First Impressions: The New ChatGPT BROWSER For Accountants

Pretty big announcement from OpenAI this week, the launch of Atlas, a web browser that is like ChatGPT first to control your browser for you. Does this mean now that I can cut ChatGPT loose with my day to day Internet browser based tasks? Yes. But also no. Today we're talking specifically about what this means as it applies to like accounting work, accounting firm work. What would this work for, what will it not work for? And what does it, what does it suggest about the future? The, the ever so tenuous. That's not the right word. The ever so confusing future that we are going into headlong whether we like it or not. Let's talk about it. If you're new here, welcome. This is the podcast feed where we talk about all things running accounting firms and, and doing the work that accountants do in accounting firms. And as you may imagine, big theme of the last couple years has been AI. A lot of listeners here are very heavy ChatGPT users. You're not tasting the rainbow yet, man. Pick up that chatgpt for business. Kind of the greatest Swiss army knife app we've ever had. ChatGPT. But what does it mean to have like a ChatGPT browser? Right? So that's the big announcement, that's the News. It's called ChatGPT Atlas and it's a web browser just like Chrome or Internet Explorer or Firefox. Right now it's available on Mac only. And this has always been the case. Anything that OpenAI has launched, it comes to Mac first and then Windows after that. Which is so funny because like Microsoft is the biggest investor in OpenAI. Historically, when they do roll it out on Windows, the functionality is largely the same. Like occasionally Mac will get some features a little bit sooner than Windows folks. But either way, even if you're a Windows user, you will have this in your hands soon. So I think it's worth talking about. So let me run you through exactly what it does before getting into, like what does this actually mean for accounting work? So to be clear, this is a desktop app. This is a web browser that you run just like any other web browser. And when you're logged into it with your ChatGPT account, that is you log into the browser using your ChatGPT account. Then when you first open up Atlas, what you get is this kind of search box that kind of does anything. So you can type in a web address, but you can also type in a question for ChatGPT and it'll jump straight into a conversation. So it's almost like, like Google had a Baby with, with Chat GPT that is the main interface when you open the browser. Now one interesting thing that it does, I, I don't know about you. I probably Google on Chat GPT more than I Google on Google these days. So most of the searches I would have run on on Google I now use ChatGPT for if you're the same way. One nice thing here is if you ask ChatGPT a question like find me the best XYZ and it gives you the results when you click on the like the citation in the text of what ChatGPT responds with, the browser opens up that website but it keeps that chatgpt conversation on the side. It's actually kind of nice. It does this sort of split screen thing where it's like, you know, if I ask it find me the best hair products for accounting influencers and it gives me a little list and each of those items usually will have a citation to a website. If I click on that website, the website takes over the screen. But I still have the ChatGPT conversation over on the side. And there's been a few versions of this out there already like this AI assistant alongside your web browsing experience. Microsoft Edge has actually had this for a while where it's got Copilot sort of on the side. The trouble is it's Copilot, so it sucks. And as you're browsing with Atlas at any time you can sort of pop out this side conversation bar and you can talk with ChatGPT and it can see anything that is on the screen in the website and you can talk about it. And that's, I don't know, fine. That's not super interesting to me. Another feature in Atlas is any text that you select in the browser. Like if you're drafting an email, you can select the text and then tell ChatGPT to do something with it. Like you get this little pop up over your selection and it's like what do you want me to do with this text? And that's like eh, I don't know, not that interesting to me. The one thing that is very interesting, like there's a whole bunch of things you could go into with this browser. Uh, like the way it actually does actual Google searching. Like it'll literally do a Google Google search for you and just show all the Google results. It'll show you all the video results. Like there's some interesting things about the browser, but big picture. The only really cool thing that gets me excited beyond what other browsers already do is agent mode. And that is the ability of ChatGPT to take over the browser itself or a specific tab as you're doing stuff. So I'm, I'm in QuickBooks, let's say, and I'm, I'm doing something for a client. I can have this sidebar kind of companion chatgpt conversation going on alongside of it and say, hey, go do xyz. And it will take over the browser and you get this kind of cool blue sheen over the browser window that indicates that it's working. You can see it moving its little cursor around and clicking stuff and it's like, it's pretty cool. Like the, the user interface, like how they visualize the agent working in your browser. It's, it's very interesting. Also with this browser like part of the reason why they built their own browser is to enable better management around your logins. So anytime you cut the agent loose to do something on the web, you can select whether it should go out and be able to use your logins or not. This has always been the limitation of the earlier versions of this like chatgpt operator is it couldn't save logins so you couldn't really give it like a recurring task and ask it to go out and do this thing because you'd have to log it in and take over and do that whole hokey pokey every single time. Now that they have their own browser, the browser itself has like cookies, which is how the browser remembers it's okay to leave you logged into a website. And so it's made some aspects of password management a little bit better. The reality is for most accounting firm work, especially if you do tax, if you have two factor authentication enabled, which you should, you are still going to have to do two FA Olympics. Like you're still going to have to put in the code and all that stuff. And no browser can like fully automate that for you. Unfortunately, unless you've got like a password manager that will do like the one time code sort of stuff, then you in like reality in practice, you're still going to be fist fighting logins. Unfortunately, even if you are using Atlas to its fullest and cutting agents loose to go and do things like, you're still going to be fussing with logins. Now the big question is how good is agent mode? Like how functional is it? Does it actually do what you want it to do? And people dunk on it a lot. And I mean it's, it's got a long ways to go, but it is still useful. I would argue. And part of the problem is that we use agent mode the same way that we first, that most of us use chat. The very first time we saw it is we jump straight out to the very hardest, most technical thing that we know. That like we're impressed, we're impressed that we know how to do it. And so we go to the machine and we say, ah, surely you cannot do this. And then it can't. And we pat ourselves on the back and we go about our normal lives. But there's a lot of tedious, menial things that it can do just fine. And I don't want to be the bearer of bad news, but there's a lot of things we still have to do that are manual and tedious, like moving files from one system to the other system, updating the status of a project in one system to the other system and that sort of thing. It's already good enough, it can do that stuff. And Atlas can not only see across all of your tabs, it can see into your history. And this is just, this is taking problematic browser history to an entirely different dimension. I mean, not that we need to go too far down this rabbit hole, but some people are going to be really into the idea that they've got this AI that'll that'll kind of be their, their buddy as they browse the web, just always there hanging out. And like Pepperidge Farms, Atlas remembers, if you want it to. So there's a few different versions of what it remembers. First, you've got cookies, which you always have in your browser and at any time you can delete cookies and that will like, you know, make sure you're not auto logged in the website, stuff like that. So that's the first way it remembers stuff and any old browser does that. Second is your browser history. And this looks like any other browser except ChatGPT makes your history more accessible. So if you're like, man, what was that thing I was looking like? And you got like 20% of the details you can tell it to try and go back and find that in your browser history. Also you can give it tasks to do that. Pull from your browser history as well. You can say you can tell chatgpt, hey, that thing I was looking at yesterday that I was doing X, Y, Z with, pull some of that information into this thing that I'm working on right now or go tell me what the last five clients that I worked on in qbo or give me the timestamps or something like, like maybe you're filling out a timesheet. So it tracks your history just like any other browser. But you can actually like use it as a source of, for getting ChatGPT to do stuff. The third way it remembers what you do is, is what it calls browser memories. And these are a lot like quote unquote memories inside of ChatGPT where if you want, it will remember things across conversations and in theory gets smarter the more it works with you because it has those memories. And the idea is that like over time it gets to be a better, more personalized assistant for you because it can remember kind of fragments of all these past conversations that you've had. The same applies to how you browse. And chatgpt like chatgpt right now, I would say has the best version of memories. Like almost all chat apps now have this, like Gemini and stuff like that. I would say Chat GPT in my experience, it has like, it's still limited, but it has the most useful version of memories. But what would that look like in your browser? And if you want, you can disable browser memories. You can delete them at any time as well, but by default it will remember aspects of how you browse. And where that's interesting for work is if I'm in QuickBooks and I've got this thing that I do for every single client every single month, or maybe just a single client every single month, and I tell it go to XYZ and maybe does half of it and I have to help it in a few spots, I want it to be better the next time I want it to do that same thing, right? So a well executed version of browser memories ought to feel like cross training another person. Because so many of the things that we do are repetitive. There's potentially great power in those memories if it's able to recall the last time you had it do a similar thing. And so that's kind of exciting, right? Because the first time you ask that agent to do a task, if it doesn't do it perfectly, and most of the time it won't, it would be really nice if the second time it was just that little bit better. And that's kind of the wild thing about agents. And where this stuff is at right now is this is like the floor for tasks that you will never have to do again. Like, just think about what this is in ChatGPT in general, how it is just like the most general tool imaginable, where with no context, you just tell it to go to X, Y, Z and tries to figure it out on its own ChatGPT and the Atlas browser and agent mode. Like this is the floor for just how intelligent software can be. Because we're not educating it, we're not really giving it specific directions now. We're just saying like, go do this thing for me. And in many cases it can do that, many cases it still can't. But like, just think about how profound that is, that as soon as it crosses that threshold, that's not something that any of us will ever have to do again. And so at what point can it cross the threshold to, you know, you give it a PDF and it completes the QBO bank rec or something like that. Like, these things are very important to pay attention to because it's, it is so close to that threshold of being like exponentially useful to where if you find that one thing that it'll do for you reliably, but it'll do it every single month across 300 month end closes. But yes, a big deal. You know, it's the worst, man, when you hire somebody and they start working for you and you're like, I thought this person knew stuff. The number of times I've been hoodwinked by somebody that interviews well and they have umpteen years of experience and you pull them in and you're like, what you did? You don't. What? You don't even know your stuff, man. This isn't gonna work. Let me tell you what one of this episode sponsors Cloud Accountant Staffing, is doing about that. So cloud account staffing help accounting firms find talent around the world, be it the Philippines, South Africa, Latin America. And they have an online portal where you can go out and see folks that they have, right recruited, schedule interviews with those people directly on their platform. And one of the things now that every single one of those people have is a collection of assessments. I'm looking at one right here. Assessments for business tax, for individual tax, for critical thinking, for understanding instructions, their personality type. And so many of these things like individual tax, this person I'm looking at, they're in the 93rd percentile. And so cloud account staffing, they have a huge volume of folks that they have go through these assessments and they hand pick the highest performers and they serve them up for you to make your wildest dreams come true. Cloud Account Staffing's backstory they actually ran an accounting firm, hired some offshore folks, had an amazing experience, and then just started helping other accounting firms do the same. Interested? I thought you might be. Check out the link down to Cloud Accountant Staffing in the show Notes. Let me tell you about an app that perfectly combines three very important things that normally you need a bunch of different things tools for the app. It is one of today's sponsors, Client Hub. Let me tell you the three things they combine. Number one, an amazing client experience. Because the whole portal experience, the client facing stuff in Client Hub it was actually thought out from day one. It wasn't like this bolt on where most practice management systems are like we're gonna make the ultimate operating system for your accounting firm. And then later on they're like oh it looks like we all need to have client portals. Let's make that now ClientHub was built from the ground up to be super easy for the client to use. To use. That is trifecta point number one. Trifecta point number two is it is also super easy for you to use because my gosh, do practice management systems now require a PhD in learning how to use a super complex thing? Not Client Hub, they're keeping it simple in like a good way. Third part of that Trifecta deep bookkeeping workflow. Sorry tax people, sucks to be you. Client Hub very good for bookkeeping, controllership, casual stuff like that because it's got super deep integrations with QuickBooks and Xero and even has like some built in month end close automation stuff now so you can sign off on like all the all like each account on the financial statements like right inside of Client Hub. Very cool. They're taking it to a very different level of depth specifically for bookkeeping than most practice management systems are. And you get all that one platform buddy. What more do you want me to do for you worth checking out if you're doing accounting bookkeeping stuff. Learn more about Client Hub. See the link down in the show notes. And so oftentimes I think it's easy to be hard on these like early agents that can fall over pretty easily and don't always work amazing. But I mean imagine the alternative like if you do try it and it is like oh dang, it's actually really good at that, like you know, tax research right now it's really stinking good at tax research. And so the day when you use agent mode to try that more kind of end to end task or that, you know that even that one small little thing in whatever tool it is you're using and it does it as good enough for that, that's a big deal. That's like tell the team the humans don't have to do this thing anymore and it's already good enough to where we honestly could probably be finding new versions of that every single week. This task that we normally do that it can now take off of our hands. Now where are the boundaries with this, like what is appropriate here and what is not? And I'll give you some of the same guidance that I gave folks around ChatGPT's agent mode and that is to consider the scenarios where it's navigating the web in a read only context versus a read write context. So it is still very hard to blindly trust an AI to do like a, a sort of task that takes, I don't know, 10 to 30 minutes and just cover my eyes and hope that the end result is, is something I'm going to be happy with. And so if you think about something like QuickBooks, where I mean you can, you can issue payroll, you can do quite a few things where if, you know, I cut my 4 year old loose with that. In fact, I was just scrolling through my camera roll earlier today and there's like 90 pictures of somebody's leg that I think was him and he got a hold of my phone at some point. And you kind of have to assume worst case scenario, I think with AI for now. And so the scenarios where I would cut it loose is I'm fine cutting it loose if I'm watching, like if I am keeping an eye on it and like from a productivity standpoint, is that saving me time? Almost almost certainly not. But right now it is really interesting to learn what can it do reliably and what can it not do. So if I'm watching it, I'm not that concerned if I'm not watching it, if I just want it to run in the background. Right now I think I probably only give it access to like read only environments where I can see stuff and I can do stuff with that information. But I have a hard time just cutting it loose in a sandbox. That could be problematic. Now to be clear, it is cautious. Like the examples they give of like go out and order me all of these things, like it will always come, come back to you. I shouldn't say always. It has a good sense of when to come back to, to ask for approval and stuff like that. And so we're in this kind of sticky transition where it's like, can I actually trust it to go do my bidding yet or not? But get ready for this to be just a core part of work, like having to navigate this discomfort. And at the end of the day I always come back to how it is not too dissimilar to how we employ other human beings to do stuff for us. Because the new intern, Ricky, he's not great either, and he's created some issues that we had to end up going back and fixing. But not all people are capable of getting the same productive output out of Ricky. There are some people who will say, ricky stinks, man. He doesn't know anything. Easier to just do it myself. There's other people where they look at Ricky and they see a piece of clay that needs to be molded. I would say AI is not too dissimilar. Where there's people right now that are finding ways to get tremendous use out of AI, even in its limited form, and the things that it currently doesn't do all that well, people are finding where it can be helpful and where it isn't helpful, and they are finding a productive path forward. I think there's. There are traps with AI just like there are traps with humans where it can be easy to write it off. And this is. I mean, this is evidenced by right now the huge spectrum of opinions on AI, where there's people that are like, I use ChatGPT all day long and cannot imagine a world where I get my work done without it. And then there's other people still who are like, I don't even know what to do with it. Like, can somebody just tell me what buttons to push? So, like, all things AI, like, there's an element of learning how to work productively with agent mode, and you're not going to get out of bed one day and, like, know the right way to do that. Now I say this is becoming increasingly, like, normal and just part of the way that we work, because there is an inevitability to just handing over more of our tasks to AI. And so the progression here, I mean, it started with app level AI, where now we've got, like, an AI assistant built into every single app that we use. Right. And that's actually kind of annoying because I don't want to talk with this AI assistant over here. They can't see into that AI assistant over there. And we have all this siloed information. What I want is the assistant that can see across all that stuff. Right. So what's the solution? Like, integrations? Like, no, that hasn't solved any of this yet. What is the more abstracted version of an AI assistant? Okay, how about an AI assistant at the browser level? Now, at the browser level, that AI assistant can see across all of my web apps that I use. Right. Makes sense that Then that is the next logical progression. Well, what about after that? I got other stuff that I do that's not in the browser. Well, the next step is os operating system level, AI assistant. That like this feels inevitable to me. Like this has to be where we're headed. I've actually been saying this for I don't know, two or three years now and thought it would be Microsoft that like stole everybody's lunch money here with a really good AI assistant built into Windows and whew, boy, have they not delivered on that. You can use the built in like Windows Copilot to like turn dark mode on and off and do some system setting stuff but like that's, that's about it. But you can see where that's kind of inevitably where this stuff goes, right? Like you have a desktop assistant that can be sort of the companion on the side to help you, but it can also take over when needed, maybe run some stuff in the background. And that is just gang like that is the future of work. Right now we're in a very messy transition, kind of like self driving cars where it's like, yeesh. I don't know that I can really trust it to do this because it doesn't get it right all the time. But I think the opportunity still lies in like finding what can it do for us reliably and where do we still need to wait now ChatGPT Atlas. I don't know that this is like if you're a Mac user, I don't know that this rises to the level of like a run out and try it right away sort of thing. I do think agent Mode inside of ChatGPT is really, really good specifically with doing spreadsheet stuff. I actually think the most universally useful for accountants feature in ChatGPT right now is using agent mode to complete stuff in a spreadsheet. I mean it can do. We've, we've had examples on the YouTube channel of where it will go in and take like over a thousand values from a payroll journal export and pull it into like a payroll allocation spreadsheet and it will do that stuff flawlessly in like less than five minutes. Examples where we give it a pile of tax documents and it put it, it puts it into a structured Excel like workbook. It's not very good at creating Excel files and spreadsheet stuff from scratch. It's really good at taking your spreadsheet template, like a work paper template and completing it based on the instructions you give it. That to me is still the coolest juiciest feature in ChatGPT for like our recurring workflows atlas, the new browser. It's cool. And I get like, take it for a spin. I don't know, tell me what you think of it. But for me it doesn't yet rise to the level of like this is a no brainer buy. If the agent functionality was better, where I could like schedule tasks inside of QuickBooks, go in and do this thing, you know, every single morning and it does it reliably, that starts getting really juicy. And honestly I think that stuff is close. Like beginning of this year, my predictions for 2025, I had what some people thought was a hot take saying I didn't think 2025 would be the year of the AI agents. That felt to me a little more like 2026 or 2027 and I would stand by that. Like right now we're seeing very early AI agents and it's amazing the notion that they can go out and plan how to do this work and kind of sort of do it. But right now it's like not good enough to trust in like high stakes applications. But will it be there 12 months from now? That feels much more likely, right? And when we have that and we're almost like becoming these sort of puppeteers that are cutting all these agents loose to go and do all these things for us at the same time and we're just kind of orchestrating all that, man, that's going to be a very different type of work. And you can see a world where there's people that are going to be phenomenal at that and people who just cannot get their head around it. But that reality, I mean we were probably within 24 months of that just being the most productive way to work. Managing a bunch of agents going out and doing all this stuff for you and the folks that are going to crush that, they are going to be like feel native to that new way of working are the people that are testing and playing with this stuff right now. So I would say don't wait, like have these conversations in your team, have somebody that's maybe more ahead of the curve and like doing a bit of testing and they report that information back to everybody else. The riskiest AI plan that you can have is to not have a plan is waiting. People like to think that that's like the safe approach. That is, that is the very riskiest approach to not do anything. If you've played with this yet, tell me what you think. Drop a comment. Perplexity. The kind of AI search engine. They have a similar browser called Comet. Google has said that they will ultimately be releasing a version of Gemini inside of Chrome that can do some of the same stuff. But it just seems like a logical progression where we start with app level AI that eventually gets to browser level AI and then that gets to like entire machine level AI and the future is like a combination of us doing the stuff alongside the AI, doing the stuff. Right. Isn't that that just feel like where we are inevitably headed? Let me know your take. I'm actually very curious. Also, if you are a Intuit Pro Connect user, the the really the only browser based tax prep platform, I would love to hear how Agent Mode or chatgpt Atlas is at Navigating proconnect. That's another example of a situation where I probably wouldn't let it do that autonomously without my supervision, but would be curious to hear you throw some, you know, throw five or six tax docs at it. W2s nothing super hairy, 1099 interests, give it a pile of information and task it with entering all that into the tax software. That's the kind of thing where it's like at least agent mode, working with spreadsheets, it can do that really well. So if I can do that really well in the tax software, that's pretty darn useful, right? That's pretty interesting. Gang, that's all we got for today. You got any AI questions or other questions about running accounting firms, pop that stuff down in the comments. Thanks for being here. I'll catch you in the next one. A heads up for my podcast friends. We put together the biggest, scariest project my team has ever worked on. A big old run of live events around the country headed up until year end to ensure that your 2026 is just the best year ever. Let me run you through the cities real quick. We're coming to Austin, Oklahoma City, Phoenix, Seattle, Portland, San Francisco, Los Angeles, Atlanta, Nashville, Chicago, D.C. philadelphia, New York City and Boston with some super cool half day workshops that roll into dinners. You get CPE for it. It is an event that is built for your entire team to get the whole squad feeling inspired so that if you are a firm leader, you're not feeling like you gotta put everything on your back. Headed into 2026 gonna be a super fun experience. It's gonna be a surprise guest at every city. We're gonna record an installment of the podcast live. At each of these events you're gonna be part of the podcast itself. What? And then I would love to see you there to learn more. Head over to jasononfirms.com events and I hope to see you there.