Randy Johnston and Brian Tankersley, CPA,CITP, explore the rapidly emerging Model Context Protocol (MCP) and why it may become one of the most important AI developments for accounting firms in 2026. MCP acts as a standardized connector that allows AI tools to securely communicate with accounting and business systems without building custom integrations for every application.
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:
==-==
Transcript (Note: There may be typos due to automated transcription errors.)
SPEAKERS:
Randy Johnston; Brian F. Tankersley, CPA.CITP, CGMA.
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. One.
Randy Johnston 00:11
Welcome to the accounting Technology Lab. I’m with my co host, Brian Tankersley in his home office in Knoxville, Tennessee, and something that we want to talk with you about today that is super important for this year is model context Protocol, or MCP. Now, we had this in various episodes in late 2025, and it is a significant trend. Now, an MCP is a standardized connector that lets AI tools securely talk to business systems, typically accounting software, without any custom integrations for each of the apps. So you can think of an MCP like a universal adapter. Instead of plugging in a different cable into every system, MCP lets you use one cable. Some people have been saying USBs is an example of this, but in a separate accounting Technology Lab, we were fortunate to have Stephen Edgington talk about DEXT and he actually thought it was more accurate to talk about Bluetooth, which is a little more janky.
Brian F. Tankersley, CPA.CITP, CGMA 01:18
Yeah, it’s less, much less reliable. You have to fidget with it sometimes. I mean, on my iPhone in particular, the Bluetooth is just is a challenge, yeah.
Randy Johnston 01:27
So the MCP matters to accounting users in several ways. First, consistency, it’s one standard way for an AI to interact with accounting data. This, really, in my mind, for most of you, is going to replace API access. Two, it’s security. The access is permission based and auditable. Three is control. Your firm can decide what the data the AI can see or act on of. And it really is a practical AI. It enables real tasks, drafting reconciliations, explaining variances, summarizing transactions without exporting data, spreadsheets. Now, MCP is not several things to it’s not a replacement for your accounting system. It’s not an AI that thinks on its own, and it’s not an uncontrolled data feed. So the MCP you can almost think, is kind of a magic layer to allow you to talk to not only one system, but multiple systems at the same time. And the day before we did the recording, we were working with a colleague on MCPS, and we had many different interfaces set up at the same time. One demonstration particular, we had four different systems, and we’re going back and forth between all four and all four systems were recognized by the MCP query. Now, does it always get it right? No, so maybe I think Brian, I can see you. I’m just going to step in on that.
Brian F. Tankersley, CPA.CITP, CGMA 03:06
I was just going to say that it’s, it’s very much like in the, you know, back in the good old days before everything was as automated as it is today, when you needed to do a big analysis project, for example, like you’re doing M A or something like that, you had to go talk to five or six different people and interrogate them to get the information you needed to do that work. And this is an MCP is very much like this machine to machine conversation that you used to have in person to get the information you needed, whether it was I need you to write a query, get this out of this database, and then put this into a text file, and then you’re going to need to transform it into this format. And then we need to get this in a database. We need to get this additional table in this database and this other information so that we can compare these two data sets together and figure out what the overlaps are and what the other you know, find the exceptions as we think about this MCP tool really is designed to is very much like it’s gonna it’s asking the questions that you need to ask of the databases and of the data sources to get it now, it may ask them in the wrong order. It may not get the data structure exactly right. It’s going to make mistakes. So it is going to be kind of clunky, but it is a vast improvement over having to do all the stuff manually yourself, and it said it will actually solve a lot of prompt news many problems, let’s say 60% of problems in the background automatically.
Randy Johnston 04:27
Yeah, so I’ll probably ask you to also listen into the next discussion with Steve, because Steven was very factual about things. But I thought you were going to have a cow, Brian when he was talking about going to the gym and exercising, but you were just like, function at the bit, saying, Yeah, I’m thinking, No, I don’t want to
Brian F. Tankersley, CPA.CITP, CGMA 04:47
go to the gym well. But the thing about it is, there is a good analogy in that, you know, if you’re going to master a musical instrument, if you’re going to if you’re going to get better at any kind of anything that requires skill, like my son’s learned how to do machining. And welding. And, you know, I’ve worked with wood for decades. And dad, you know, and again, I’ve learned, I used to teach people how to weld. And so it’s a as we’re thinking about this thing, you know, I know that you have to work at something to get better at it. Okay, I’ve never felt it was important enough for me to spend an hour every day doing a repetitive walking motion on a treadmill to make my life more complete. But that doesn’t mean that I don’t understand that you do have to work at things to get better
Randy Johnston 05:29
at it. Well, you know, it’s the old 10,000 hour rule, yeah. And one of the things that in our work time here over the last few days, frankly, you know, we ran into the new c2 term, which basically says that c2 means that you speak like a native speaker in the language, and that takes about 1000 hours of training. And you know, there’s that old 10,000 hour rule that was out there about becoming a master. Well, we’re not asking you to become a master, but we know you’re going to have to do a little bit of work in the gym, a little bit of physical training on these images.
Brian F. Tankersley, CPA.CITP, CGMA 06:06
You’re going to have to try and fail at these things, okay? And as you fail, you will learn things that will make you successful.
Randy Johnston 06:14
But we’re going to try to get you through some of the fundamentals today. And for setting up a general MCP flow. There’s really five steps that applies pretty much to all platforms. First, you have to choose some sort of an MCP enabled AI tool. Now you can select an AI platform that supports MCP connections, or you can use vendor provided MCP connectors. And we were looking at libraries in the last 24 hours of various MCP pre written interfaces, which were fascinating, and various platforms that can be used. Then number two, you have to define the scope and permissions. So you have to decide, is it read only, or is it read write? And you know, if you think about general ledger and AR and AP and payroll, for most of you, you’re probably going to be far more interested in read only. And through the years, as we’ve taught ODBC, for example. You know, we talked about read only, type of authentication, and very few of the interfaces we actually allowed to write back budgeting in particular, we talked about that for years and years. Number three, you’re going to have a way to authenticate securely. So whether you’re using OAuth or a take token based access, you need some sort of an authentication. You shouldn’t be using passwords to make these connections, because those would be imminently hackable. Number four is, you want to test your setup with non production data. You want to run basic queries like, summarize last month expenses before you turn it out in the wild for broader use with your team. And number five, you want to set your government governance, you want to document and you need to update your policies. And most of you listening have a written information security plan to comply with the 5293 5447 and frankly, this is going to take an update, as I’d see it, I would expect it to be mandated this within this year, but we’ll see how that comes out. But also other things, like your AI usage policies, your vendor access logs, and other items have to be done as well, so that you can select MCPS for lots of different platforms. So before we decided to record this session for you, I did the research on MCPS for common platforms. So we looked at an MCP for Xero and an MCP for Zoho books and the MCP for QuickBooks Online. And all of those were available, including MCPS for many other systems, including productivity systems like Outlook as an example. So what I’d like you to begin to picture here is MCP is kind of this swiss army knife of connecting everything to everything. It overlays a system so we can ask things about it with AI. Now, when it comes to these governance recommendations, it’s pretty clear in my mind, starting with Read Only is a wise thing to do here. These systems are good enough that they can write data back. And the last thing we did before we went to dinner yesterday was write something back. And it was pretty stunning how easy it was to read in the original and write the new, updated things back that had been created by the AI.
Brian F. Tankersley, CPA.CITP, CGMA 09:54
But imagine how much of a disaster it could be if it didn’t write accurately. And you. Actually wrote a few 1000 records back, and you then had to manually go fix them. I mean it, that’s the thing about writing back to data is that you do it at your peril. You know, unless you’re positive that it’s going to work right, and it’s been adequately tested, you just don’t
Randy Johnston 10:14
do it. Yeah, I think another major thing, and many firms are dealing with this. Brian, in fact, I answered three questions in the last 24 hours on this topic, alone, client consent. What do we do with client centric data? And we’ve been on a bit of a tier in the last 12 months on many of these AI systems that are taking great liberties with client data. So you’re probably for MCPS actually going to have to update engagement letters and other AI disclosures. You know, one of the fascinating turns this year that you and I talked about in the last few weeks was the AICPA engage conference, where, you know, we’re both scheduled to speak, that they updated their engagement letter, where it has a clause that none of the content that’s presented can be developed on anything with AI. And I read that very carefully, because it’s like, you know, AI. Images seem pretty benign.
Brian F. Tankersley, CPA.CITP, CGMA 11:19
Spell check is AI, grammar check is AI, but the
Randy Johnston 11:23
document says no AI usage. So you and I are actually going to be kind of scoping things out and saying that looks like AI. So you need to watch your audit trails as well, so ensure that your MCP activity is logged and reviewable. You’re going to have to do even more vendor due diligence than before. We have talked about that in prior accounting Technology Labs, but you’ve got to confirm data retention and the training use and the sub processors, all topics that we’ve talked about in prior labs, also make sure that you got your policy right. So we’re super excited about MCPS, why it makes AI practical and more flexible, and it gives you capabilities that you’ve just never had before. So now I guess we could also go down into deeper into tools. But Brian, what other things do you think our users need to know about MCPS? I’m just going to predict MCPS are in your future.
Brian F. Tankersley, CPA.CITP, CGMA 12:30
I think they’re definitely in your future, because that’s how you’re going to get the data out of your systems of record and into the AI engine so that it can be used for to come up with conclusions, for advisory and to identify anomalies and to find fraud and do all kinds of other things like that. The challenge, of course, is going to be again, just like everything, with AI, making it reliable and working with it. And by the way, it’s not you’re not just going to have one MCP hooked up. You may have, like we did yesterday, four MCPS hooked up to four different systems that are all passing data back and forth, almost like a weird card game, you know, at the where these this data is just flowing all over the place in response to all these questions to solve this bigger problem. Yeah, so
Randy Johnston 13:14
at the risk of interrupting you, Brian, that’s one of the questions I wanted to ask Stephen when we were interviewing him. Explain how you do an MCP between zero and DEXT and Bill and, you know, because I was going to throw four or five different systems, many of which they interface with, but it’s like if you had an MCP for each of those systems, and then could make a query, explain how this receipt?
Brian F. Tankersley, CPA.CITP, CGMA 13:38
Yes, this here’s this receipt? Is this receipt a duplicate of anything in the ERP system? Is it a duplicate of anything that is indexed that hasn’t been posted yet? Has it been turned in on an expense report in Expensify or Zoho expense or some other application like that? You know, there are things like that in fraud detection and in everything else where we want it to be recorded once and exactly once, and so those kinds of things are going to require more sophisticated MCPS, and that’s one of the reasons why we think you ought to be experimenting with some of these things and trying to get your hands a little bit dirty, because you’re going to learn things working with them, and it’s going to take a while for you to become an expert and For you to be able to create information that where you’re satisfied and you should be satisfied at that point, you know, you should have a level of expertise where before you become satisfied. That basically says that this is reliable information and it is accurate. So you’re going to have to try and fail and, you know, check things 15 different ways, and then you’re going to be able to trust it after that point. But it’s a it’s an investment of time. You know, just like when you learned how to use Windows and had switched from DOS and how you you know, when you started working online in a web browser, you had to throw some time at it. When you started working with email, you had to throw some time at so this is kind of the same thing. I think,
Randy Johnston 14:59
yeah, there. Are. It’s clear to me right now that there are things that accountants are doing in Excel, in Power BI, in BI tools in general, in financial reporting tools, many things that you’re doing today that are going to be done differently because of MCPS. So I think the key takeaways that I’d like for you to maybe lock in on, and we encourage you to get involved now rather than later, because you need to begin your learning now first that MCPS standardize AI access to accounting systems, it reduces the risk while improving the consistency. And the real work is going to be in your learning and in the governance, not so much the technology, things like permissions and policies and review procedures are going to actually come into play. And I think MCPS are making AI usable for accountants today, right now, without exporting data or losing control, the MCPS allow that data to stay in your system of record or systems of records, and if you read it out now, you can report on it in ways that you just haven’t before one final twist, which is not MCP, but my vision is that we’ll talk to AI, and the AI will do the work for us, and you can talk to the models that orchestrate the MCPS. So think about not having to type a query anymore, but to talk a query, you know, show me all the meals in the past year where I’ve spent more than $100 boom, and it just goes and gets that data. In the old days that have been a real pain to do.
Brian F. Tankersley, CPA.CITP, CGMA 16:50
Yeah, it would have been a mess. Well, we’d like to thank you for joining us here at the technology lab. Folks. We appreciate you spending your time with us. You know, we look forward to seeing you sometime in the future at a future episode of the accounting Technology Lab. All the best. Good day. 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. You.
= END =
Thanks for reading CPA Practice Advisor!
Subscribe Already registered? Log In
Need more information? Read the FAQs