028 Discussing Karbon's New AI Launch

Another day, another AI release for accountants. And buddy, they just keep getting better. So today we're talking about what Carbon just released. Six cool new AI based features. And it's not just marketing AI, it's actual AI. It's cool AI. We're going to talk about what they do, we talk about how they work, the nuts and bolts, and we're going to talk about how to make them even better. Little, little, little preview, little peeper preview of what could this stuff all look like in the future when it's like it's fully realized, kind of maximized version. Let's do it today on Jason Daly. Feels like just yesterday we were talking about Kanopy's AI release, right? Funny how those things work. Well, just a few episodes back, you'll see that on the feed we talked about the awesome generative AI stuff, Generative AI email stuff. Canopy just launched. They did a great job with it. It was my, my number one tippy top AI thing I wanted most and they nailed it. And it's scary to be the first to move and so Canopy should get credit. They were like, they did it, good job. And they're gonna keep cranking on it yesterday or today. Let me tell you, recording these shows the day before they go live, it's just a whole thing. At some point, Carbon announced a AI. For me, it was yesterday, quote, unquote. Carbon AI automate tasks and work faster with AI and GPT securely integrated into your collaborative practice management. If somehow you're not familiar with Carbon, they are one of the leading cloud practice management systems. The main differentiator for me is Carbon is, I would say the most email centric practice management system. So they were the first ones to kind of marry an accounting practices practice management system with like an email help desk tool. So it's kind of like combination of those two things. Makes total sense when so much of what we do is built on email. But that's always kind of been Carbon's big differentiator. And if you've listened to my rantings about AI, email is going to be so central to how AI changes the ways that we work. It's arguably the aspect of firm running that will be most improved by AI is email and how we request information, all that stuff. So if anybody's suited to benefit from all this stuff, Carbon's right up there, Right? So let's talk about what they shipped first off, something that they put their finger on that nobody else has addressed yet. Specifically the way you can work with GPT is OpenAI the company behind GPT and ChatGPT and all that stuff, company that developed these models, they have their own API that developers can use to, you know, say you want to generate an email, the user gives it a prompt. The software you're using takes that prompt and sends it over to the API for GPT. It gets that generative output, takes it back into your software and shows it to the user. So behind the scenes, there's just an API that they're working with that's provided by OpenAI. Now OpenAI has this big old partnership with Microsoft. Microsoft has invested an outrageous amount of money in OpenAI. And so you can also do the same thing through an Azure OpenAI service. So if you need to do something with GPT, you can either bounce it off of OpenAI's API or Microsoft now has its own, like Azure environments where you can do OpenAI GPT stuff. And when it comes to super security conscious workflows, this is generally how that's happening right now. So, like there's, you know, health groups like putting health and running health information through GPT, like all these very high stakes, secure, like very secure, sensitive types of workflows. That stuff right now is largely going through Azure because you can lock down the ways that you work with the model there a little more so than you can with OpenAI's API. Now that doesn't mean that the other approach, OpenAI's API, is fundamentally not secure. But in the case of what Carbon's built here, it's the first time in an accounting application that we're having a vendor saying, hey, here's how we're using AI with your client data. So it's a little different, right? What Canopy did was, hey, write me a prompt and I'm gonna generate an email based on that prompt. And it may have like a name in it, like a first name, like we're writing an email for Steve here, but it doesn't pull in like the past email history or anything like that today. And that's the difference with what Carbon's rolled out. So from a security standpoint, it's a different approach, it's a better long term approach because then you don't have the same concerns with sensitive information that you may have if you were just using the other API. Okay, I know. Nerd alert. Get back on track. Jason. Okay. Stuff that it does, first and foremost compose an email, it's largely the same as pulled up on screen here for the YouTube folk, largely the same as what Canopy does. So I think it Gives you a pop up and you provide a prompt. Or maybe it just grabs whatever is in the email box already and it composes a draft for you. I actually still think that's really helpful. It's something that I do in ChatGPT a lot. Having it built into the tool is handy because you don't have to context switch back over to ChatGPT. There's some people that still stand by. If I have to write a prompt, a detailed prompt to generate the email, why don't I just write the email? Buddy, maybe you're some sort of email legend where that stuff just flows freely on the first try and as professional and all that, but that ain't me and that definitely was never my team. Like, I think you're probably the best, you know, if you're running the firm, you're probably the best at phrasing that stuff in a good way on your first try. Whew. Do you remember when you were a junior staff member? Oh, the embarrassing emails I've written, they're usually about five times longer than they should be and you take half a day to write them. So not everybody's as fast as you at this stuff. And so this is nice to have how it works basically the same as what we talked about with Canape. And you can actually do largely the same thing in ChatGPT. What happens here is in the program. So in Carbon, you give it some prompts for how you want to generate the email. What goes to the API has a little bit of quote unquote prompt engineering around it. So if the prompt that you gave it was deadline's coming up on Monday, need you to provide those two documents. What actually goes to the language model has kind of a wrapper around that. It says it's like write a professional concise email in a paragraph or two that takes into account the following information and then it injects the prompt that you gave into that and then the language model returns the email that shows up in the app. So you could go to ChatGPT right now and do that. And that's straightforward and not hard, but it is nice having that built into the app, right? So that's like the mechanics of how it works longer term. How do you make this even cooler? Oh, baby, so many ways. Think about all the stuff that Carbon knows, from every past conversation you've had with that client, to all of the open projects that client has going on, to all of the files you have from that client, to the tone of voice that you use in communications with that client, I had clients that I talked with in a wildly different way than other clients. And a big part of the overall quality of the user experience. When I say generate a draft email for me, a really big factor in how positive of an experience that is for me is not only how good the email it is, but how true it sounds to me to my voice. Because I don't want to start shipping off stuff that, like, doesn't sound true to how I talk and the way I communicate with people and think about it. Carbon already has all that stuff, right? It has all of your email history, more importantly, your email history with that specific client. So if I have a weird, silly tone with this client or a more professional tone with this other client when I'm generating an email to that client, it can genuinely compose something that is of that voice. I don't think it does this now, but Carbon has all the information they need to do that to be able to look at those past emails with that client. If you saw the very first ChatGPT video I did on my main YouTube channel, which is one of the things we did there, was I fed it a big block of my writing and said, generate a paragraph description of my tone of voice. And then I had that descriptor and I could take that. And the next time I wanted to write something, I could say, use this tone of voice. And then I put that paragraph in and it would create writing that sounded just like me. So if you think about this Carbon application, totally do the same thing. Generate a couple sentence explainer of my normal tone of voice or even the tone of voice I use with that client. That's part of the prompt that goes to the language model. And then the generated email that comes back is like, sounds just like you, specifically in the way that you communicate with that client. How cool is that? Right? Totally personalized to you, even down to, like, the client voice level. Super cool. Okay, next thing, adjust your tone. Canopy had the same thing. And it's a preset list. I guess the options are a little bit different. The preset list of options they give you right now are neutral, apologetic, conversational, and direct. Apologetic is kind of funny, right? What are contexts where I write apologetic emails? I don't know. I guess when you screw stuff up, but so it's a preset list. People kind of glossed over this with the Canopy launch. I actually think that's really helpful because the emails that are hard to write sometimes, what's hard about writing them is that you're not coming across the way that you want to come across. It's not hard to lay out a set of facts, but to lay out a set of facts and not sound like an a hole or to like to write that to be perceived a certain way, that's kind of what's hard about writing email sometimes. Right. So being able to modify your tone, that's really cool. Something I use quite a bit if I just kind of get stuck. Other interesting spins on this is when you. Oh, hang on. Garage doors coming up. Yep, I do this in a garage. Don't worry about it. Talk amongst yourselves. Oh, hello. What do we got here? Come on in. Say hello. Your name? Quick introduction into the microphone. My name's Heather. Yep. Heather. What stats? Yep. What do you do? Tell us about yourself. I stay home and take care of our three children. Anything else? That's all I do. Where are your kids right now? Well, to two are at school, one's in the car. I gotta go get one. Oh, my gosh, he's in the car. Yeah. Now you gotta go. Special coffee delivery. Anyways, sorry about that. She does that a lot. Okay, so really cool spin on the adjust your tone thing is actually the ability of GPT to assess the tone of something I've already written. So if you kind of turn that on its head, you may write something and be like, yep, I sound like an a hole here. I need to do something. But what if you write something that could be perceived as a hole? Ish. Without realizing it? That's actually a really valuable application of GPT as well. And so some contexts where I've seen that is some folks will use in ChatGPT, they'll use GPT as sort of a critic, or they'll write something and be like, pick holes in this. Give me. If somebody's gonna build a counter argument to this, what would the top five counter arguments be? And that gives you an opportunity to kind of strengthen the underlying writing there. And you can totally do that on emails here. So like a little flag, like, could you be a little more friendly here, bub? GPT is actually really good at assessing that stuff. Where it gets maybe a little tricky is like, what is the point at which it sends that thing to the language model? Right? Like, it's not gonna do it on every single keystroke because it costs money to make an API call to the language model. So is it like when you go to say send, does it run through almost like OG Outlook spell checker where you would hit send and then it would give you A pop up to like replace or fix any typos. Right. Maybe it's something like that that just happens on submit. Maybe at an organizational level you can set an expectation for the parameters that that tone needs to stay within. So let's say your company has kind of a friendly vibe. You can absolutely do some prompt engineering that will give you like metrics on a certain message. And it's kind of arbitrary, but it will tell you like assess from one to a hundred this bit of writing on these different scales. And I could organizationally set some sort of threshold that those messages need to be within. And then when somebody in my team sends something off, if it's not within those parameters, maybe it flags it, maybe it offers a suggestion that will get it within those parameters. Some of the. I've thought about some sort of top down aspects of enforcing this on like your team and how they communicate and some good things about that and some weird kind of Big Brother things about that. The idea that if somebody is speaking in a tone that I don't want, do I want to be notified of that? So like I said, I'm not sure, not sure the right answer there, if that's kind of getting weird or if it's actually a really helpful thing. Because you can also assess how people email across your organization. What is their general tone, Are there specific clients that they're always grouchy with? And so maybe those are helpful metrics to have. Maybe that's overreaching and getting kind of weird. I'm not sure. Okay, next feature that was number two, next number three is compose a reply. Now we're talking. Okay, so somebody sends you an email, when you go to reply, it can generate a reply for you based on that email thread. And this is the first example, like we talked about, of something that's actually pulling in client information. So for the first time, this is some information that ought to be proprietary to you that is going to the language model. And in my book this is totally fine. Doing this through Azure. In my book it's actually not that bad doing it through the OpenAI API. But that's a deeper discussion. But this is the first example we're seeing of a practice management system bouncing client information off of a language model. And there's people that are spooked by that. Here's the thing, I don't think you need to be spooked by that. I think a lot of the scariness of language models right now stems from not from us not knowing what they're capable of. And not knowing what an AI, what a language model is capable of does not mean that you don't know what information it has and retains. So where it's problematic is when it retains information and it does something with the information that you give it that you don't want it to do. What's like, that is problematic. And so you need to understand, is the model being trained on the info I give it? In this case, no. Right now that only happens. I'm talking about OpenAI models here. Right now the model's only trained on what you give it. If you use ChatGPT and you don't turn that setting off. If you're working with the API, if you're working with Azure, like Carbon is here, the model is not trained on the stuff that you send it. It's just like any other API call, like making an API call to QuickBooks to push a bill over with Bill.com like there's nothing there that is retained. And so while there's ambiguity around, like what all are these models capable of, there is not ambiguity around how is the data managed and retained? That's cut and dried. I think people can sometimes use ambiguity to spook people away from new stuff. And it was this way big time with cloud adoption, right? People didn't understand it. And you can create this sort of element of ambiguity around like, oh, we just don't know. So we can't do it when like, no, there are answers there and you can like research it, you can learn that yourself. You can know exactly what's happening with it. Like, there isn't this. I don't know. I think that's like, it's one of those just like in my mind sort of misinformation things that makes it hard for people to make decisions. Like we saw it with COVID and around elections and stuff like that. Like the more that people like to say like, oh, here's this truth, and then someone else is like, yeah, but do we really even know what truth is? Like, it's a way of just like throwing a wet blanket over the whole thing when it's like, no, like just do research, get to the bottom of it. And when it comes to AI stuff, like, unless you're using this like fly by night model that doesn't have like, have any retention policies, which is the same as using like a super immature cloud app, right? But like the mainstream models these days, like they're GDPR compliant, like there isn't really any ambiguity around how they manage data. So this is exciting. To me, the fact that you can now generate an email according to the thread itself, that's a huge time saver. Even if you're still the I'm not gonna save any time if I gotta write the prompt guy. Composing an email from the reply totally skips the notion of having to write a prompt. Right? There's gotta be like, there's gotta be some aspect of prompting that needs to at least be optional for when somebody's emailing you looking for you to make a decision so that you can like indicate what that decision ought to be. But I love that it takes more of that conversation into account and synthesizes that into what it generates. Okay. Fourth, this is a really cool one. It's gonna help you prioritize your inbox. So based on factors like the urgency and the tone that the person's using when they emailed you, sentiment analysis, are they happy, are they grumpy? And the general tone of the email, it will identify for you emails that are high importance versus low importance. How cool is that? So somebody sends you a super sweaty email with something that they think is urgent, it's gonna prioritize that one for you whether it's urgent or whether the client just thinks it's urgent. Like, that's still gonna be for you to decide. But if somebody sends you an email with the expectation that you're going to get back to them faster, this will escalate that. And I think that's helpful. Obviously, what's urgent for external stakeholders isn't always gonna be what's urgent for you. But I'll tell you what I do like is the ability to prioritize that stuff without having to drill into every message. I think that's helpful. The devil will probably be in the details and like how well that one performs for you. And are you going to have clients who just. The way that they phrase things, they just always sound like something's on fire. And so those people always get flagged. But I can tell you I'd rather have it than not have it because it does save you from having to check every single message. If something that could be spicy actually comes through, how this one gets smarter, longer term, it's probably learning more client specific stuff about how they communicate. If you have a client who always seems to think what they need is urgent, but then you always say like, no, this actually isn't important, then the ability for the model to learn that over time that's valuable so that you're not then like making the same changes every single Time and Steve's emails aren't always coming through in this on fire state. Okay, fifth cool thing, compose an email from a task. So in your Carbon workflows you kind of like you templatize these workflows that have all of these different steps in them. And now one of the steps that you can create is to like compose an email, for example, if part of your workflow template is to send a client a like a status update. So if you got five steps in there and the third step is to send them a status update, you can put a step in that workflow template that will generate an email with a prompt. So it's not going to send that email off automatically. But I think actually like building status updates into your workflows makes a ton of sense. The fully automated ones, honestly I'm wishy washy on because in some ways that's noise. I don't want them to get those automated emails and that increased the likelihood that they disregard my human sent emails. But I do like the idea of generating, helping me generate an email that is part of a workflow still subject to human approval and a human sending that off. I think that's really cool ways that that gets better over time. Some of the stuff we talked about before, tone of voice that is client specific, maybe it's more intelligently seeing what's happening within the scope of that project. I don't know if it sees into that right now at all in the way that it generates the email. Another thing that always comes up here as you try to automate communications is did you just send that client an email like 9 minutes ago and it's really weird to send another one now without acknowledging that? Or did you send it to them while you were having lunch with them? Like there are some other sources of context that could make that email a little odd. But Carbon can see all of that, right? So they can see all of the other communications you've had with your team if your colleague just emailed them five minutes ago, you know, add a qualifier at the top there saying hey, I know you just heard from Tina, one other thing, something like that. And so I think longer term it'll get more intelligent in that way. Cool thing about Carbon is it's already got all that information. So. So it's a matter of like what is the best way to engineer that into how that stuff is generated. And it's why also like, I don't think we ought to fixate on V1 of what exactly this looks like today because Just like all things AI, it's getting better really fast. And the value of shipping, it is like you learn what works and what doesn't work and can start building on the best things. And then the things that were like so. So maybe you just ditch those. And then last. This is super cool. Summarizing emails and comments. So I think in like in carbon and I, I see this in a lot of places, like in ClickUp and stuff like that, you have like these sort of activity logs for a specific client or for a specific project and it's kind of like this little like timeline of all the things that you've done with them, right. And with an email or even with a long comment in this case, it takes like drilling into that thing or unfurling this giant thing to see exactly what was said. So like if I'm doing a month end close for a client and the project kicks off on the first of the month and a couple things happen and then Tina sends him an email and then a couple other things happen, that's not really helpful to me until I hop in and see what it is Tina emailed about and read the contents of that. What this is gonna do here is it's gonna summarize those big meaty things into a more bite sized version so that quickly at a glance, you can get a better understanding of what was communicated all in a single view. Pretty cool, right? You think of all the situations where you have to open up an email or drill into an email to remember what it was. If there's a version of that that I can look at that's more concise, that helps me from having to drill into that, that's like really helpful, right? Longer term, how does that get more helpful? I think there's probably other applications for summaries, like documents. Oh, good grief. Document naming, huge pain in the butt. Nobody does it well. And there honestly isn't really like a single way to do document naming. Well, that will tell you everything that's in the document. The summarization of files and documents, that sort of thing. Super helpful both in your own internal file storage, but also in the crap that clients send you do. Your clients fastidiously rename all of the files that they send you that are a mix of smartphone pictures and PDFs and all these different things. Absolutely not. So stuff like dynamically renaming the files that have garbage names that the client sends you and giving, like giving like a one or two sentence explainer of what it is so you don't have to open up every single one of those files. That's going to be super cool. So I think there's other, other interesting applications for summarizing things, but this is a great first step. So bravo to carbon. I think this is a great reminder that you never know what folks are working on behind the scenes. Because when people saw the canopy announcement the other day, there was a lot of carbon folks that were like, wow, when's carbon ever gonna ship something? I wish they'd get to work on the same thing. And they got something really cool here. They've shipped arguably even more stuff. So all that is to say, you know what, like our little tribe here and the fact that we're talking about AI and all these things and encouraging people and rewarding the people who are shipping stuff, it is genuinely making a difference. So carbon shipping more than canopy did. That's great. Like balls and canopies court. Like, this is what we need to. So we need to do in a responsible way. And I don't see anybody doing anything irresponsible here. But we need to keep pushing the envelope because when Microsoft Copilot's sitting out there and you've got other like, big mainstream options that are super powerful, like, we gotta keep the ball rolling. Ultimately the best thing for us long term are the accounting specific solutions. But if people stop using them because Copilot and stuff like that are faster to move, that's bad for us long term because Microsoft Copilot's never going to see into my client's accounting ledger. And like anything accountant specific, that longer term could be really cool application. Really exciting time right now. What a week. Like we've been talking about this stuff back to last year and kind of concerned that people would be slow to move in our space. And I feel so much better than I did a few weeks ago. People are working really hard on this stuff. I got another killer AI scoop that is every bit as exciting as this. I hope we can talk about it tomorrow. I don't know if it'll be ready for public consumption yet tomorrow or not, but there's even more cool stuff around the corner, gang. So thanks for coming and hanging like and subscribe. Share this with your friends, with your mom and all that stuff. Bravo to Carbon. We're gonna keep highlighting the people that are doing good work on this stuff because I think it's important and it's ultimately helping us a lot. Like this stuff will. Will enable kind of a totally different approach to working and helping us to focus on the important stuff. Helping people rather than did we already ask for that document or not? So good stuff. What a fun week. I'll see you tomorrow.