AI Automation Ask Me Anything 📱

Hello, hello everyone. This is Isaac, and guess what? I was bored. Well, not really bored. I just got out of a Q&A in an interesting community. Shout out to Future Firm for those of you that know it. And I've been meaning to just starting to go live a little bit more regularly, seeing, seeing if anyone joins. Uh, we don't have anyone here at the moment. But I just want the practice of going live and starting to answer questions. Because I just happened to come out of a Q&A session, I thought I'd go live and share some of the questions that came up there in case they're useful for any of you. Some of the questions that came up in that community of accounting firm owners were, what's my current take on Claude versus Copilot versus ChatGPT? What's currently the best? AI tool and how they compare. What is the current best practices on security and can we put client data in there? That's another question that came up. A use case came up, a specific use case around connecting Anchor, which is a proposal software, with Sage Intacct and how to think about that. And I have a good framework for that that we been putting— trying to really put in place to help people understand what are the steps you take to assess how something can be automated in the first place. And then I would like to wrap up with that particular framework and some examples on things we are actually doing at the moment in the agency that might be— might benefit some of you. All right, let's get started. Claude versus ChatGPT versus Copilot. Here's the reality. We are really in a race, right? The AI race for the best tool. And as such, it's a very developing space. And the, you know, first person that's winning, it's really a moving target. From our perspective, however, and the bet that we're making, Is that Claude right now it's winning and it will continue to do so for the purposes of specifically accounting firm? A couple of reasons for it. Um, one, we really like Claude's— uh, well, for context, for those of you that don't know, actually 8 people have joined. Come on, let's go! If you have any questions, could you let me know? But I'm answering right now. Actually, perhaps I should filled with energy and answering right now, uh, Claude versus ChatGPT versus Copilot, like what's the best tool and what's the bet that we're making? What's your current take on that? And here's where I was going. If you all don't know the history behind Anthropic, the maker of Claude, it actually resulted as a conflict between employees within ChatGPT about how they were treating security. So I guess let that sink in. Like Claude came as a result of people being concerned about how ChatGPT treated security. A couple things that have— we've seen from that that makes it more comfortable using it. Not only is it better, let's be honest, if you had put all of them to the test right now for accounting from work, Cloud is the best, bar none. It is. But also, cloud has a bigger share of the enterprise market. Think governments, think nonprofit institutions, think banking, think finance, that are arguably industries that are even more highly regulated than some of the things that accounting firms do, as an example. So our take is that that leads, leads us to believe that Claude will continue to be better from a performance standpoint. But even if it fell behind and ChatGPT surpassed it for a little while, we're still betting that Claude will treat security more seriously because of the deals they're trying to get, which is with more enterprise companies as opposed to ChatGPT. Regarding Copilot, however, I mean, I don't know what to tell you, but it's just not as good, unfortunately. Like, in every test we've done, it just doesn't perform as well at the moment, especially if you're on the free version. But even though the paid versions and tests that we have performed and the feedback we've gathered, Copilot, it just really sucks. It's really far behind. And the other 2 options that are not here, that they have its place for certain use cases are Gemini and Perplexity. Gemini, we found it to be very good at honestly research, specifically in the context of things that are within YouTube, things that are within Google already. And Perplexity seems to be also like a good search engine from the perspective of being able to pull different sources and then making sure that everything it tells you actually has a source. To cite. So that's our current take on Copilot versus ChatGPT versus Claude. The other question I wanted to ask is like on the topic of security. I'm actually bringing a security expert. So if you're seeing this June 25th, in about 2 weeks, we'll be talking— 2 to 3 weeks— we'll be talking to an expert on the topic of AI security to tell us like the truth about which You know, tool is the best from a security standpoint. But our current take and the current consensus is that if you are on either ChatGPT or Claude's Teams plan, many firms are comfortable putting sensitive information from clients, right? As long as in their own accounts. And of course, all of the other IT security practices would apply from the perspective hey, you're giving each user an individual account, so that allows you to have an audit trail backend. You're also making sure that they have like 2-factor authentication so bad actors don't log into their accounts, because ultimately that's the biggest risk, is another human logging in as opposed to, you know, some AI agent doing something weird. And that's something, perhaps another note worth highlighting there. Many AI frontier models and frontier companies actually have an incentive to eliminate personal information. Why would they want your Social Security number or your client's bank account for being trained to the model? Their answer is actually they don't want that data sometimes, and they have ways to try to eliminate and remove that data if it is at all being used to train their models. But part of why people want to make sure you opt in for the Teams plan or the version that teams for the relevant AI tool is that you'll be able to turn off the data being trained or the model being trained on your data. And you'll also be able to have some certain, like, administrative and permission privileges settings that would allow you to either not allow something for your team, like a particular product, like loading Chrome as an example, something that we disabled, or like a connector, right? You're connecting Slack or you're connecting Gmail, and you'll be able to remove that from the remove certain actions from Claude or ChatGPT to be able to do without a human approval at the end of it. Now the other thing I wanted to talk about is like, how do you know what to automate with? And that's a big question, but I think that after years thinking about it, I have a good answer. That is Useful and widely applicable. Okay. And I'll walk you through it. So we have a little bit of a slide that walks us through it. But let me give you the context. I mean, as we've grown as an agency, obviously we're on a daily basis making decisions about, okay, what's this process? what do we automate? And it's something that actually sometimes landed on me as the owner to make some of those architectural decisions. So the more and more I sat down and thought like, okay, there's gotta be a framework, it's gotta be a decision tree for us to properly select what tool to use when. Okay. And I think I've landed on a spot that I'm happy with for now and that I want to share with you. Ultimately, there's one prerequisite before I share this that you have to keep in mind, is that regardless, the first step is always understanding what to automate. And I would encourage everyone here to not try to like look for the latest you know, 17 AI use cases YouTube video, but to pay attention to their own calendar, their own time, and the things that they're currently doing in their day-to-day as that source of information to find the use case, to find what to automate. Turn on your automation radar because there are things you've done today that you've done yesterday that you've done last week that are probably ripe for automation, and you're just like looking and Googling For something else when you have things already there that you can use to start buying back your time. That said, once you know what to automate, there's really, from our perspective, and this is the framework I'll talk about now, there's really 2 other questions to answer: who and where? Who's gonna do it or execute on that step, and where is the data? And now what that gives us is two factors, right? Two variables from which to assess how do we go about automating something. I'll give you a couple of examples. As part of that Q&A that I just dropped off of, someone asked, "Hey Isaac, I want to." Ensure that every time an invoice is issued in Anchor, which is my proposal software, an invoice is also created on Sage, which is, uh, the software that this particular firm was using. So she knows that what— when an invoice is created on this software, I want to create it on this software too. How do we use this framework to know what to build with and when to use what? The directive here is to pick from left to right the cheapest who to do the job. Prioritizing agents that are AI native and will be able to work— you'll be able to develop them more easily. Following it out with automations. be able to just run in the background. And following that with, hey, maybe there's not a feasible way and we need to have a human in the loop and some sort of interface for a human to interact with and make that job easier for them. So as an example, for that firm that they were cool, we want to move from Anchor to invoices from Anchor to Sage. The first question I asked, do they have a native feature? So this first row, it's out of the question because there's not really an off-the-shelf or native feature that allows her to create that connection and automate that process. So as you can see here, you're going to go from left to right, but we also want to go from top down because these are the wheres. This is like where you can get access to the data, and you want to prioritize from the simplest one To the most complex. So for her, again, no off-the-shelf features available. Cool. What's next? Native connectors. In this case, this is the next piece we would evaluate. Does Sage and Anchor have native connectors on any of the ed tools you use? And the answer for her was no. So the next step was to evaluate some of the automation platforms to see if in this case those 2 apps that she wanted to automate had the actions she wanted available. To explore that, we simply pulled a— the directory from Zapier. Let me see if I can share that as well. Let me share that other screen. Here we go. Show window. Okay. So what do we did with her? Again, none of these things were an option. This wasn't an option either. So you just have to go row by row, left to right, until you find the tool that's going to get the job done. And actually, there's something missing here. Here we have Chrome extension. Is custom Chrome extensions as the bot-related human-in-the-loop solution of choice. But in this case, again, none of these made sense. This didn't make sense either. So this is where we landed because this person had used Zapier before. I just thought, cool, let's log in and see if Anchor is available, which is one of your apps. So this was part of the, like, live research we did. An anchor was indeed available. We aligned that for her. It was not just when a proposal was accepted, but specifically when an invoice was issued. So this is the trigger that she wanted for that process. So I tell her, hey, this will be your 4th step, and your last step will be— guess what? You want to create that invoice in Sage. So we look for Sage Intacct, which is the things you want to use, and you have here the ability To create an invoice. So first step will be when an invoice is issued in Anchor. Last step will be create an invoice in Sage. There'll probably be some steps in between to be able to make that work, but this is how we landed and say, hey, this is the cheapest who— not only who— and easiest where the native connector to automate that particular process for her. If that would not have been possible, we would have simply gone through and keep going in this order to see if, okay, Sage didn't— if Sage wasn't supported by Zapier or Hanger wasn't supported by Zapier, we move down a step and say, hey, is there any APIs or scripts that we can now create to be able to access that data and to create those steps? In this case, create that invoices that she needs. So this is something I've been refining probably for years, and it's taken a long time to get to this point. And this column of agents was just added probably 3 months ago because previously it just was— it was only automations. But if you've always— if you've ever struggled with, hey, I know what to automate, but I want to know exactly what tool to use This is the best framework we have. And I'm sure the apps in this table will evolve over time, but in my attempt to systematize and document how we think about what tool to select, this is the best answer I have to date. Another example, just earlier this week we onboarded a firm and they are They work for the restaurant business. They do accounting for restaurants, and they have to do a daily sales reconciliation. What are we thinking for them? Let's actually look at the slide again, right? Well, first thing we ask is like, well, currently would they do that in Excel? Good. And they're using a specific software. That's pulling the daily PNL to that Excel. Great. But they also have to pull Toast from that to that Excel, and more importantly, then they have to perform a variance analysis to understand when do those daily saves from QuickBooks and Toast don't align, and when they don't align, understand why. This is an instance where I would literally Laid out every step. Step 1, get input from QuickBooks. What do we use for that? We'd recommend a connector for that that's able to pull that data. Then second step, we want to pull the data from Toast too, and that's one that we're currently evaluating. Is Toast— does Toast have a cloud connector? Does Toast have the Zapier Make integration? We haven't done the research yet, but this is literally the path we'll go down. to assess what's the best and easiest way to get the Toast data. Once we have the Toast data and the QuickBooks data, guess what? Well, what are we going to go and do? Again, starting from top to bottom, left to right, create a skill to do the variant analysis now that all the data is there available. And ultimately, that's probably what we'll land, right? Get the data from QuickBooks, get the data from Toast, Do the variance analysis, and there would be different tools we use to automate each one of those pieces. Again, just thought I'd share that other example. Well, glad for the handful of people that joined as I came in live just without any notice whatsoever. I'll be doing more lives probably on Thursdays afternoons. So if you all want to join in, we'll probably have some communication shared either on our email newsletter if you are opted in for that, or on our socials. So follow along because I do want to do a little bit more hands-on training and MNAs as people have more and more questions about AI automation. Other than that, thanks for joining me. If you did, thanks for watching the replay. If you did, and I hope to see you on the next one. Take care and stay efficient.