Interview with Enrico Palmerino, Founder & CEO of Botkeeper

February 18, 2019

Read my follow-up article to this interview, โ€œItโ€™s not a bot โ€” the truth about Botkeeper, the Google-funded, AI-powered โ€œbookkeeper replacementโ€

Blake Oliver interviews Botkeeper Founder and CEO Enrico Palmerino on February 18, 2019 (transcribed by Sonix)

Blake Oliver: The question that I have had for a long time that this incident has resurfaced, which is I'm very curious to know what is the extent of artificial intelligence, of machine learning, of automation - those, I understand, are really three separate things - in Botkeeper that you were using ... How are you using it?

Blake Oliver: I guess I should preface this by saying that I went back, and listened to one interview in particular that you did on Fox Business News on November 27, 2018, in which you spoke for several minutes about Botkeeper. In that segment, you spoke at length about how you're using artificial intelligence to do bookkeeping, and accounting, but I don't believe in that interview you mentioned that there was a human element, at all. Is that not, in some way, misleading, given that you are using a team in the Philippines to do the work?

Enrico Palmerino: I think a better question there is we ... It is not misleading, because we are not using a team in the Philippines to do the work. We're using a team in The Philippines to validate the work being done, and to assist in other administrative components, similar to how we tried to schedule this call, or meeting. I have a personal assistant in The Philippines that facilitates meeting, and scheduling for me. We've got a dev team in The Philippines, a pretty large one - dev, and dev-ops - that build a lot of the software that we're doing. We also have data-enrichment team. Data that comes in, they'll basically go out, and try to find additional details, or information, whether it's on the web, or elsewhere, to enhance the information.

Enrico Palmerino: It's kind of a common practice to use a team to, whether it's take a company name, for instance, and then find as much information about that company that you can, either that may not be able to be pulled effectively from screen-scraping, and pull that information into a centralized database, and then enrich it with other fields, such as maybe the location, employee count, other pertinent information that could help, or enhance our machine learning, and software.

Enrico Palmerino: The interview you referenced was ... I'm not sure exactly how many words I was able to get in on that. Obviously, it's a pretty short interview with Fox Business News. They asked a few questions on the fly, which I was not made ... It was not prefaced on what those questions would be, so I did my best to communicate as quickly the topic, or answer the question that they had as possible.

Enrico Palmerino: There was no question asked about whether or not we used a combination of people, or software. The question they asked in particular was about the software, so I spoke solely about the software that we use. Me not speaking about the human assistance on the software is not a matter of trying to be misleading, as much as it was me just trying to speak specifically to the point that was addressed, and having very limited ... As you mentioned, I might have spoken for a minute or two, and I've probably already, in this conversation, spoken for a minute or two. It's a very limited amount of time to answer multiple questions.

Blake Oliver: Fair enough. We should probably define what we mean by the work, doing the work.

Enrico Palmerino: Botkeeper, we define ourselves as being a bookkeeper. We are not an accountant. There's a reason why we use the name Botkeeper, because it was supposed to be a combination of robot, and bookkeeper. What Botkeeper does is a really great job at entering data, and properly classifying that data, and reconciling that data across multiple systems.

Enrico Palmerino: We have a team of human accountants. You'll notice on our site, it's ... The idea of Botkeeper is that we deliver a solution, which is accurately doing bookkeeping, and accounting for you. We do that through a combination of machine, and human. The machine is our own machine that is then built to sit on top of many other platforms.

Enrico Palmerino: The easiest way to think about that concept is, if you're an accountant, or let's just say you're a business owner, and if you bought Expensify, Bill.com, QuickBooks Online , Receipt Bank, Hubdoc, let's say, a variety of other tools, and just by purchasing those alone, your accounting doesn't get done. You can't just clap your hands, and walk away. You still need a bookkeeper to reconcile the data between those different tools to properly classify/categorize that information, and then apply the appropriate accounting workflow, or principle to it.

Blake Oliver: Let me stop you there for a second. You mentioned these third-party tools. Would you mind listing the third-party technology that you have integrated into Botkeeper? Whatever you can remember. I know there are often a lot of tools. I, myself, when I had my own outsourced bookkeeping practice, I probably used a dozen, at least.

Enrico Palmerino: Our tech stack includes, I think, a lot of the prominent tools that I think most accountants are familiar with; tools like QuickBooks, Bill.com, Expensify, TSheets. We just launched the first QuickBooks-HubSpot integration. That's going to be made available, I think, soon on HubSpot's platform, as well.

Enrico Palmerino: We build out a bunch of these different integrations into various CRMs. We connect them to Google Analytics, other social media platforms. We link into different banks, credit cards ... A lot of like basically where financial data is going to come into. We'll connect into payroll platforms; think like the ADPs, Paychex of the world. Gusto, for sure, is a big partner of ours ... A plethora of accounting tools that serve various aspects of the accounting process.

Blake Oliver: I'm imagining, of course, that the types of services that you're providing under the bookkeeping umbrella would be bill pay; I imagine expense management-

Enrico Palmerino: Yep.

Blake Oliver: Payroll processing, right? Do you do [cross talk]

Enrico Palmerino: Not necessarily the processing, but the administration on the payroll platform that does the processing.

Blake Oliver: Would that included onboarding new employees, and offboarding employees?

Enrico Palmerino: Yep.

Blake Oliver: Typical ... The kinda things that I might have done, and, well, that I did do in my practice, which was not just the financial statements, and the categorizing of transactions, but also the processing of bills, and accounts payable. Do you guys do accounts receivable?

Enrico Palmerino: We will do invoicing on behalf of our clients, yep.

Blake Oliver: You do invoicing, and then, of course, administration of payroll, but it sounds like you have the clients running the actual payrolls.

Enrico Palmerino: We'll have the client running ... We may run the payroll, but I still consider us a processor, because like I think of ADP, Paychex-

Blake Oliver: Gusto's the processor.

Enrico Palmerino: They're the processor [cross talk] We're using their tool to do the process end.

Blake Oliver: Got it. Same thing as Bill.com is the processor, the payment processor. We are just administrating it.

Enrico Palmerino: Correct [cross talk]

Blake Oliver: Let's talk about the automation, or, specifically, the artificial intelligence aspect involved in each of these processes - in the bookkeeping, and the accounting, in the bill pay, in the payroll, and the expense-management reimbursement. Where does the artificial intelligence come into any of this? So far, this sounds very much like the practice that I had, which had no artificial intelligence, just human intelligence.

Enrico Palmerino: The artificial intelligence comes in at the point of inference, and interpretation of data. If you think about it, right now, data comes into your ... Say it would come into your practice. A lot of data might hit QuickBooks staging area, where it then gets properly classified, or categorized.

Enrico Palmerino: Bills might come in, but you would send them to Bill.com, and then someone in your firm would have to manually key in the category, the ... Select the vendor; put it in the appropriate workflow; enter the due date; other pertinent dates in Bill.com, before it could be processed. You would have sales contracts that might be sent to you that you would then turn into an invoice to be sent on behalf of the client.

Enrico Palmerino: Where the AI comes in is actually understanding what that piece of data is that was sent to us, and then, determining its proper categorization, classification, and other pertinent details to then perform an automated action. This is where it goes from being AI, which uses machine learning to do the processing, into a more robotic-based process automation, automated workflows, decision trees, et cetera, and then, actions that trigger an outcome.

Enrico Palmerino: The difference where what we would do versus your firm is your farm would typically have a team of bookkeepers tasked with taking all that data, and just identifying that data, putting it down, and identifying, determining what that data is, and then, what workflow it should go down. We do a really good job with our AI at automatically identifying what that data is, and then putting it down the appropriate workflow. Depending on what that workflow is, then, that determines the level of human assistance required.

Blake Oliver: Let's maybe walk through an example workflow, if you wouldn't mind [cross talk]

Enrico Palmerino: Yeah, let's take bill pay.

Blake Oliver: Yeah, let's do bill pay. I'm a business owner, and I have a bill in my hand. I imagine I scan that bill, right?

Enrico Palmerino: You can use our application to take a photo of it, submit it, and then it enters the Botkeeper system. I'll do a comparison, and analysis of my firm versus your firm. In your firm's situation, someone would take a photo, attach it to an email, send it off to you, or take a photo, attach it to an email, and send that email off to the Bill.com inbox. They send it to you.

Enrico Palmerino: You would then have one of your bookkeepers key in all the appropriate, and pertinent data off of that bill into suggested, or appropriate Bill.com workflow for approvals, and payment. If you send it to the Bill.com inbox, you could choose to pay Bill.com ... I think it's like 50 cents per transaction to queue up that workflow for processing on your end. Then, Bill.com oftentimes would sync that data into QBO, though there might be an error, or something, and your team might need to jump in there, reconcile, fix the error, clear the error log in Bill.com.

Enrico Palmerino: In the Botkeeper scenario, you would send that input ... You would use our mobile app, or web application to upload, or submit that bill. Alternatively, you could send that bill to a bot name at Botkeeper.com, which would receive that bill. That bill then runs OCR software against it to extract the appropriate fields off of that bill, and the characters, and information associated.

Enrico Palmerino: Those go into a database that then runs against our machine learning to say, "What is this, and why, or how should we care, and what should we do with it?" The answers that come out of it are, "This is a bill from such-and-such company, so match it to the vendor. If the vendor is in Bill.com, match it to the vendor that is currently in Bill.com if it has the same name. If the vendor is not in Bill.com, create that vendor on behalf of the client ..." [cross talk]

Blake Oliver: Specific to that example, matching the vendor, how fuzzy is that? Is that a strict match, or, are you saying, "If it's close enough, we'll match it"?

Enrico Palmerino: The machine learning, and the AI only know to do things that they're like pretty much exactly ... The only time we will allow it to do 100 percent of the workflow is when it 100-percent matches. Let's say the vendor bill comes in, and the vendor name currently in Bill.com is ABC Company. The bill comes in, and says ABC.LLC. Then, our tool would say, "This is either a new vendor ... It looks like this other vendor, but we're not sure if this is the same vendor," so it gets flagged for review by one of our staff in The Philippines, who then says, "Yep, I was able to identify that they share enough of the same characteristics; this is in fact the same vendor. Let me correct the name in Bill.com to be the actual accurate name of the vendor moving forward.

Blake Oliver: Okay, that's fair enough. Then, in terms of the ... I imagine that after you've extracted the OCR data, you're also picking the expense account in Bill.com, or the ...

Enrico Palmerino: Yep.

Blake Oliver: Now, and of course, [cross talk]

Enrico Palmerino: -the way our machine learning works is it picks an expense account from QuickBooks. I'd like to assume that the expense accounts are properly mirrored in Bill.com. That's not always the case, so the way that our tool runs is it basically looks for ... We're doing real-time syncing with the chart of accounts that's in QuickBooks. It runs against that. It picks the appropriate categorization class/subclass, so on, and so forth.

Blake Oliver: Right.

Enrico Palmerino: It looks in Bill.com to see if that's there. If it's there, it uses it. If it's not there, it attempts to create a new one, but once again, here's where the human assistance comes in, because what we don't want to do is create a whole bunch of similar fields if. for whatever reason. there is an old synced field that we just could update instead.

Blake Oliver: You're looking into QuickBooks for the history, right?

Enrico Palmerino: Yep.

Blake Oliver: That's not particularly exciting, because, I can do that, or, actually, if I'm just using Bill.com, say, as an individual, once I've coded a transaction a particular way, or invoice a particular way, next time it comes in, Bill.com's going to do that anyway, especially if I'm using their data-entry service. What about [cross talk]

Enrico Palmerino: -could pay for a data-entry service-.

Blake Oliver: What about [cross talk]

Enrico Palmerino: -find most accounting firms ... I don't know if you use their data-entry service, but pretty much-.

Blake Oliver: I have in the past. Not for a while, but I did in the past-.

Enrico Palmerino: -was that it was either cost-prohibitive, or it did one piece of the equation, and they weren't 100-percent certain that it was accurate. They might come in, and the accountant felt like they either needed to constantly correct it, or they just couldn't afford to make an important to turn it on.

Blake Oliver: Other than the transaction history, what is your system looking at to determine the coding, if anything?

Enrico Palmerino: There's two machine models that run in parallel. The first one runs against the client's history to determine key value pairs, and the mapping of those key value pairs. Basically think vendor name, some detail level of the transaction - dollar amount, date, date of invoice, other descriptive info - and says, "Let me learn, and understand where you put things like this in the past.".

Enrico Palmerino: The nuance there is Bill.com requires you to teach it every single time, and the same thing with QuickBooks. You have to do all the vendor mapping yourself. The other nuance there, in that statement, is vendor mapping, not transaction mapping. In QuickBooks, and in Bill.com, you can say, "Everything from this company should be treated in this format," which is good.

Enrico Palmerino: It works often, or many times, with some vendors, but more and more, these days, as vendors expand, or extend their product offering, and lines - they might sell a variety of tools, functions, widgets, what have you - each of them having- should have different categorizations, and classifications. What we're doing is we're basically mapping it down at the transaction level.

Enrico Palmerino: The second machine learning that runs, runs against the collective of all our clients to then say, "Have we seen something like this before? How have our clients treated this something before, and what is this something inherently?" Not, "Is this something Amazon," but, rather, "Is this something an apple, or a desk, or what have you," and then teaches the first machine learning, which understood where to map that key value pair, that, in reality, what you're mapping is this client, in the past, has actually put all apples, which are a fruit, under this category. So, in the future, when you see fruit, or apples, whether it's an Amazon apple, or another apple, put it under that same category.

Enrico Palmerino: That's an important nuance, because it basically is then dynamically creating rule sets for the client automatically, without a client or accountant having to do them, and also at a higher level of reliability, because it's not doing a broad-sweeping mapping at the vendor level, but rather looking level deeper at the transaction.

Blake Oliver: To make sure I understand, it sounds like you are OCR-ing the document, and you are pulling line-item transaction detail into your system, and then you are-

Enrico Palmerino: Yep.

Blake Oliver: Are you, say, in Bill.com, creating a bill in Bill.com with the line-item detail?

Enrico Palmerino: Yep.

Blake Oliver: Okay, and then, each line is run through the AI to determine, so, if it is apple, then it always gets categorized to the apple expense code, or whatever that should be.

Enrico Palmerino: Correct. Yep.

Blake Oliver: If it's an orange, it goes to orange.

Enrico Palmerino: Why that's important is whether you bought that ... Whether you buy in the future. Let's say currently you buy apples, and oranges from Amazon, and tomorrow, you decide you want to start buying apples, and oranges from Star Market. It's the same apples and oranges; they should get the same treatment, regardless of the vendor being different.

Enrico Palmerino: In the past, if you switched vendors, the accountant, or bookkeeper, or someone on your team would've had to say, "Oh, what is this new thing? Okay, it should be treated this way." With our system, if, across our collective base of clients, or everyone, we have seen Star Market oranges, and Star Market apples, we would automatically know it's a like-for-like swap, and here's how to do the appropriate categorization, or classification.

Enrico Palmerino: Why that matters is like, so, for one client, if you are a business, and let's say you're buying food, and your company sells software, you may ... It's going to be more logical that your food gets categorized, or classified under meals and entertainment, and that's where the first model that runs against your history is learning their nuances. That first model, if it ran against the history of a restaurant, would understand that food, because it ...

Enrico Palmerino: More, and more, it's understanding that this thing is a restaurant, but what it's really understanding, presently, is that you put food under COGS, and you're different. For whatever reason, you have categorized, and classified food purchases under COGS. I don't care why, or how, I just know that's how you've typically done them. In the future, you start buying things from Star Market, those are going to go under COGS, even if it's a different vendor.

Blake Oliver: Tell me then, what else in the bill-pay process, other than the OCR, and entry of the bill into Bill.com with that line-item detail, what else is automated, using AI, if anything?

Enrico Palmerino: I'd say, beyond that, from true AI, there isn't really ... The AI that we use is to uncover what this thing is. Then, from there, is kicking off a workflow, like, "What do we do with this thing? Now, I know that ..." The AI says, "Hey, this is a bill. This is a bill for these things, and they're categorized in this way. Because we understand it as a bill, put it into Bill.com workflow; depending on what workflow it is, put it here, and assign this approver, and get the notification off to them."

Enrico Palmerino: If that thing you sent was instead an expense, it would ... " I understand that this thing has been paid. Here's the categorizations, and classifications; this should be booked as an expense, and shouldn't be put in Bill.com." The distinction: I think a lot of people like to bucket, or like to use AI synonymously with automation. In reality, the two are totally different.

Blake Oliver: Yeah, I think that is something that people tend to miss, and machine learning, of course, is a subset of AI.

Enrico Palmerino: Correct. I would argue that we cross the bounds of AI ... The key distinction for AI is AI does situational awareness, and even, I would say, predictive actions, which, in accounting, you're not really doing a lot of predictive actions, but you're using past-, and present-day events to determine how to treat this present piece of data that came in.

Enrico Palmerino: On the AI scale, we're at the early stage, or beginning of AI; whereas, things like Watson that are trying to predict the future markets, and how to invest today to gain, and receive benefit down the road, that's a whole other level of AI that we don't play in yet.

Enrico Palmerino: We play in the 'how do we ...' How does our past interpretation of data influence the present-day actions on new data that we see that maybe we've never seen before. The client has no history with this vendor; our collective base of clients have never worked with this vendor, or bought this thing. What's the machine's best guess at what it is, and how to treat it? Then, the data-validation team comes in, where they basically either affirm that that guess is correct, and that trains up the model.

Enrico Palmerino: AI, the way it works, or, machine learning, for that matter, the way it works is it all starts with a confidence rating. The machine's never seen it before; in general, it's going to have zero confidence on its accuracy of prediction. Now, there'll be other influencers, there, or other influences, components that come into play as to whether or not it thinks, "Hey, I'm 66-percent confident that this is the answer." Then, if the human affirms, "Yep, that's correct," then the confidence level increases; maybe it goes up to like 70-something percent, or 80 percent. After enough affirmations of a similar piece of data, it reaches a really high, or what we would consider a confident rating to automatically [inaudible].

Blake Oliver: Let's make this a little more complex I'm on the accrual basis, so, how does your system know whether or not, or does it know whether or not to accrue expenses, or to defer?

Enrico Palmerino: The AI is not really doing the determination on whether to accrue or to defer. That's more of a rule set. Client may sit with us, and say, "Moving forward, any time an expense crosses this threshold ..." Sorry, "Any time an expense is s over $2,000, we want to assume to accrue it over some period of time. Over a year. Anytime we purchase software ..." See, once again, it's setting the rule set. "Any time we purchase software that's over $2,000, within a year, we want to accrue it in that year that it was purchased," then that would-

Blake Oliver: Got it.

Enrico Palmerino: -trigger an automated workflow action, not AI, which would be applying our AI, because the AI would have said, "What is this thing that was purchased, and how do I treat it," then kicking off the workflow.

Blake Oliver: Got it. What about approvals? What about the payments? That's another step of the process in a Bill.com type situation. How does that happen?

Enrico Palmerino: Similar rule set. What are your rules around how you want to structure workflows? Who's approving what bills, when? A lot of that structure, like Bill.com does a pretty good job at setting up different levels of approval-

Blake Oliver: I mean once I've set up an approval workflow for a particular vendor, it assumes the same one going forward. Is that what you're leveraging?

Enrico Palmerino: Yep.

Blake Oliver: What about payments? Is a human being clicking the pay button, or is it a machine?

Enrico Palmerino: We like to push the pay button responsibility on the client. We like to say, "Hey, look, from a control standpoint, we'll get it in here. It will have passed our level of approval for this [cross talk] business ..." so on, and so forth-

Blake Oliver: Totally makes sense-.

Enrico Palmerino: Then, "When it comes to issuing payment, you have to click the pay button."

Blake Oliver: What about payroll? You onboard/offboard employees. Is the machine learning, or artificial intelligence involved in any of that?

Enrico Palmerino: No, the AI's not doing the onboarding/offboarding of employees with the ... I'd say some machine learning might come into play on getting payroll data departmentalized, or broken out, and synced into QuickBooks by that appropriate department.

Enrico Palmerino: Depending on what payroll platform you're using, they may, or not may not have that functionality, so the way we basically do that is we might have a trigger that basically runs an export against your payroll platform, downloads the data, and then applies an appropriate rule set, and using some learning along the way, knows how to - as, potentially, your business changes, as a department grows - how to treat it, or apply it accordingly.

Blake Oliver: Accounts receivable, any AI involved in accounts receivable?

Enrico Palmerino: On the, "What is this thing. What is this item being invoiced for?" Yes. Back to the data inference, or interpretation, then, an automated workflow might be the thing that creates the invoice-

Blake Oliver: Sorry, step back; where are the invoices being created? What's the initial workflow?

Enrico Palmerino: An employee in your company might take a photo of a sales ... Not your company, sorry. Employee of a company takes a photo of a sales contract. If, in Botkeeper, we've worked with that company to say, "What are the parameters by which you do invoicing? What are the invoicing schedules? Are you doing milestone completion?"

Enrico Palmerino: There's milestone completion with the client having to give feedback on when a milestone is crossed, versus just, "Okay, we bill upfront 50 percent of the total amount. We bill 25 percent at three months, and then, maybe 25 percent at three months after." That's something we can create an automated workflow around.

Enrico Palmerino: What is the item on there? On that sales contract that you gave us, there might be multiple sales items that are being purchased, or are sold, for that matter. Then, grabbing those sales items, saying, "Where do we apply ... For this item has some level of revenue in it. Which revenue account does it go in?" is more of a mapping exercise than it is an AI exercise-.

Blake Oliver: It sounds like what you were describing earlier would also be more of a rules-based exercise than anything AI.

Enrico Palmerino: On the payroll side?

Blake Oliver: No, no ... You're saying I snap a picture ... It sounds like everything starts with your app, right? I'm a business owner. I've got a sales order. I need to create an invoice. I snap a picture of the sales order. Is that correct?

Enrico Palmerino: Yep.

Blake Oliver: Then, you get it. It comes to you guys from there, and then, is an AI looking at it first, and determining what type of transaction it is?

Enrico Palmerino: Are you saying on the bill-pay side? On the bill-pay side of the equation, on basically data inference, when it's [cross talk].

Blake Oliver: No, I'm talking about the accounts receivable side. I'm assuming that if I snap a picture, and send you guys something, I don't have to tell you what it is, do I, or do I say-

Enrico Palmerino: No.

Blake Oliver: Okay.

Enrico Palmerino: No, and a lot of that has to do with the fact that we know the things you sell. If we know the things that you sell, then we know where, and how to treat those things, based on where, and how you treated them in the past. If it's something new, the AI might say, or make a suggestion as to where we think you should treat it. We still like to try to run the AI model as often as possible against all data, even if we're still ... Even if we require ...

Enrico Palmerino: Like on the receivables side, it's going to be an unknown item that has never been seen across any of our clients before because it's the thing that you're selling, so it's going to have its own unique code associated with it. It's going to likely have a completely unique description associated with it. It's always going to start at that low confidence rating, and then, it will have to be trained up.

Enrico Palmerino: How it trains up is then someone in our team says, "Hey, we noticed this new thing that you sold here. Where, and how do you want to recognize this? Is this a service expense? Is it a product? Is it a service revenue? Product revenue? Give us some additional detail, and info. Then, based on that additional detail, and info, we train up the model accordingly. Then, the model's really starting to look for anomalies more than on the A/R side, than it is trying to actually automatically figure out where, and how to treat these new items, or services that you just started selling.

Blake Oliver: Okay. Let's talk about the meatiest part of this, which ... We don't have to go on. I know we've been talking for a while. I don't want to take up too much of your time.

Enrico Palmerino: Not at all. I hope it's helped clarify.

Blake Oliver: It is. Let's talk about the actual bank reconciliations, and categorizing of transactions in the GL. Where does your ... What exactly is Botkeeper doing, from a non-human standpoint, from an AI standpoint, when it comes to reconciling the books in QuickBooks Online?

Enrico Palmerino: If you think about how QuickBooks Online works, you're going to pull in bank, and credit card data into what goes into QuickBooks' staging area. A staging area, then, a human logs in, usually, and applies the appropriate categorization, and classification at that point. Then, you're reconciling basically the data feed that you got from the bank with all the information that you've received to make sure that there's a one-to-one match.

Enrico Palmerino: The AI's doing the categorization, and classification, so we basically pull information out of the staging area. Goes into our system, runs, comes back with the categorization, and classification, does the appropriate booking. Then, we run a secondary variance analysis.

Enrico Palmerino: You would call variance analysis machine learning. It's ... If you think about it, regression analysis is quasi machine learning. I look at it as like just statistics, but in the land of machine learning, anytime you run an algorithm, that's the piece of data that are in machine learning.

Enrico Palmerino: We run a variance analysis to basically uncover what appear to be duplicates, what are uniques that don't have an associated item with them, and then, those get flagged to then be reviewed by a human to do the final tying out, or to say, "Yep, in fact, these are duplicates," or, "No, these are uniques," and do the reconciliation.

Enrico Palmerino: The nice thing about that is it basically cuts the time down doing a reconciliation, and gives an additional spot check on it. Because everything was properly categorized, and classified, the human didn't have to do that.

Blake Oliver: Is your tool running the reconciliation report, and saving that in QuickBooks?

Enrico Palmerino: You could still run the reconciliation report in QuickBooks. We do that, frequently, still for our clients, just as like a second validation of 'this was done.' Because QuickBooks is storing it in QuickBooks, it's easier for the client to refer back to, or auditors to go, and get. We have a secondary variance analysis that's being done on our end that's basically flagging for our team, or to spend, or pay attention to, and to look at, and chase down. Then, they'll run the reconciliation report as a final tie-out to make sure that our software didn't miss anything, and to give you the final affirmative check.

Blake Oliver: Let me summarize to make sure I understand. Transactions come into that staging area we've talked about. I think it's called Banking, in QuickBooks Online, right?

Enrico Palmerino: Yep.

Blake Oliver: Transactions come in from the feed. QuickBooks already will suggest matches, and suggest categorization, so-.

Enrico Palmerino: Based on the rules that you've created, or applied in the past, yep-

Blake Oliver: Right, or the history of what you've done in the past. I believe it looks at the history for ... Even if you haven't created a specific rule, it will start to suggest them, right?

Enrico Palmerino: Yeah. It tends to be more high level, like the way ... When I look at what QuickBooks suggests, it looks a lot more like kind of what American Express will do for categorization, and classifications. Where it'll be like, "Oh, this company name looks like software. Apply it here?" Sometimes, that company sells skis, so it's not the most accurate. Whereas ours runs against a true model. It's modeled not on just matching a database matching, but actually is doing machine learning against all that client's prior transactions, plus our other clients' transactions.

Blake Oliver: Got it ... I could certainly believe that it might be more advanced than what Intuit has so far built into QuickBooks, when it comes to the expense categorization, since you can do it across multiple clients, and I'm not sure that Intuit has gotten to that point yet.

Enrico Palmerino: I'm not sure. I can't comment on Intuit's [cross talk]

Blake Oliver: The bank feed comes into this staging area. QuickBooks would suggest matches. Normally, you would then actually either create the transactions in the GL from there, categorize them, change the vendor names, whatever; or QuickBooks will suggest matches if there's a one-to-one match. It's usually pretty good at figuring out dates, and amounts, and matching those to transactions you've already got in QuickBooks.

Enrico Palmerino: Yep.

Blake Oliver: You're telling me that, instead of that, you guys are pulling in the feed into your system, into your automated system, which then goes, and does the matching, and then pushes those transactions back into QuickBooks. Is your system also, for any transactions that already exists in QuickBooks, is it creating the match right there? I'm not sure you could do that [cross talk]

Enrico Palmerino: -yeah, from the duplicate standpoint, we're not being able to say, "Hey, this, in fact, is the same thing. Let's ..." [cross talk]

Blake Oliver: Let's reconcile it.

Enrico Palmerino: -yeah, "Let's reconcile it, or not ..." but, what we can do is, when we run that end-of-month variance analysis, it says, on our end, "Hey, look, this appears to be a duplicate. Someone take a look at this one particular item." In our application, it makes very quick for our team to do a correction, or tweak it. Then, at the end, one of our accountants basically reviews everything that's been done for that client at the end of the month, pushes the reconciliation button in QuickBooks to get the affirmative - everything is reconciled, and complete - and then, gives the approval to our client, or partner that this set of books is ready for the end-of-month close.

Blake Oliver: Got it, so somebody is clicking "Okay" in QuickBooks to match any transactions that need to be matched to the bank feed. The part that is automated is transactions coming in on the bank feed that do not have a match in QuickBooks. You're creating that transaction automatically.

Enrico Palmerino: Yep, and sometimes, we've seen a lot of accountants disable the rule sets, because ... The downside with relying particularly on a one-to-one mapping rule structure is it's really good, so long as there's total consistency. The second that there's inconsistency - so, a vendor that sells a few different things - it's really good at automating error. Then, spotting, or finding that error becomes difficult, because you basically have to go through all the GL detail to uncover it, and then you have to make ... Back out the entries, and reclass them.

Enrico Palmerino: That sometimes takes more work than just putting it in manually, yourself. That's why our model is ... Most accountants aren't comfortable with using the rule set too extensively, because of that automated error. They can rely on our model, because it's going to get a wider range, and variety, and level of complexity correct.

Blake Oliver: ... An important thing to do in QuickBooks is to run the reconciliation workflow on a monthly basis. View the statement balance online, or view the PDF statement, run the reconciliation in QuickBooks, and if you've matched everything properly, QuickBooks should just tie out. The GL should reconcile โ€” transactions should tie out. Is a human being doing that at Botkeeper, or is that the machine doing it?

Enrico Palmerino: For doing the final tie-out, and clicking that button, and running it, a human's still doing that on our end. Big reason being is the whole process of looking at the PDF bank statement, and doing the final assistance, and tie-out, the way that the guidelines around doing things appropriately, and accurately are, you need to look at that, and you need to essentially manually tie it out, because there's this idea that data feeds aren't always going to be correct.

Enrico Palmerino: We do that today; we're striving to eventually get to the point where we can ingest all the data on that PDF, but we're still trying to ... We're still running various degrees of testing to say what if we ran OCR, like three, four, five, six different models against that data set, and if all five, or six said the exact same thing? Then we could confirm that we've properly captured the info on that PDF, and then compare the info that we had in that database to the info there, to be able to then automatically do that click. Right now, we're not doing that, because we just haven't gotten ... We're still building, and developing, and I think it'll take some convincing of the market that you can truly rely on software to do that piece.

Blake Oliver: If it actually generated the reconciliation report, and basically created the work paper, then I could review that, and know that- and be confident; the same way I could rely on a staff accountant, or a bookkeeper doing that sort of thing for me.

Enrico Palmerino: Yep, but there's a ... As you know, there's a lot of nuances there; hence, why we still have a lot of people in the accounting industry performing those tasks.

Blake Oliver: Yeah, well, nobody's figured out how to do that yet, as far as I know, at the small-business level.

Enrico Palmerino: Yep [cross talk] single large business, now you're introducing consistency. It's easier to build robotic-process automation to do that, because it's one entity doing the same thing a million times.

Blake Oliver: I'm getting the impression that the ... I'm very glad that we had this conversation, because I'm getting a much clearer idea of what exactly is using machine learning, and I'm going to call it machine learning if that's all right, because it is. Specifically, that's the type of AI, right?

Enrico Palmerino: Yeah, or I'd say machine learning is a level on the AI spectrum, and you start getting into where ... Some of what the machine learning that we're doing would be early AI, technically, if you look at the AI spectrum, but whether you call it machine learning, or AI [cross talk]

Blake Oliver: All right. Well, then, I'll just call it AI, because that's easier. It sounds like the AI that is in use currently for Botkeeper is basically paired with OCR, in that you are able [cross talk]

Enrico Palmerino: -automated workflows, like decision trees, and robotic-process automation.

Blake Oliver: Right. Those are separate, right? Rules-based bookkeeping is very different than AI bookkeeping.

Enrico Palmerino: Correct.

Blake Oliver: I can go, and I can create a bunch of Zappier connections, and I can create automations. I could even probably, at this point, if I wanted to, spend a lot of money on RPA software from one of these vendors - UiPath, or Blue Prism, or whatnot - and I could try to figure out how to do it myself, right?

Enrico Palmerino: Yeah, it's just very costly, and very time-consuming, and it gets hard to actually work across many companies. That's why- that's where the AI comes into effect. AI does a really good job at solving a piece of the equation. Then, once that piece is solved, then using other tools, or technologies for automating to pick up the next leg.

Blake Oliver: Just to summarize, again, my understanding is that the AI in use is ... You're doing OCR on documents that are submitted to you, via your app, or via email, or whatnot; they go into an inbox. Your AI ... You're OCR-ing those docs, pulling out all the text fields. Then, the AI looks at those text fields, and tries to figure out what kind of field this is - what is this data? Is it a date? Is it an amount? Is it a description? All that good stuff. So far, am I right?

Enrico Palmerino: Yep, and then, the proper categorization, and classification of that data.

Blake Oliver: Okay, and that's based on a mix of the individual company transaction history, the transaction history across all of the Botkeeper clients, and-

Enrico Palmerino: Yep.

Blake Oliver: Is there anything, like if it's a de novo transaction, does your AI try to figure out what it is, or does that get kicked out to a human?

Enrico Palmerino: Say that again? What [cross talk]

Blake Oliver: Sorry, if it's brand new; if it's never been seen before, or maybe it's only been seen in like one or two other instances, does your AI try to go search Google, and figure out what it is, or does it ...?

Enrico Palmerino: AI doesn't search Google. That would be like doing a web-scraping process to then ingest in a database, to then run AI against a database. What the AI would do is it would look at this new transaction from a never-before-seen vendor, like an unknown unknown. It would say, "This looks similar to this.".

Enrico Palmerino: If you go back to, truly, what AI is, think about it as like AI is an algorithm that maps a line. All an algorithm is, is an equation. It's basically doing a mapping of correlated data. What is the relationship of this data? What the AI is doing, in that sense, is it's saying, "I've never seen this vendor ..." This is the difference between AI, versus just database mapping, versus just one vendor mapping ... Saying, "I've never seen this unknown unknown before, but it has the characteristics of these other things I have seen, and therefore, I think it's this." Then, the human comes in, and affirms or denies the accuracy of the AI's assumption.

Blake Oliver: The database that it's leveraging right now is limited to the transaction history of the Botkeeper clients, the QuickBooks stuff, or is there anything else that it's using?

Enrico Palmerino: It's the transaction history of all Botkeeper clients, and then, we also grab ... We've done other database imports, and stuff, of different vendors, and transactions, and information that we've used to do testing, and also train up the model.

Blake Oliver: Where did the AI come from? Did you develop this internally, or is this something that you have purchased?

Enrico Palmerino: The AI is developed internally. We have roughly a 30-person dev team that's built this out. We still do things like fringe every AI shop, which is leveraging AI to almost to help, or assist in fitting AI models. There are companies out there, like DataRobot, or Algorithmia, that do a great job at taking the machine learning that you built, and structure it, looking at different data sets, and then, determining whether or not - because usually, we build multiple models - which model, under what circumstances, and what situations is going to be the best fit for that data, so that, then, you basically build a second level of AI that says, "Hey, if I see something that looks like this, run it down this model to accurately predict it.".

Enrico Palmerino: Yes, the AI that we've built is our own AI. We leverage Amazon to do a lot of our computing. We've also used Scikit, which is like a python library, and looked at other machine-learning processing tools, like SageMaker, and others, to do some of the ... Where we build the models and then, be able to leverage that platform to do the computing, and processing.

Blake Oliver: What are you using to do the OCR? Is that something you've built?

Enrico Palmerino: No. The way we look at OCR is OCR ... There's a variety of OCR out there. The tools, for the most part, like the actual OCR, I think, is kind of similar in terms of accuracy from one tool to the next. We have a nice partnership with AutoEntry to assist in doing the character recognition, and extraction off the documents. Then, we run it against our models to do as I mentioned.

Blake Oliver: Got it. You do AutoEntry. You send the docs AutoEntry; it pulls out the information, and then sends that to you guys, to your model.

Enrico Palmerino: Yep.

Blake Oliver: All right. I asked a question, where, and this may be difficult to answer, but I said what percentage of the work you do for clients is human, versus machine, at this point right? You said, at the low end, it could start at 30-35 percent, and then trend upward toward 100 percent machine, as, of course, I assume, as you train the AI, and the RPA.

Enrico Palmerino: Kind of the way to think of it is there's a few variables that come into play here. One is that client's individual, and particular history. How long has this company been in business? How many transactions did they process on a monthly basis? That gives a larger data set to learn.

Enrico Palmerino: A company that just signs up with Botkeeper today that processes a few transactions a month, and if the transactions vary drastically from month to month ... Let's just say they're an antique shop, and they sell a chair today, and a lamp tomorrow, and a glass the next day. It's going to constantly be like learning new things.

Enrico Palmerino: The model is going to train up as best it can, based on the collective of info that it's seeing on that particular client, plus all the other clients that we've operated in the other transaction library that we've built.

Enrico Palmerino: A company that comes to us with a year's worth of transaction history - it processes, call it, 50-100 transactions a month - that's going to be a much richer data set. It's going to allow the models to learn a lot faster. It's also going to allow our human team to train the model a lot quicker, because there's going to be more interactions with data to be trained up against.

Enrico Palmerino: The way to think of it is there's the longevity of your business, the transaction volume that you're processing, and then the duration that you've been on Botkeeper that will also determine how much we automate. Then, a fourth component, which is the current count of clients, and the current total transactions that we process.

Enrico Palmerino: Almost inherently, a company that ... Let's say if you were to go back like two years, we may have started out 15-20-25 percent out of the baseline, being totally automated. These days, I'd say we're probably closer to at least 30-35 percent being totally automated, off your first day, and then it trending up pretty quick to that.

Enrico Palmerino: Usually, getting to 100 percent's really difficult, unless the scope of what we're doing is limited. If you're having us send weird, complicated invoices on your behalf, we're probably never going to get to 100 percent, because there's always going to be some level of human involved, but layering our little bit of human is way more cost-effective than you trying to staff a human, or contract a human, otherwise.

Enrico Palmerino: If you were saying, "Hey, Botkeeper, I want you to do all the categorization, classification, reconciliations, and processing for me, especially if you're a cash business, we can get to pretty much 100 percent or very close to that point in little or no time. If you are accrual basis, it's going to take longer, but if we're at automating over 90 percent of all the workload, or 90-something percent of all the workload, then I think we're certainly doing our job.

Blake Oliver: Speaking of accruals, the monthly accruals, deferrals, journal entries, and all that, that's still being done by a human accountant, right?

Enrico Palmerino: Correct, yep.

Blake Oliver: Any work papers that need to be done to reconcile those accounts also would have to be done by a human.

Enrico Palmerino: Yeah, and that's why Botkeeper's a great complement. We don't really look at ourselves as a competitor to other accounting firms. We think of ourselves as a partner, or a complement. Use us to do most of the day-to-day, mundane tasks, and grunt work. Let your team of accountants do the more skilled critical-thinking tasks, such as making those custom end-of-month journal entries, putting together work papers, doing the critical thinking, reviewing, advising, et cetera. We at least removed a really good chunk of their time that was spent doing manual processing.

Blake Oliver: I think that is all the questions that I have. Any questions for me, or anything you'd like to add?

Enrico Palmerino: No. I'm really excited about where Botkeeper is today. Investors are really happy with the software that we've built, and where it's heading. We are fortunate to get Google as an investor, and working with them, and understanding the AI ... I'm sure you've seen some of the Google illustrations, but understanding areas that we can improve, like models that we can do better, [inaudible] data piping, and workflow is a big thing. I'm really excited over where we're going to be over the next year, now, with this new round of capital, some really smart people assisting, and advising us.

Blake Oliver: That's great. One thing that I'm still trying to understand is I understand that the employee was disgruntled, but why would they go, and say that the AI, or the tool doesn't work?

Enrico Palmerino: This is a good, classic example of when you have an organization of our size, not every employee interacts with every aspect of Botkeeper. If you're the human that is doing- it tasked with, say, onboarding clients, or learning, or understanding all the nuances of that client, and then you hand it off, and your job is solely to literally understand everything that goes on with your business - the nuances, the workflows, the rules sets that we should be applying.

Enrico Palmerino: Then, after I learn that, I hand it off to our team in The Philippines, which is, like I said, dev team Philippines, data validators, administrators, so on, and so forth. Then, it's up to them to then take it, and build the automated rule sets, workflows, and run the [inaudible] against it. It could be perceived that, "Hey, I'm just sending all this workload to The Philippines, and then they're doing it." That's not actually a false statement. You are sending work to The Philippines, and then it's getting done. How it gets done is the more nuanced answer, or question, which is, we use machine learning, and AI to do this component of it. Then, the workflow happens, and a human reviews, and validates it.

Enrico Palmerino: Either someone who didn't fully understand all the workflow, or didn't interact, or specifically leverage the AI themselves could interpret it incorrectly; or someone who's purely frustrated that they were getting terminated. They felt like they put in a lot of work, and were responsible for helping get Botkeeper to where it is may want to take Botkeeper down with their departure.

Enrico Palmerino: Then, the other part of the equation is that we internally build different tools, or applications to try to facilitate the humans in our equation - the jobs, and roles- the jobs, and work that they're doing. That's our testing ground. If you think about it, naturally, rather than build software first, and test it on clients, we build software internal, test it on our team, get their feedback.

Enrico Palmerino: That oftentimes means the first interaction with new tools, or feature sets that we build that our team has is it's buggy, or it still needs work, or improvement. Then, we get their feedback, and we build, and iterate a fix, and then release to our clients. They could perceive that as things that we're building aren't working, or what have you.

Enrico Palmerino: I think, for this particular individual, they just were really frustrated that they weren't going to be here anymore, and like other employees who were let go in that situation, went, and tried to do as much damage ... Make false, malicious claims, violate employment contracts, and so forth.

Blake Oliver: Makes sense. One question I forgot to ask is: in your app, your clients, do they have the ability to speak to Botkeeper - communicate back and forth?

Enrico Palmerino: Yeah, they can. There are two ways you can communicate. The way we like to distinguish it is there is the bot, which does have human assistance validating, and verifying; send most requests there for processing, but if you need to get ... You're like, "I need to speak to someone. I want someone to email me back. I want to get on the phone with someone to talk about something or get ... I have a suggestion," you could email our human@botkeeper[.com] [email address], and that basically routes directly to one of the US accountants that we have reviewing the different client accounts to basically jump in, and say, "Okay, let's have a call, get on the phone, talk about this question that you might have."

Blake Oliver: You are using a chat bot; that is one of the options.

Enrico Palmerino: We have limited chat bot functionality. We do have the ability to send chat. We aren't doing a ton with natural language processing right now. We do a little bit. If you request a report, something, that's a very obvious request. It can serve it up.

Enrico Palmerino: The downside of chat bots ... Chat bots work great in most situations, and scenarios, where you're basically trying to serve up, "Hey, I think you're asking about this. Am I correct?" or "Was this helpful or not?" Then you say yes or no, and it would either serve something else up, or route you to a human.

Enrico Palmerino: With accounting, the downside is that we basically leverage natural language processing to do the interpretation of the requested action, and then we use AI to take the thing in that requested action, and try to identify what it was, and then, put it down the appropriate workflow. You introduce basically two levels of potential error, and they combine, and they kind of like multiplier effect ...

Enrico Palmerino: We've taken the approach that we don't want to use a lot of natural language processing yet; we're working towards it, but let's have the interaction on email, or communication be predominantly human-driven, where you're at least communicating, so you're getting the responses that you want in a logical format, and we're not potentially multiplying error, or introducing frustration.

Blake Oliver: How many clients do you guys have?

Enrico Palmerino: I want to say it's somewhere around 1,300 or so. Might be more.

Blake Oliver: How many accountants do you have working for you?

Enrico Palmerino: We have-.

Blake Oliver: Accountants and/or bookkeepers.

Enrico Palmerino: We have ... I think it's like 23 accountants.

Blake Oliver: You've got 23 accountants for 1,300 clients. That means that, on average, each one is handling 56-57 customers?

Enrico Palmerino: Yeah, that's about right. What's crazy is we used to actually have ... If you went back six ... This kind of just goes to show you the rate at which we're developing, and increasing automation. You look back six, or eight months ago, I think we even had 33 accountants supporting probably half as many clients. We're able to continue to very rapidly increase the ... We've basically moved those people out of accounting roles into other roles in our organization. We've been able to rapidly increase the amount of workload that that accountant is able to oversee, manage, review, et cetera.

Enrico Palmerino: The other thing to be noted is the size, and range of our clients. Smallest client might be an Uber driver. Largest client, Fortune 5000 company that replaced a team of accountants with Botkeeper. It's a very wide range. It would be hard to say you support 50, or 60 really small companies with one accountant. It's more like there are some companies that have replaced six to eight-plus people with Botkeeper. In that regard, it's ...

Enrico Palmerino: We're basically able to get a lot of leverage on an accountant, and that's the same sort of dream, and solution that we sell to our accounting partners is: let Botkeeper be the last bookkeeper you hire, and allow your firm to scale, and grow, and increase margins, et cetera.

Blake Oliver: How do you feel about when a company does that, when they hire you guys, and they terminate six to eight people that were previously working in accounting?

Enrico Palmerino: Oftentimes, the terminations don't happen ... It's not like they bring us in, and they say, "Everyone's out." More likely than not, they've lost an accountant, or two; they bring us in to replace that role. I think the turnover in bookkeeping is really high. It's not typically accountants that we're replacing, but more bookkeepers. Because there's really high turnover, and I think ... I'm sure you've heard this a million times; this is like the number-one problem in the accounting industry is that hiring, and retention of good bookkeepers is so difficult.

Enrico Palmerino: We'll come in, maybe replace ... If you had lost one bookkeeper, you probably weren't on the market to solve your bookkeeping dilemma, but if you lose two, maybe three in a short period of time, you're now Google searching, "How do I automate this and not have to deal with this headache," because it's really painful right now. Those tend to be the companies that come, and bring us in. Then, as they grow, they either don't hire additional bookkeepers; as a attrition bookkeepers over time, they backfill them with Botkeeper.

Enrico Palmerino: There was a study that came out that basically showed the number of accounting firms ... Here's the number of accounting firms today ... The new accountants that are graduating, there's not even enough of them that are going into bookkeeping to take over those firms, let alone backfill all the other people that were operating them. I look at it as we're trying to meet the supply-and-demand curve, more than we are replacing the supply of accountants.

Blake Oliver: Have you followed the QuickBooks Live Intuit news?

Enrico Palmerino: I've heard, yes ...

Blake Oliver: What do you think?

Enrico Palmerino: It's causing a lot of rumblings I think among the accounting space. I don't know. I look at it as ... I'm kind of personally ... Well, at one end of the spectrum, what QuickBooks did with QuickBooks Live is exactly what they did with TurboTax. You'd kind of say you could see it coming.

Enrico Palmerino: At the same end of the ... At the other side of the equation, TurboTax didn't rely on accountants to go, and bring clients on to a platform, so, the fact that Intuit said, "Hey, here we are. We're supporting the accountant. We're here to make your life better," and then they released a competing product? I could totally see why a lot of people are really frustrated.

Enrico Palmerino: I don't think, from my understanding of the product, it's designed for a really, really small business owner to do really basic tasks. I don't see it replacing the relationship, and the accounting value that an accountant offers.

Enrico Palmerino: Personally, I can't think of any businesses I know, firsthand, that do their own taxes on TurboTax. I feel like the company that does their own taxes on TurboTax probably would take advantage of this QuickBooks Live offer, and that's probably not the client that ... It's not the client that we tend to typically support, nor is that the client that our accounting partners tend to typically support, so, I don't know. It's shocking. We'll see what happens. Intuit's going to do what they want.

Blake Oliver: Do you view them as a ... Do you view that service as a potential competitor to Botkeeper?

Enrico Palmerino: Not really. I think the service does ... There's certainly some level of overlap there, but I think the individual or company that uses that service is really ... They're looking for a real basic level of accounting. It's kind of like saying would we consider Bench a competitor of Botkeeper? Not at all. Bench has their market, and focus, which is the S in SMB, and cash-basis-only. This QuickBooks Live offer, I think, is geared to a very simplistic business-.

Blake Oliver: Cash basis, as well.

Enrico Palmerino: Yeah, exactly. Whereas most of our clients are accrual ... Your business will be cash basis up to a certain point, usually, if you're trying to grow, and scale; you're actually [cross talk]

Blake Oliver: I know there's a line for taxes, but where do you draw that line? When do you think a business should switch to accrual?

Enrico Palmerino: I think it depends on the business model. There's a requirement around the revenue, where you eventually have to make the switch, but, from a business-model standpoint, there are some businesses where cash basis does just fine work for them. I personally am a big fan of accrual. I just think if you're going to ... Whether or not cash basis does, is fine. I think you'd benefit from accrual, and then use the cash report to give you the cash-basis understanding, and what your cash flows look like, because accrual just really lets you plan your business, and see ... Massage out some of the ebbs and flows-

Blake Oliver: I have to say ... Sorry to interrupt, but in listening to you, and talking over the last hour or so, and knowing everything I know about QuickBooks Live over the last week or two, it seems to me like if Intuit wanted to build QuickBooks Live, and do it right, they should buy you guys.

Enrico Palmerino: Maybe. I'd say it's not a far-fetched idea. I don't think we're necessarily trying to go on the market just to sell the company, right now. I think we're really happy with how we're growing. We also ... I think there's a lot to be said for ... you know we. We didn't want to ...

Enrico Palmerino: There are firms that'll pop up that are trying to do automated accounting, and their goal is to take clients away from accounting firms, and accountants. We just saw that as, like, that's an uphill battle. Everyone's going to be rooting against you. Your goal is to basically put all the other companies, and firms, and accountants out of business.

Enrico Palmerino: What if we were like the, call it ... I'd like to think of us like the knight in shining armor that comes in, and says, "Hey, plug us into your platform, or your firm, and let us be ... Let us allow you to plug and play AI into your practice."

Enrico Palmerino: For us, it goes from being a land grab of clients to just a landmark. We're just making it that much harder, or difficult for another company to extract a client from our partners. We're allowing them to realize better margins, grow faster, and deliver a better service.

Enrico Palmerino: I think we would strongly ... If someone were to buy us to try to put everyone out of business, it doesn't really mesh with our values. If someone were to buy us to try, and continue our vision of being an underlying platform that ideally has many accounting firms, and accountants use to do more work, and grow their firm, then I'd consider it.

Enrico Palmerino: I'd say that would be my answer. I'd have to understand what, and how, and the reason, and the logic, and how this was intended to be sold. If Intuit said, "Hey, we'd love to buy Botkeeper, and we're going to license it solely to accounting firms to use on their clients," I'd say, "All right, that sounds a lot more in line with our goal, and vision." If they were going to buy it to put all of the other accountants out of business, I'd say, "This probably isn't a good fit."

Blake Oliver: All right, I'm going to go ahead, and hit 'stop recording,' now. Thank you so much for your time, Enrico.

Enrico Palmerino: Don't mention it.

Sonix is the best online audio transcription software in 2019.

The above audio transcript of "Blake Oliver interviews Botkeeper Founder and CEO Enrico Palmerino on February 18, 2019" was transcribed by the best audio transcription service called Sonix. If you have to convert audio to text in 2019, then you should try Sonix. Transcribing audio files is painful. Sonix makes it fast, easy, and affordable. I love using Sonix to transcribe my audio files.