Earmark Podcast | Earn Free Accounting CPE

Attention: This is a machine-generated transcript. As such, there may be spelling, grammar, and accuracy errors throughout. Thank you for your understanding!

Jeff Seibert: [00:00:00] No one's going to be sort of outcompeted by the AI itself. You are going to be outcompeted by firms that really adopt this aggressively and sort of bring that future forward. And so I would just be really eager to try as many things as you can find and start with your simpler clients, start with your sort of more tech forward, risk taking clients and see how much of their process you can automate.

Blake Oliver: [00:00:24] Are you an accountant with a continuing education requirement? You can earn free Nasba approved CPE for listening to this episode. Just visit earmarked in your web browser, take a short quiz and get your certificate. Welcome back to the podcast I'm Blake Oliver. Today we're tackling a big claim an AI native general ledger that says it can automate most bookkeeping work without losing control or trust. My guest today is Jeff Seibert, co-founder and CEO of digits. Jeff recently unveiled Digits, autonomous G.L. and new accounting agents. And we're going to unpack how they work, where the human stays in the loop, and what a smart first pilot looks like for your firm. If you've wondered whether to switch from a legacy general ledger or how to price an AI powered client accounting services offering. This episode is for you. Jeff, thanks for joining me.

Jeff Seibert: [00:01:17] Blake, it's so great to be here. Thanks so much for having me.

Blake Oliver: [00:01:20] Jeff. What is an autonomous general ledger?

Jeff Seibert: [00:01:26] Yes. So we started this company in 2018. It's been a long journey with the goal of reimagining accounting in the age of machine learning. And that's the underlying technology. Ai, of course, is a term didn't really exist in 2018, but it is fundamentally machine learning. And what really struck me was the opportunity to try to automate just all of the tedium that goes into the core accounting workflows, particularly during the monthly close. And so the goal with the autonomous general ledger is to automate 90 plus percent of the bookkeeping of the statement reconciliation of the flux analysis, the sort of insight generation, the report generation, so that you can focus on the actual value added work that really helps your clients. And the tedium is sort of taken care of for you.

Blake Oliver: [00:02:13] So how is this different than a traditional GL like QuickBooks or Xero?

Jeff Seibert: [00:02:18] Yes. So it is a fundamental architecture difference. And this might get slightly technical for a moment. But the way to think about QuickBooks, Xero even all the way up to NetSuite, Oracle, etc., etc. is they're all fundamentally relational databases. They were all architected 20 to 30 years ago, and they treat each transaction as basically a row in the database, and they treat it as text. And so let's say your client took an Uber ride. Quickbooks is just going to see Uber. It just sees text. It sees whatever you type in. And so you type in those transactions and you've probably noticed they sort of sit there in the database. And then when you run a report, it prints that data back out to you and of course aggregates it by your chart of accounts, etc. but it doesn't understand the data, it doesn't know what Uber is. And so digits is a fundamentally different technical architecture. We use what's called a vector graph data model. And this means everything is an object. And so Uber is an object. Airbnb is an object. Your marketing category expense category is an object. So is travel. Your bank accounts are objects. And so through this, we're basically able to map the flow of money through each business. And that allows the AI to build what's called a semantic data model, a semantic understanding of the money flow. And that's how it can get so accurate at booking transactions. And so the best way to summarize it, it's a it's a data architecture difference. That's pretty fundamental to how the game works.

Blake Oliver: [00:03:49] So we've gone from a relational database to a vector graph data model where each of these vendors or I mean, is an account part of this too? They're all objects.

Jeff Seibert: [00:04:05] Yes. So think about your vendors, your customers, your bills, your invoices, your chart of accounts. And then each transaction is basically a vector, a connection in that graph. And so then you can build a clustering a mesh network of all the financial transactions in that business.

Blake Oliver: [00:04:23] This sounds a lot like how like an LLM is structured. I don't know much about it technically, but I know there's vectors are involved and sort of like each of these concepts like the, the the words in a in an LLM are like connected by in space in a way like that. Is this similar to that?

Jeff Seibert: [00:04:43] It is very similar to this. So in Llms they tokenize the input and output. So each word is converted into tokens. Each token could be a short word or a piece of a big word. And those are converted into what are called vector embeddings, which are literally just numbers. And so it's it's plotted in n dimensional space. You can think of your typical 2D graph with a line on it, but in n dimensions it's sort of hard to think about. Um, we use a very similar model here. And so we're able to convert each transaction into this vector embedding. And that allows the AI to understand how it sort of lives in that n dimensional space. I realize this gets a little weird.

Blake Oliver: [00:05:23] Well, maybe you can, um, help make it real for us. Uh, like, by walking us through a transaction. You mentioned Uber. Yeah. So. So let's say I, I'm going to a conference. Uh, I'm going, um, today to Seattle for work, and I'm going to use an Uber. I'm going to take an Uber to the airport. Yep. So I spend that on my, uh, spend card. Like, how does that show up in digits compared in, like, let's compare it to a typical GL.

Jeff Seibert: [00:05:53] Yeah. So in a typical GL right you would connect to your credit card. It would come into the bank feed. It would then sit in the bank feed for your review. Or perhaps you've set up a manual rule that would then book it to travel let's say right in digits. We also connect live to all your banks and cards. The transaction comes in, there is no bank feed and there are no rules, right? What happens is the transaction is then basically processed and understood by the AI models. And we spent five years training these models. So I built our machine learning team way back in 2020. And we have trained all these models internally on 170 million plus transactions, um, almost like $1 trillion in real small business financial activity. And these models basically understand all of this, all of these transactions and how they were booked before. So when Uber comes in it converts that Uber transaction into this vector embedding. And then it looks into this model and it's like, oh what is that. How has it been booked previously. And then what is similar to it. And so what's really fascinating is the emergent patterns here. So when you have that scale of data and you see how everyone is booked Uber before you start to see, oh, wait, Uber and Lyft are often booked the same way. And so the model starts learning and it's like, oh, actually Uber is very similar to Lyft. And so if it sees Lyft in your accounting for this client it then knows how to book Uber. And so you can sort of look at this pattern detection that it starts building. So the way to think of digits I can pause there for a moment.

Blake Oliver: [00:07:34] Oh yeah. Yeah. So um, so because of the vector graph data model, Lyft and Uber are similar vendors. The model understands this or digits understands this. And I think the thing that is is big here for me. Is that in a traditional relational database like QuickBooks, part of the reason that Intuit struggles to do this kind of work for us is that each each client company is its own thing. Yes, there's no visibility across all of these relational databases to learn from, right. So, so, but digits is like all the transactions for every company. It can it can learn from all of them and apply those learnings to your company.

Jeff Seibert: [00:08:23] So that's exactly right. So the way to think of digits is a layer cake. And so it's not just one model. We actually have different tiers of models that allow us to be sort of as smart as possible in the moment. So the top tier of model is trained on each individual client's books. And so if a transaction comes in that that client has seen before, it's effectively perfect because it mimics exactly what was done with that transaction before. And you never need to create rules. None of this is manual. It's just fully automatic. If that transaction has not been seen by that business before, we then have our global model which trains across our entire data set. And it looks at, oh, how did another business in digits book that transaction? We then translate that into the chart of accounts for the client in question. And so digits works with any arbitrary chart of accounts. You don't need to standardize anything. That's what you get from our global data set. If we have never seen the transaction before, it's completely novel. Let's say a business just opened down the street. We then fall back to an LLM to basically go, and this will get into the agents, I'm sure, in a few minutes to go and build a dossier on that transaction. So it's like, what would you do as an accountant? You would probably Google it. What do our agents do? They literally Google it and they research the transaction, build a dossier about it, and then we give that to an LLM to do what's called zero shot classification on that transaction. And so it has no prior history of it, but it makes the best guess and gets the fall back.

Blake Oliver: [00:09:55] It has your chart of accounts. Yep. And it it builds that it does that web search that I would do as an accountant. And it takes the information and it, it finds the best fit expense account.

Jeff Seibert: [00:10:08] Exactly, exactly.

Blake Oliver: [00:10:09] I got it. Okay. So it's basically working the same way a human bookkeeper would work. Um, so that takes us through to the categorization stage of this transaction life cycle. Let me ask you one more thing before we move on to reconciliation, when we're categorizing, a lot of times even the web search doesn't give us enough information, for instance. Right. Uh, I'm going to Seattle for a particular event. And maybe the nature of that event determines how we're going to categorize this Uber transaction. Right. How do you get all that context Into digits AI. That is not there. Like a bookkeeper might have to ask me, like what event are you going to? Or look on my calendar or something like that. How do you handle that?

Jeff Seibert: [00:10:56] Yeah. So so two different ways. Um, so we do allow there to be additional context stored on the client that the LM has access to. And so it's like, oh, where is this client based, what type of business are they in and so on. And this is just plain text. This isn't like some big profile you need to fill out. Um, that allows you to sort of bias the LM towards booking it correctly. But you're exactly right. Your example is perfect. Some things, like you just don't know it's something unique to the business that they did that month. And it's not going to know that. And so the way this would show up is if there's a transaction, the business's history that's booked in different places, right. It's uncertain it moves around. Then our models are going to be low, confident in their booking. Anything that's low confidence we immediately throw out and it's put in the inbox. And so then you, as the accountant or the client can go in and go through the inbox and book it how you want to.

Blake Oliver: [00:11:50] And what percentage of transactions are going into that inbox.

Jeff Seibert: [00:11:54] So right now about 4 to 5% relatively small.

Blake Oliver: [00:11:58] That's pretty good because I'm thinking back to when I had my cloud based accounting firm. We we were able to use rules to automate a lot of this coding. But it was like 80% maybe.

Jeff Seibert: [00:12:09] Yeah. Okay. Yep.

Blake Oliver: [00:12:10] So you've gotten it down to like four. Did you say 4 or 5? Four.

Jeff Seibert: [00:12:15] Yeah. So I can share some news here. So we've we previously published our benchmarking digits was 93.5% accurate in the spring. As of the August close we hit 96.5. So we're continuing to march up.

Blake Oliver: [00:12:29] And is that calculated based on when people change what digits did.

Jeff Seibert: [00:12:33] Exactly.

Blake Oliver: [00:12:34] Got it. Okay. So 93% of the time digits codes it and it doesn't ever get changed.

Jeff Seibert: [00:12:41] That was the spring number. Spring as of August 96th, 96.

Blake Oliver: [00:12:45] Okay. And it's probably just going to go up from there because these models keep getting better and better. Yep. And I imagine it's getting easier to connect the external sources for context. Like I noticed that in ChatGPT. Now I can connect my Google calendar.

Jeff Seibert: [00:13:01] Totally. Yeah. So so we can keep bringing in more and more context. Exactly. And this is the right way to think about it. Basically, if you're able to share the context with digits, it can get really, really smart. And so our goal is to just keep adding additional layers of context.

Blake Oliver: [00:13:17] So this is exciting because I've always thought of AI powered bookkeeping as limited in the sense that sure, it can do standard bookkeeping with a standard chart of accounts for a typical business, like a startup, for instance. Yep, startups are pretty easy, right? It's just like cash burn.

Jeff Seibert: [00:13:36] Right? Exactly.

Blake Oliver: [00:13:37] Payroll. Standard expenses. A lot of SaaS. Yep. It's not complicated. And so all these startups start to do bookkeeping, and then they go and they try to serve other businesses like construction companies and they realize or nonprofits where it's fund accounting and you realize, oh, this is this is way more complicated. Totally. But I could see with the context of like a construction companies, projects that they're working on that the AI could, like figure out if it has enough context. Okay, this purchase was for this project.

Jeff Seibert: [00:14:13] Exactly, exactly. And you hit on all the hard examples. I will say each additional percent gets harder and harder for us to achieve. But I think just long term, directionally, you can imagine if you're able to bring in all that context, it's theoretically possible.

Blake Oliver: [00:14:27] That's where you're going.

Jeff Seibert: [00:14:28] Exactly.

Blake Oliver: [00:14:29] So right now, like what are the businesses that work best on digits?

Jeff Seibert: [00:14:34] So I mean, you hit it. If your business is straightforward digital based right. Like relatively reliable month to month. We have many examples in our system where digits is 100% perfect every single month. And like, that's the dream, right? Because everything's predictable. Um, we're currently really pushing to improve our support for nonprofits. Um, we're bringing in other sort of sources and connectors and so on. Um, and so we'll just keep expanding. We'll have more to announce at the end of this year on sort of new verticals that we support.

Blake Oliver: [00:15:06] Okay. So great fit for. Yeah, standard cash basis expenses. I'm not doing a lot of project tracking like that's you've nailed that. Yep. And we're we're going to start doing those. I guess one way to think about them is just additional dimensions in the GL. That's how I've always exactly. Yeah. Got it beyond just expense accounts. Um what about. Well I want to talk about journal entries and stuff, but I think when we talk about agents, we'll get there. Let's go through this. Let's not get sidetracked. Let's keep going through this transaction lifecycle. So we've got the Uber transaction. Uh, it's come through into digits and it's been categorized. And my business, because I'm a consultant essentially is pretty simple. So we assume it's gone to travel expenses. We're good. Um, what about, like, reconciliation? Because categorization is just one part of the process.

Jeff Seibert: [00:15:57] Yep.

Blake Oliver: [00:15:58] The reconciling may take I mean, these days, it definitely takes more, more time. So like how how does digits how does digits different than a traditional GL when it comes to actually closing the books?

Jeff Seibert: [00:16:09] Yes. Great question. So reconciliation is the other major time sink in the close. And this has been our major effort this year. Um, so you can already reconcile the accounts in digits. You can go and mark them and check everything off and so on. It matches what is coming shortly, as in the next few weeks. Um, I'll sort of tease it here is completely automated reconciliation. So we go and pull the statements from the banks automatically if the bank supports that. Um, if your bank doesn't support that, you can just drag and drop the PDF onto digits. You don't need to worry about converting to a CSV or anything, whatever. Um, we have done a lot of work in document extraction and getting really good at understanding bank statement structure. All the major banks have slightly different statements as you're aware of, and digits now pulls the transactions from the statement automatically reconciles each account, and the magic of it is we reconcile to a pixel bounding box on the statement. And so you can literally click on the transaction and see it on the statement and vice versa.

Blake Oliver: [00:17:13] That's that's beautiful.

Jeff Seibert: [00:17:15] That's really that's the power of AI, right? It's like it's so tedious. No human would ever do that.

Blake Oliver: [00:17:21] But you need to be able to click through and to see exactly where on the PDF we're mapping to. That's that's always been the issue with these AI reconciliations is how do I trust it exactly 100%. So that so can I ask you a really nerdy question about how he's doing these reconciliations? So one of the fundamental differences between zero and QuickBooks is how they handle reconciling QuickBooks allows you to clear transactions, and then you can reconcile them by doing a month end report where you reconcile. For a long time, zero didn't have that. They just do a 1 to 1 match. How does digits approach the transaction matching aspect of reconciliation?

Jeff Seibert: [00:18:06] So that's a really interesting question. So we we match the transaction to the statement. So the status of each transaction is captured right. Does that transaction match the statement. Then at the period level or whatever time interval you set you basically it's like does the full reconciliation tally with what's in the ledger. And then that portion is marked closed.

Blake Oliver: [00:18:28] And it's locked like these transactions. Exactly. That's the issue. Right. The big problem is we reconcile and then we delete a transaction or somebody messes with it.

Jeff Seibert: [00:18:37] Correct. So the key thing is you're right. So basically once it's reconciled, once you close it, that is locked. We do of course allow you to manually override the lock if you need to for some reason. In that case the reconciliation will automatically reopen and it shows in our UI why it reopened, as in this transaction was deleted, this transaction was added whatever it might be.

Blake Oliver: [00:19:00] Okay. Reporting.

Jeff Seibert: [00:19:03] Yes.

Blake Oliver: [00:19:05] How is it different? The same than QuickBooks or Xero for reporting? How does digits handle reporting?

Jeff Seibert: [00:19:11] Yes. So our goal with reporting was this is your opportunity to storytell. You want to you want to bring all the data together as easily as possible to tell each client what they need to be focused on. And I always felt as a as a sort of serial entrepreneur, I felt really confused and constrained by what I was getting from my accountant, because it would just be the standard Excel export of here's your P and L balance sheet, etc.. Um, and I wanted like, visualizations. I'm very visual. I wanted to sort of understand the business. So we went and built what I would call Google Docs for finance. So digits reports are live in browser, fully customizable. You can change the cover, photo, titles, etc. etc. you can drag and drop any graph or component into the report. So let's say you want to deep dive on marketing spend. You literally just type marketing m-a-r-k and drag it into your report and you get a graph of your marketing spend.

Blake Oliver: [00:20:05] That's beautiful.

Jeff Seibert: [00:20:06] It's very cool. And so you can you can tell a story to a client in seconds just by drag and dropping a couple components in. And then, of course, we do have the full traditional PNL balance sheet, cash flow, aging reports, etc., etc. all of those do the analysis for you, and so you can just turn on the color coding and it highlights what change and by how much. And then we also automatically do insight detection on each line. So it's like why did marketing jump 20%? Well, you don't need to go run a different detail report. You just mouse over. It's a feature we call hover to discover, and it tells you, oh it spiked because of an increase with X. And so we do vendor and customer level attribution. You can see the top transactions that resulted in the delta from the prior month. Et cetera. Et cetera.

Blake Oliver: [00:20:51] That's great because I think variance analysis was the most time consuming aspect of what I did when I was in client accounting services. It's just so tedious to pull up the the GL detail and to try to figure out what it was that was different.

Jeff Seibert: [00:21:08] Completely agree. And this is actually a funny note for you. This is not AI. It is fascinating to me how many startups there are claiming like we do AI fpna and like bring up your variance analysis and so on. We do not use any AI for this. Variance analysis is statistical analysis, and we just do brute force stats over the entire data set for the month and surface it. And computers are really good at crunching a lot of numbers very fast.

Blake Oliver: [00:21:34] And it makes sense because, uh, what? Ai is statistical?

Jeff Seibert: [00:21:38] Yep.

Blake Oliver: [00:21:39] And and variable. And we want our variance analysis not to be variable. We want it to be deterministic.

Jeff Seibert: [00:21:48] Correct. I actually I met a founder a few months ago. I won't say the name, who told me she was building an AI dashboard, and she was really excited because it was 98% right. And my reaction was, that's useless. Like, if it's not 100% correct, what's the point?

Blake Oliver: [00:22:04] Yes, we were just talking about this on, uh, the accounting podcast, David Leary and I, where it has to be like 99.9%. We're talking, you know, or 100%. I mean, it.

Jeff Seibert: [00:22:16] Should just be 100%.

Blake Oliver: [00:22:17] Should be 100%.

Jeff Seibert: [00:22:18] The math has to be correct. Yeah. And so this is one of the other big things we've done with digits is we we fundamentally don't use llms that much. And when we do have to use them, we prevent them from doing math. And so instead, we built our own financial modeling engine that we give to the LMS as a tool. They can request to do math. And it's done in our financial modeling engine so we can guarantee it's correct.

Blake Oliver: [00:22:42] That's the right approach, is you use the LMS to like drive the tools to to use the calculator to evaluate the output of the calculator, but not to actually do calculations.

Jeff Seibert: [00:22:55] Exactly correct.

Blake Oliver: [00:22:56] Just like people. Right? We don't expect you don't give me a complex math problem and expect me to solve it perfectly right every time, unless I'm a human calculator. But there aren't that many human calculators around there, right? There's there's very few. Yeah, it's funny to me. Um, just side note, Jeff, I'm curious about your thoughts on this. The more I explore AI generative AI, the more I come to realize that it is. To effectively use it. You have to think of it like it's a person.

Jeff Seibert: [00:23:24] Yes, 100%. And right now, basically a junior employee? Yes. And so give your junior employee all the tools and all the context they need to do a good job. And don't assume they can do something that they obviously can't do.

Blake Oliver: [00:23:39] The phrase we've been using a lot in accounting, I feel like, is, you know, a clever intern.

Jeff Seibert: [00:23:43] Yeah.

Blake Oliver: [00:23:43] Because we we have a lot of interns that come into accounting firms and they have knowledge that they've acquired in school, but they have no practical experience.

Jeff Seibert: [00:23:52] Correct.

Blake Oliver: [00:23:53] And so when you hand a task to an LLM, it lacks context. It lacks real world experience. You get an output that's very similar to a clever intern.

Jeff Seibert: [00:24:03] It's a very, very eager, very overconfident, clever intern.

Blake Oliver: [00:24:08] A very. Oh yeah. I went to a really good school. Yes. You know, paid a lot for that education and, uh, it thinks a lot of its abilities. Yes, exactly. That's great. Okay, let's talk about the accounting agents. This is the. We're in the current hype cycle of accounting agents or or just AI agents in general. And of course, it's getting hype in the accounting space. There's all this talk about it, you know, of course, every time something comes along like this, will it take our jobs? Right? It will totally automate all of this work. I'm curious to hear your thoughts on just AI agents in general in our space. And then let's. And then after that, let's talk about like how you're doing it at digits because you had a big announcement about this.

Jeff Seibert: [00:24:56] Yeah. Let me first just say what they are because I think they're also caught up in this whole hype wave. And it just it drives me crazy because of course crypto was big and every company had crypto. And then AI is big and every company has AI. And now agents are big and every company has agents. And it's it's largely just sort of hype. Uh, the reality here is what is the difference between an LLM, which now folks are generally familiar with, with ChatGPT and an agent. An agent is simply an LLM that you run in a loop, and so you give it a task. The LM attempts to do the task. You then ask it at the end. Did you complete the task? If not, what are the next steps? And then it says, okay, here's what I did. Well, here's what I didn't do. Well, here's what's next. And then boom, the loop continues. And so it runs in the loop until it believes it's completed the task. And this is a relatively simple mechanic that can have sort of profound behavioral, uh, like benefits and like it can achieve some very interesting things. But as we were just saying, it's not going to do anything that you don't give it the context or tools to accomplish. And so the prime example in digits, we we are pretty sure we were the first in the accounting industry sort of at all to run agents.

Jeff Seibert: [00:26:10] We've been running them in production since January of 2024. So almost two years now. And their first task was exactly what I described googling transactions that our system didn't know about. And they do it actually shockingly well. And they do it with less human bias than many of you would, because they don't have just this instinct of, oh, I've seen that transaction before. It has to be X, and I'm just going to do a quick search and then book it. They're thorough every single time, and they never get tired. And they research them 24 over seven. And it's actually it does drive remarkable accuracy. Um, so that was our first one. And then the second one is this reconciliations, which we spoke about reconciliations, a very tedious sort of looping process. You want to go through all of the transactions, find the best match for each one in order to reconcile the books. Um, and so that's how we think about them. We try to use them in tedious work where it's not really a time bounded project. Like, it just needs to figure out this transaction. It needs to figure out how to reconcile this statement. And that's what they're great at.

Blake Oliver: [00:27:10] And you mentioned the human in the loop aspect, where you're surfacing the low confidence transactions to a human to review it, approve it or, or change it, I assume with the reconciliations. It works that way too.

Jeff Seibert: [00:27:26] It's exactly the same. So if it does achieve the reconciliation, you actually have the choice of whether it should sort of auto finalize it or you as the accountant still want to review it and finalize it. That's a choice in the product. Um, obviously, if it doesn't successfully reconcile it, it's immediately surfaced for your review. And so you'll open it and it'll be 90% done and matched, and you can see all the pixels align. And then you'll have to help it figure out what it couldn't get.

Blake Oliver: [00:27:51] Um, just like, uh, one of my human bookkeepers.

Jeff Seibert: [00:27:55] Yep. Exactly.

Blake Oliver: [00:27:56] Okay. Um, what about, like, other. So so we've got researching transactions. We've got reconciling transactions or reconciling statements. Yep. What other agent use cases are you excited about working on? What are you hoping to automate with agents?

Jeff Seibert: [00:28:18] Yeah. So just like QuickBooks, digits does offer full bill pay and invoicing built in, so it makes it very easy to migrate clients over. Um, and they're already relatively automated. So you upload a bill, you drag and drop it onto digits, and we extract everything, including the line items. Um, and it's remarkable because they're way more accurate than the old OCR approaches everyone's familiar with. And so you can just click on the bill, look at it, click approve or send it to the client, approve, click pay, etc.. Um, what we're really excited about and this is not in yet, but sort of development is thinking through how we automate collections and nudging the recipient to pay and like, how did they manage the RFP process for you? Because it's the same thing, right? If you have if you have a clever intern who you've set to do AR, AP ops for someone, we can look at those processes and sort of guide an agent to take those steps.

Blake Oliver: [00:29:09] So getting approvals on bills, following up for those approvals.

Jeff Seibert: [00:29:13] Following up nudging. Yep.

Blake Oliver: [00:29:15] Because that's that's that's that was always probably the the most day to day time consuming tasks that we did was was that but that that totally seems like something an agent could do because like you said, it just runs in a loop and its objective might be get the appropriate approvals right, get approved.

Jeff Seibert: [00:29:33] Get paid, whatever it might be. Yeah.

Blake Oliver: [00:29:35] And so you just have to teach it well like you would an employee. Okay. How often should I follow up with somebody. Right. So I don't annoy them too much. Right. Obviously not every hour.

Jeff Seibert: [00:29:48] Right. Right.

Blake Oliver: [00:29:49] Maybe maybe once every few days. And then it starts to get urgent and then every day.

Jeff Seibert: [00:29:54] Right. And it's like, start very polite and gentle and then maybe get a little more direct in the emails.

Blake Oliver: [00:30:00] So. So how much are you allowing digits users to customize those agent instructions? Because that's something that as I build my own agents, I'm very aware that the instructions can really change how the agent behaves.

Jeff Seibert: [00:30:16] Definitely, yes.

Blake Oliver: [00:30:16] So how much are you hard coding on your side. And how much are you exposing to the users?

Jeff Seibert: [00:30:22] So I don't want to tip our hand on our team of a of an upcoming announcement, but there will be a lot of customization options for our accounting firm partners. Got it. So digital ship with sort of a default behavior. But we are going to let the firms pretty thoroughly customize how the agents operate.

Blake Oliver: [00:30:38] I love that it's always good to give people a default setting and then allow them to customize.

Jeff Seibert: [00:30:45] Exactly. Yep.

Blake Oliver: [00:30:46] So that's that's that's the difficulty of building a product. Right? Is like how do you keep it simple but also make it complex enough for all use cases in accounting especially.

Jeff Seibert: [00:30:57] This is one of the hardest parts. So I way 20 years ago I started my career as an engineer at Apple and I love basically opinionated design, right? You want to take a stance. The product should stand for something, but you do need to strike the balance because if you can't customize what's important, it's not going to fit everyone's use case and they're not going to like it. And so that's probably what we spend most of our product design time on, is the right balance of exposing features versus not adding complexity to the UI.

Blake Oliver: [00:31:27] It's funny you said Apple, because that's exactly what I was thinking of, where a very non-technical, non-tech savvy user can pick up an Apple device and it just works. And the default behavior is great for them. But then somebody who's a little bit more sophisticated, like you or I can go in and and tweak these settings.

Jeff Seibert: [00:31:46] That's exactly right. And this is what I tell our team, like my goal is to, is to build the sort of accounting platform that Apple would if they ever did that. Apple's never going to build accounting software. But imagine they did. What could we build?

Blake Oliver: [00:32:00] Let's talk about trust security. This is always a question I get every AI presentation I do, every time I go out in person and talk or on a webinar. How are you approaching the Security of sending data to LMS of the models you're building. One of the fundamental differences of digits that you've just talked about is that the the database is structured so that you can learn from all the companies, and it's not siloed. So how do you maintain security in that environment?

Jeff Seibert: [00:32:39] Yes. This has been our major engineering effort over the seven years of the company. So as quick background, this is my third tech startup. I was previously head of consumer product at Twitter. Our team has worked at Twitter and Google at immense scale. We have basically architected the entire system for this because it's so important. Um, and so what we do, all of our data is encrypted at rest. In Iowa, we use Google Cloud. We host everything in Iowa. Um, it does not leave our systems. And this is the big difference between digits and other players in the space. We train our own models. And so we are not sending raw data to OpenAI to these third party foundation models. We are not letting them train on the data. Everything I talked about, we've trained in-house. And so that's been really fundamental to us. Um, that allows us to get really specific. So we use something that's called object uh, object envelope encryption. What that means is each object in our system is encrypted with its own key. And so even like, heaven forbid we get hacked, you can't just steal a key and steal all the data. You'd have to steal N keys for each object in the system. Um, so it goes really, really deep. And we developed these approaches at my prior company, Crashlytics, where we were basically doing mobile crash detection for all of the major apps in the world. Crashlytics runs on 6 billion smartphones. And so we had really sensitive data like we knew Uber's Dow MoU. All this public company data that was not.

Blake Oliver: [00:34:09] Right because you're collecting logs for every app like this is this is the This is the most. This is everything.

Jeff Seibert: [00:34:16] That was the keys to their kingdom. Yeah, exactly. And so we've taken the same approaches here.

Blake Oliver: [00:34:21] Well that's great. So yeah. And at Twitter I mean everybody's like DMs.

Jeff Seibert: [00:34:28] Yes.

Blake Oliver: [00:34:29] You know from a security standpoint.

Jeff Seibert: [00:34:31] Under my watch they they previously weren't when I was head of product we started encrypted DMs.

Blake Oliver: [00:34:37] Okay. That's great to hear. And the audit trail right. It's important that it's, uh, secure that, you know, if something if there's a leak or a hack or whatever, that it's it's locked down, like you said. Um, but we also need to be able to see what happened and why it happened. Audit trails are super important to accountants. How are you approaching that when it comes to these agents doing the work when it comes to auto categorization? Like like if something gets auto categorized, how do I know what happened?

Jeff Seibert: [00:35:06] Great question. Um, so everything is logged and recorded within digits. You can click on a transaction and see the activity log. See who edited that transaction when. So this solves the common issue of like no one loves having their clients in QuickBooks because you don't really know what they're doing. Um, in digits, it's very clear what's happened. And we have restricted roles. So you can just have view only access and in digits you can actually do per object access. What I mean by that is like you could invite your head of marketing and they only see your marketing expense category. They don't see payroll, they don't see your bank accounts. Anything else you deem sensitive. So it's very locked down.

Blake Oliver: [00:35:45] So that's and that's impossible to do in a traditional GL. So you're saying I can invite somebody and they could just see everything to do with marketing. Yep. Including all the transactions, everything that's been tagged.

Jeff Seibert: [00:35:59] So so imagine this right. Like you're used to saying, oh no, we're over our marketing budget. Why? Well, we didn't find out till the end of the month and we didn't know where we were and so on. With digits, you can give them live access to marketing expenses and remember its book 24 over seven as transactions come in. So there's no excuse. And yeah, they can see marketing. They can see all the transactions book to marketing and nothing else.

Blake Oliver: [00:36:20] I love that.

Jeff Seibert: [00:36:22] Um, so yeah, to finish your your question. So we've taken the security extremely seriously. It's Soc2 type two certified. You can go to trust.com and get all the details. Um and so yeah happy to answer any questions from anyone with concerns there. We've we've taken that extremely seriously.

Blake Oliver: [00:36:39] Let's talk about moving clients to digital first. Actually I want to ask you about taking on Intuit.

Jeff Seibert: [00:36:47] Yes.

Blake Oliver: [00:36:47] Because that I cannot imagine a more ambitious goal than to try to take business from Intuit, which is, I think it still stands as the only multi-billion dollar company that has been built on, like, direct to, uh, small business software as a service. There are there are very few and they have done it and they just dominate. They are 8,090% market share. Zero tried to come to the US after being very well established in Australia and New Zealand, and has struggled to even get 10% market share. And here you are digits. Yeah, I mean you have this incredible background Twitter and Crashlytics, Apple all that. Um, but still it's like it's it's you're a you're a David against a Goliath or you're a, you know, like they are a giant. Remember when they had those commercials with the giant robots? Yes. So I just want to ask you about that. Like like, uh, what why why go after this giant and, um, you know, it just seems so challenging. Like to to accountants. Just, like, stuck on QuickBooks.

Jeff Seibert: [00:38:07] Yeah. It is a really great question and all credit to Intuit. I mean, they have built a truly impressive business over many decades now. And so that was hard by by all means. Um, I would say what got me very sort of obsessed with this space. Remember, I had no background in finance or accounting. I was building developer tools. We were doing consumer product at Twitter. I got so frustrated by the difference in data quality between the product engineering side and the business side. And so in product and edge, we have Google Analytics, Grafana A B, testing, dashboards, log analysis, like you know exactly what your servers are doing, whatever user is doing in real time as it happens. And then as the business owner. I was waiting 2 to 3 weeks to get a black and white PNL, an excel sheet from our accountant. And so at first I thought that was just a small business issue. And like big companies had it sorted out. And then at Twitter, I literally went to corporate finance and I was like, oh, I want to run an event for our consumer org. What's the budget? And their actual answer was, oh, we haven't run those books yet. Can you give us three weeks? And I was like, you have 100 people in corporate finance. Like, what are you doing?

Blake Oliver: [00:39:15] Yeah.

Jeff Seibert: [00:39:16] And so that was literally the seed. And so when I left Twitter, I was like, I need to figure out what's going on. I talked with tons of business owners, spoke with tons of accountants, and I realized the problem was not the people. Everyone is very aligned on like wanting the data. The problem was the software. And so what I saw was an opportunity to disrupt the status quo because of the tech transition. And so, like each of my companies have sort of ridden a tech wave. My first company in creo, back in 2000 and 708 was basically web 2.0 document collaboration. We built the first capability to to render files in a web browser. We ended up getting acquired by box and our tech powered boxes document display for years. Um, with Crashlytics it was the rise of mobile and apps were crashing. And we could tell you literally in two seconds what line of code you had to fix. And then at Twitter, as had a product. I launched the algorithmic timeline, which was one of the world's first global deployments of ML, now called AI. And so I understood.

Blake Oliver: [00:40:14] And for our listeners, like, the reason that's a big deal is because previously the timeline was like just it was whatever was most recent.

Jeff Seibert: [00:40:23] Correct. It was a strict chronological. Right.

Blake Oliver: [00:40:25] And and the algorithmic timeline, like, changed everything because now it's surfacing the stuff that it thinks will be most interesting to you.

Jeff Seibert: [00:40:33] Exactly.

Blake Oliver: [00:40:34] And that's that's what really I mean, I feel like Twitter really took off after that.

Jeff Seibert: [00:40:38] Both Twitter and Facebook did roughly the same time. Exactly. And in social media, it has great power. It also has great problems that it causes. I helped make the Netflix documentary The Social Dilemma to sort of explain all the downsides. And so I was looking for a business. I could go into that used machine learning, but had no downsides. And I figured if we automate accounting like that would be amazing. Like then we could have a live finance dashboard. And so that was the that was the premise. And yes, you're going head to head versus into it versus versus this monstrous company. But I felt that as a startup, we could move a lot faster with a brand new tech stack and basically literally code the architectural advantage into the core of the product.

Blake Oliver: [00:41:23] And that's really the thing that Intuit can't change.

Jeff Seibert: [00:41:28] Correct.

Blake Oliver: [00:41:29] They are stuck on that 30 year old code base.

Jeff Seibert: [00:41:32] Yes.

Blake Oliver: [00:41:33] Well, in the case of QuickBooks online, I guess it is like 20, 20, 30 years old. They haven't changed.

Jeff Seibert: [00:41:39] And here's an example for you. As you know, in QuickBooks, you can't have a vendor and a customer with the same name.

Blake Oliver: [00:41:46] Right?

Jeff Seibert: [00:41:46] Right. Why? Because they apparently chose name as the primary key in their database.

Blake Oliver: [00:41:51] Oh. Big mistake. I mean, they should have known better. Exactly right. You use a unique ID and obscure that behind the scenes, right?

Jeff Seibert: [00:42:02] Exactly. Those are the type of things that are just incredibly hard to change. That'll be a years refactor.

Blake Oliver: [00:42:08] Oh, I mean, to do that across I mean, how many, you know, businesses, how many, how many of these little databases do they have like to actually make that change?

Jeff Seibert: [00:42:17] They likely have one monster database, but I don't know their internal architecture.

Blake Oliver: [00:42:21] Yeah, I always I always assumed that was the issue because like, that's why they've struggled to like do the categorization correctly.

Jeff Seibert: [00:42:30] That could be two, right?

Blake Oliver: [00:42:31] There must be something architecturally that like prevents it because I think this is the problem where they have, you know, every company has its own chart of accounts. Yep. And there's no like, unified chart of accounts like yeah, Xero tried to do this. They have a like higher level chart of accounts as a practice that you can map all your clients to. Yep. But it takes a lot of work. I assume that like your vector data vector graph data model, just like does that.

Jeff Seibert: [00:42:57] That's what it does natively. Exactly. Is that's really our secret sauce. Like you've probably heard of some hype. There's a whole bunch of YC startups trying to do AI, accounting and so on. But when you look at what they're doing, they have a standard chart of accounts, which they then send to GPT.

Blake Oliver: [00:43:12] Yeah, you have to be able to have the customization. All right, Jeff, so let's talk about getting clients on to digits from QuickBooks. Yes. So what is the ideal client in my firm to try with digits to pilot.

Jeff Seibert: [00:43:30] So we support a wide range of clients to get the most sort of impressive first time experience. I would start with one of your simpler digital clients, right? Relatively consistent books month to month. You should be able to see it almost fully automated almost immediately. Digits has two modes. You can obviously start fresh with digits, and you can either upload a chart of accounts if you want, or digits will actually suggest a chart of accounts and build it as it sees transactions come in. So that's sort of a cool option if you have a brand new client.

Blake Oliver: [00:43:59] Yeah, that's really nice.

Jeff Seibert: [00:44:01] So this is I should have mentioned another thing the agent does. It can literally look at your transaction feed as it comes in and build the chart of accounts up from nothing based on how it thinks those should be booked.

Blake Oliver: [00:44:12] That's that's always one of the most tedious aspects of starting with a new client is trying to figure out what is my default, what is my chart of accounts I have to take. I would take a big template that I built with hundreds of accounts, and then I would have to whittle it down for that particular client.

Jeff Seibert: [00:44:27] So under the covers. Yeah, the agents actually doing the opposite. It's starting with a massive list and instead of whittling it down it's just turning them on.

Blake Oliver: [00:44:36] I love that.

Jeff Seibert: [00:44:37] Um, but of course. So that's for for brand new folks. Uh, for folks coming from digits, you can bring in all their data, right, in the ad client flow. And it takes 2 to 5 minutes super fast.

Blake Oliver: [00:44:49] What about, uh, pricing for firms, you know, like, how do you. One of the big complaints is actually the pricing of QuickBooks.

Jeff Seibert: [00:44:57] Yep.

Blake Oliver: [00:44:58] Like, it just keeps going up every year. Yes. And you hear you hear accountants griping about it because, well, we're not doing anything more with it, but the price is going up. So like, how are you competing on price?

Jeff Seibert: [00:45:10] So digit is simple. It's $100 a month. Uh, we launched with that price back in March. We're not changing that price. Um, I don't like pricing games. That's why it's not $99 a month. It's just $100 a month. Um, and we do have volume discounts for accounting partners. So we have a full accounting partner program. You can get in touch with our team, we'll give you discounts based on your client count, etc.. Um, we also have additional SKUs. If for folks like just your right up clients, you only do taxes for. We have a ledger only SKU. Um, if you have clients that for some reason you only want to do reporting for, we have a SKU for that. Um, so basically come chat with our team and we'll work with you. But the core digits plan $100 a month. Super easy.

Blake Oliver: [00:45:54] And how are you seeing firms that are moving clients onto digits. Pricing their packages of offerings for their clients.

Jeff Seibert: [00:46:03] This is a really great question. Um, we've seen a couple different models work. So what we've seen over the past just years in general is firms moving towards more of a fixed pricing model, which is great for automation because then obviously as it gets more automated, you as the firm save time and can serve more clients. Um, we've also seen folks do a sort of hybrid model where the monthly close is a fixed fee. Um, and then they do per hour advisory work on top of that. I think both work great. If you're charging purely per hour right now, then the automation may make that challenging for you because the whole goal is it does whittle down your hours over time.

Blake Oliver: [00:46:43] So on that note, how do you see AI changing staffing at firms? I think we hinted at it with this accounts payable agent ID Idea. There's not many firms that employ people who just do accounts payable. I guess some larger ones do. But you know what? What jobs are going away is that is that one of them? You know any others? What do you see growing? What do you see shrinking?

Jeff Seibert: [00:47:12] Yeah, totally. Um, so I guess first off, high level, there is no world in which AI quote replaces accountants. Um, the. That is crazy to me. Knowing how the tech works as we talked about it, is purely probabilistic. It will continue to get tricked by edge cases, etc. etc.. Um, there is a world where it does automate 90 plus percent now at 96% of the bookkeeping tedium. And so I do think the sort of junior bookkeeping roles won't be that productive for that long. Um, what we really see is firms already sort of trying to do this is whether they're have already outsourced a lot of that roles. They're already sort of up leveling their team towards advisory work. I think that trend will continue. This will accelerate that. Um, and it'll allow you to sort of shift into more of a review mindset across the board. What gets me really excited is if you talk with junior folks, people like still in college debating whether to join the profession. I think this actually makes the profession far more attractive. Right. Because you don't want to just sit there doing data entry all day long. You want to help learn how to advise businesses. And I think this will create a faster path to that.

Blake Oliver: [00:48:24] What do you think about education? What should accountants be learning when it comes to AI? There's so much. Whenever I go out and talk to them, they're overwhelmed. You and I, we live in this. We play with it every day. But if you're a typical accountant or firm owner, you don't have a lot of time. Yeah. Where would where would you say they should focus?

Jeff Seibert: [00:48:44] So it is definitely overwhelming even for folks who are in the space. Um, the key thing is I would be disciplined at trying new tools and set aside some time each month. Obviously, like during the close, it's a nightmare, but set aside a couple days the week after you've sort of finished your monthly close to try a new crop of tools and see what could be helpful for you. And the reason I say that is no one's going to be sort of outcompeted by the AI itself. You are going to be outcompeted by firms that really adopt this aggressively and sort of bring that future forward. And so I would just be really eager to try as many things as you can find and start with your simpler clients, start with your sort of more tech forward, risk taking clients and see how much of their process you can automate. And we know from our internal data it is like 96% is straightforward. That's our median. Now you can get to 100% automation for a really simple client. And so I would challenge everyone pick one client on your firm and try to go that path and see what you can achieve.

Blake Oliver: [00:49:47] Yeah just start small. Start with one. Give it a try. I also recommend, uh, watching YouTube. Uh, you can learn so much on YouTube. Uh, I learned how to swim. Watching, like, like real, you know, I mean, I knew how to swim, uh, from, like, impressive swimming. Yeah. Like to actually, like, you know, like, swim fast. I watched YouTube videos, and there's all these incredible Australian swimmers out there just giving away knowledge online, and I. And they have underwater cameras and like it's incredible. And that's just one example I wanted to learn about AI agents. So I went on YouTube and I just started searching okay. Ai agents, you know, AI agents in accounting. But you can also just learn about them in general. You just have to be careful because there's a lot of hype also on YouTube. So you have to find try to find the channels that are less about hype and more about real stuff. Yeah, but it's like it's all out there. I mean, you can learn anything now and and you can actually learn it by asking, you know, a large language model. You can you can use ChatGPT to like learn stuff.

Jeff Seibert: [00:50:49] Now just be skeptical because remember, it is a clever intern, right?

Blake Oliver: [00:50:53] It's it's going to guess. I like to think of it too, as like I have a ten year old and he hallucinates all the time. So it's sort of like it's like that's actually a natural thing that we do is like our brains. We don't we don't know the answer necessarily, but we have like an idea of the answer. And so, you know, we'll just we'll do our best. And sometimes it's right and sometimes it's a little bit right, and sometimes it's totally off. But it goes back to that idea that AI is really modeled after human intelligence.

Jeff Seibert: [00:51:28] It is remarkably so. So I have a two year old and he's obviously just learning to talk, just learning to identify things. And he saw a fish a few months ago. He's now obsessed with fish, but I've been shocked by how he can recognize just the head of a fish. Just the tail of a fish. This shape fish, this other weird shaped fish. And he knows they're all fish. And you can sort of think, like, see his mental model Expanding. Yeah.

Blake Oliver: [00:51:50] Yeah. We're children are building a mental model of the world. And it's that's. I think that's what's hard about AI is that all previous tech innovations have worked fundamentally differently. Correct. And learning how to code is, is like all the how am I saying am I help help me say this right. It's like zeros and ones, right? True and false. That's been the foundation of all tech up to this point.

Jeff Seibert: [00:52:24] Yeah, up until now. So all coding in history has been telling the computer what to do. You're basically writing the rules. Now with machine learning, you are simply giving it the goal state, and it is figuring out what to do to get there. That's the like crazy mental model difference.

Blake Oliver: [00:52:42] And it's so exciting. It's so great to talk to you, Jeff.

Jeff Seibert: [00:52:47] No thank you. This has been amazing.

Blake Oliver: [00:52:48] It's been very educational. Um, so. Yeah. Everyone listening? Give digits a try. Jeff, where should accountants go to learn more about digits?

Jeff Seibert: [00:52:57] So go to digits. Com super easy. You can literally start a free trial. Add one of your simple clients. Or if you just want to see the magic happen, sign up for an account, link your personal credit card and it'll auto categorize your credit card. That's probably the fastest way to just like, see the AI work. Um, and then of course, you can cancel the trial. Um, but then if you click on our accountants tab, we have a full partnership program. Our team would love to speak. We can obviously get you discounted pricing, get your clients set up, uh, help train your team, etcetera, etcetera.

Blake Oliver: [00:53:28] I have been speaking with Jeff Seibert, co-founder and CEO of digits. Jeff, thanks so much for your time.

Jeff Seibert: [00:53:35] Blake. This was awesome.