is a client intelligence product, where people can dump all of the information, now people were asking for, like, can it do this, can it do this, can it do this, what. People were asking was actually like to actually do the work, so we started with the review process, automating one review in part of the preparation. Then we launched the preparation, and what we created is the agent orchestration layer, where we have created 30 plus agents where you can prepare, onboard, create work papers, create an organizer binder, do a month hand bookkeeping, book closing, you can create an R and D study. We with one of our partner from that, we brought down their 1r and D study time from 60 hours to less than 30 minutes. We people can do advisory projections and all of that. So we created this new agent architecture where people tell agent what to do and agree, and they an agent and go does that and human are in the loop making judgment call, so that’s our orchestrate the agent orchestration layer, we call it tax cpt co work. Okay,
Randy Johnston 11:11
I appreciate the clarification on that, because I remember when Agent Andrew was initially released, because I talked to Andrew about it, you know, as it turns out, but you know this, I’m going to call it end to end. We’ll
Kashif Ali 11:27
just
Randy Johnston 11:28
pick on the onboarding as an example. Onboarding being a very common problem that firms are trying to solve, but we try to get people to think about the end to end process, from PBC gathering and the engagement letters all the way to the delivery process and talk about that in the context of portals because there’s so much interest in 1040 work paper prep products, the likes of Black or Tax Autopilot or Filed or Magnetic or Solomon, and then the special DK one products, the additives, the abacus, and the like that are doing that type of work. So, do you see yourself in a where you are today and where you’re going as continuing to build out your agent library and supporting that end-to-end workflow? And tell me just how to think about that, please.
Kashif Ali 12:21
So, the best way to place Tax CPT is thinking that we’re creating the super app for accounting, tax, and advisory firms. So, you have named a lot of different people are doing a lot of different things, they all have their swim lane, someone is in research, someone is in practice management, someone is in prep, someone is in onboarding, someone is in right. What I’m saying is, we’re creating the super app that you can do everything in tax CPD, and that’s the goal, because the number one pain point when I was doing all of this discovery, and people were walking me through the workflows. I’m like, why do you have 14 tabs open to find a W-2? First of all, why do you have so many different tools, right? And maybe this industry needed an outsider like me to look at that and build something better. Maybe we don’t need a 19 step elaborate workflow to collect a last year 1040 Maybe we can do better, maybe we can do easier. Right, there are so many products that are so hard to set up that actually setting them up become a job in its own self, right. So our goal is, and we’re not asking people to, by the way, get rid of all of the tools that they are using right now. For we’re building is we’re building agents that can go and do things on your behalf. So imagine you very similarly, like prompting was such a foreign word three years ago, right? Today, agent is that word, so you ask, you tell your agent, you get, you said go in my onboarding tool, xyz, whatever that tool is, take that information, extract that information here now. Go to my preparation tool, right? Do the preparation right now. Take that information and go to Agent Andrew to do the review. Then you do a review, human in the loop, and now go deliver it to in the client portal, and you know the delivery, all of that community agent can do all of that, it can concatenate different systems, right, so that’s how we see ourselves an AI operating system, not just a. Single tool, or two tools, or collection of two, three tools that are in three different lanes, because the power of everything combining together is best intelligence brain, tax intelligence brain, and all the information, the context, and the memory, and agents doing the work, so we believe that we will be able to save people so much time and make them way more productive, and that’s the goal of Tax CPT building this AI operating system, so people can truly see the ROI of AI
Randy Johnston 15:35
well. That’s a beautiful explanation. Thank you. And I, as you’re saying that, you know the good news, bad news, and we try not to put things that I’d call timely in a lot of our podcasts, unless it’s really breaking news. But earlier this week I was teaching for a conference, and the very last session I taught was AI in tax, and the very last call I made before having you join us today was to a CPA firm in Brooklyn, tax team, right? And you know, I have to apologize to both of those groups, saying, you know what I said on the call, and what I said publicly in the session was, I am not aware of anybody that’s built a family of agents for tax that are effective yet now for those of you attending today, that’s how you learn new stuff, but it was part of the reason when cash and his team reached out, we said yeah, it would be great to get an update, because I was aware of, you know, Agent Andrew, but I didn’t realize this, we’ll call it Agent Operating Platform, which, you know is a nifty way to position it, and reality is it’s not really even though it’s a text CPT product, it sounds like it’s really of the agentic MCP layer, as the way I’m thinking about it, and I also think your comment earlier about, you know, prompts are one thing, and we’ve taught a lot of people to prompt AI engines, but that the agents are now your new prompting tool, and of course, exactly during this timeframe that we’re speaking, in the prior 30 days, of course, we had Microsoft released Agent 365 on May one, and earlier this week also released their Scout for building agents, so you know we’ve been watching for agent environments, and over the next few weeks, for me personally, I’ll be teaching how to build these agent environments, but you know, maybe for some of you listening today, build versus buy, maybe you want to buy rather than build, because a lot of firms have been down this path of trying to build agents, and you can build pretty simple agents well, but these ones that have text knowledge are are different, and you were correct,
Kashif Ali 17:57
and complaints,
Randy Johnston 17:58
the general, yeah, and complex, and these general AI tools just don’t do a real good job at all of this. We may think they do, but you also, I think, cash correctly called out the security risks, because I have watched instructors at conferences say, ‘Well, let’s just upload this tax return into ChatGPT. Like, no, that is one of the dumbest, simplest things you can do, and if you’re a listener and you just got insulted, I think I’m okay with that, because you, we shouldn’t be doing that type of work so
Brian F. Tankersley, CPA.CITP, CGMA 18:35
well, and that’s that’s no different from when people were emailing tax returns 10 years ago, you know, before they really, or 20 years ago, before they really adopted portals. It’s, you know, they.. I think sometimes people don’t know what they don’t know, but I mean, I will say that my experience, anyway, with, with, with, with AI, and especially with generative AI, is that the thinner the information is, the more likely it is to hallucinate, because these, these agents almost want to please you, because they, and so the problem, of course, of creating your own agent here is that you don’t want, you don’t want the agent to tell you what you want to hear, you know, you have, you know, you have short, short timer staff people that for that to tell you what you want to hear, you know, the eight, you want the agent to tell you the good, the good and ugly truth, so that you can deal with it, and that’s the, that’s, you know, that beyond just the complexity of getting all the right stuff into the training, it’s also trying to keep the mental health of the agent okay, so that it doesn’t hallucinate, and just dream up whole new, whole new things that don’t exist.
Kashif Ali 19:46
I would, I would add two things here very quickly. The general purpose AI, Open AI Cloud, and all of Gemini, those agents and those tools are created, they are like social media. They want your attention. They want to keep you engaged. They are not made for account rents and tax work and advisory work. They will, but Brian correctly pointed out, in order to please you, they can make up stuff. We do side by side comparison, and it’s very easy to gaslight them to say, like, this tax law exists, is like, oh yeah, I’m sorry, it does exist, but it didn’t, so I don’t know if you guys have tried it out in tax CPD, will you get a different experience, right, like a professional tool should be if someone tried to gaslight Tax CPD, it’s like, no, no, this tack, this section does not exist, and this is the interpretation of this section that exists. One thing, another thing is we don’t want your attention, we want to help you get your work done effective, effectively. If you ask Tax GPD, what is the weather is like outside, it’s gonna say that’s not a question that I’m designed to answer, right? So we, it’s a work tool, it’s a professional work tool, and this is how we like to keep it. So that’s one thing you know, that a difference between a professional or something, general purpose. The second thing is, as people are learning through this information, it’s as important is that how to you, we are at that level of AI cycle that it’s never been more important to learn to wield AI, and one very interesting thing that I see is that AI matches the capability of who is using it. We see this in our tool, and we see it overall in our company. If you are a senior person and you know what you’re doing, you get way more done as compared to someone who’s not right. So that’s an interesting anecdote. I don’t have very big learning here or anything, but just wanted to share that, that how a lot of certain people are can be 10x 50x 100x more productive, and I’m talking about engineers here that are truly willing. Yeah, there is a difference between people who really go all in and people who don’t, and that, that, that I see that in engineering, I see that in a lot of different professions, and we see that in the tool, and that gap is widening. So, my parting thought, and I don’t want to say this conclusion of what I wanted to say here is, if you haven’t ever tried AI or ever worked in agents and MCP, this is the best time to jump back in, because this gap, it gap is widening, and people should be on really learning about this stuff.
Randy Johnston 23:06
Yeah, in fact, I, as you were just saying, parting thought, I’m thinking, yeah, that is a super parting thought, because here we’ve been talking about my, and I hate to use popular words, but the sycophants of AI, you know, trying to please us, if you will, as opposed to I’ve got real work to get done, and I was thinking about the gap that you just identified, because I’ve been watching the AI gap widen, and you do have people that are consuming millions and 10s of millions and hundreds of millions of tokens in a day or a week, and they are getting way more done because they know what they’re doing, and so we’ve got casual users. It’s almost like professional drivers versus casual drivers. I notice so many people think they know how to drive, and I’m looking at them saying, oh, mg, where did you learn to drive, right? And I’m not saying I’m a great driver. Brian’s written with me, so he knows that I’m not a great driver, but I’m safer than most because I pay attention when I’m driving, and paying attention when using AI may be part of the formula here. So, Brian, questions, parting thoughts from your side.
Brian F. Tankersley, CPA.CITP, CGMA 24:21
Yeah, so, so, so, I guess I would just, you know, you know, cash, about three years ago we were talking about prompting, now we’re talking about agents and MCPS, and those kinds of things. What do you think the future looks like? What do you think the future of work looks like with AI, and what are some of the things agents are going to handle on an automated or agentic basis in the near future that people might not have expected them to be able to handle?
Kashif Ali 24:55
It’s very hard to predict the future. Sure.
Brian F. Tankersley, CPA.CITP, CGMA 24:59
Yeah, and we understand that, that you, you know, about, you know, our good friend Dr. Bob Spencer often said that 20 seconds in the future is about as far as you can be accurate, so we understand this. Okay,
Kashif Ali 25:13
exactly. I, my guesstimate, and where things are going, you know, you can see the future. What is going to happen? And I remember being on a podcast three years ago, and I gave them this example. I told them that I was an average programmer, but where AI tools were at that point, I was, I became a 10x programmer, and 10x programmer is an example, like someone who’s so productive, like a sorcerer, but the 10x programmer who learned AI was 100x 1,000x and this is trend that continue, I draw a lot of my inspiration for any future of work. What is happening in programming and engineering today? And the coolest thing is that I get to see it in Silicon Valley, and in my team, and a lot of my friends that have companies, and so, so based on that, this is my thesis. The future of work is that all the manual and repetitive redundant work is gone. You don’t need 17 step workflow. You actually don’t need to build the workflow right. Agents does thing what you want them to do, so for example, you can have one agent that is preparing and one agent that is reviewing, and then you can concatenate agent, and they are doing 10 different things, so you can spin up 1050, hundreds of different agents, sub-agents that are doing that job, you can be sleeping, and you can be traveling, and agents are working, they’re doing the work, so that’s like the swarm of agents, that’s that, and then the last and very best part of all of this, all of these agents, and including, by the way, this is I’m talking about tax CPT agents that we are creating. All of these agents are recursive learners, so they improve and they learn from their mistake, right? So, what it means for tax and accounting world. So, all the redundant work is done. All of that is gone now. The judgment of the people are gonna matter so much. Your years of experience, right? Agent is doing something, and it can tell you that this is where I want your judgment. If you make this choice, here’s the risk for this client, and here’s the reward for this. Right, the risk factor is nine, but reward factor is this. But if you want to have the reward risk, and you can, by the way, you can do the settings also. So we believe this will enable a one person, $1 million practice, and we people will be able to get free up and do more than compliance work. People will be actually able to do advisory work. I know this is a big conversation in the industry. I also believe that billable hours, we know there is a huge conversation around that, probably going to turn into outcome-based pricing. There is a lot of people are already doing it, and yeah, and we all know that there is a shortage of professionals, not enough new talent is coming in, and offshoring is a big thing, but the quality concern is a big thing also, so I think very, I’m just barely scratching the surface, because we launched this form of agent three months ago, the biggest thing people are saying is bringing everything on shore, right, scaling the firm, and because the actual bottleneck always has been finding good people in order to scale, right? So these are the few trends that I’m for seeing that’s going to happen in the next three six months, a year, or ahead, until unless there is a artificial super intelligence comes in. The things are moving so fast that you know I can only predict what I know in the next three six months a year.
Randy Johnston 29:53
So I think what I heard there, you try and tell me all bets are off if super intelligence arrives. I think I got that,
Kashif Ali 29:59
but. And
Randy Johnston 30:00
as it turns out, just thinking ahead, and this is maybe a little collateral, I do believe the days of outsourcing, I do believe the days that have of dashboards and several other technologies we’ve used, and and I’m going to think more carefully about your workflow steps, there’s several things that we thought had to be a certain way that don’t have to be that way anymore, and that’s, that’s really, I think, the big things that I learned from you today. Cash, well, listeners, we are so pleased to have you along for another county technology lab podcast. We hope you’ll be with us again next week when we speak about technologies in the accounting profession again. Have a good day.
Kashif Ali 30:46
Thank you so much, Trinity. Thank you so much, Brian, for your time today.
Brian F. Tankersley, CPA.CITP, CGMA 30:50
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 =
Sign in to get access to this free resource, and all of our whitepapers and reports.
Download this content today!
Register Now Already registered? Click here to Log In