Rishi Srivastava: AI Powered Construction Accounts Payable at Beiing Human

Hugh Seaton: [00:00:00] Welcome to Constructed Futures. I’m Hugh Seton. Today I’m here with Rishi Srivastava, co-founder and CEO of Being Human. Rishi, welcome to the podcast. Thank you, 

Rishi Srivastava: Hugh. I’m excited to be here with you. 

Hugh Seaton: Yeah, me too. So let’s do what we always do.

Let’s kick it off with what being human is and what you guys do.

Rishi Srivastava: We are a document automation solution focused on construction accounts payable process specifically in the mid-market of construction companies. That’s where a lot of time has been spent in manual AP processes. With the AI technology out there specifically document AI that we use we are able to save construction companies 80% of time and attention required in AP process.

Hugh Seaton: Let’s talk a little bit about that problem. Obviously most people listening to this have at some point. Had to process, [00:01:00] a lot of documents for whether, even if it’s just for their own expenses presumably a lot more than that. So everybody feels the pain of having to go through and find the right numbers and line them up and so on.

What sorts of things specifically and what kind of got you here?

Rishi Srivastava: I was working as an AI engineer at Bank of America a few years ago, and I saw. Big problem with document automation. We were spending so much time looking at various documents, copying, pasting, typing a lot of attention and time was wasted there.

And I realized that I’m an AI engineer at the at Bank of America and we are still doing this process. And I was. Intuitively certain that other industries specifically industry like con construction, that this problem is going to be there. And then I met my co-founder who was a construction CFOA few years ago, and we decided to go after this problem in aps for construction companies.

Hugh Seaton: So what have you found when you go out [00:02:00] there and look at what’s, obviously as a CFO, your co-founder really understood the problem itself, but talk to me a little bit about how, you go into a company and you say what are you doing now? So 

Rishi Srivastava: the companies that we go to, most of the time, the process that we are automating is completely paper based.

An AP invoice comes in an AP clerk usually takes a look at it. Let’s say the invoice is coming via email, they go and print it, and send it out for approval and get stamped on it from pm and once the approval is done, they usually come back to, let’s say their accounting system.

Which can be any of the MB accounting system, maybe foundation Delta C or QB QBD or QBO, others as well. And then they type that information in the entire process of managing this paper invoices along with the complication with approval. [00:03:00] That’s what we are here and automating.

Hugh Seaton: So if I can restate this. Somebody gets an invoice or some document, but let’s just sort stick with invoices. They get it digitally, presumably as A PDF, but it doesn’t matter. They get it digitally. They then print it out. They then route it where somebody physically stamps signs, somehow physically indicates that, yes, this is good, bad, or indifferent.

It then goes back as a piece of paper. It goes back to accounting. Who then manually either types or I guess sometimes scans, but it sounds like since there’s no way to do anything with it, they have to type in the details and then they may again scan it in for their records. Is that a good description of the process you are walking into?

Rishi Srivastava: Yeah. Yeah. And there’s besides AP invoices, the other documents that are there, of course, that go in support of it, but yeah, AP invoice is the main one. Yeah. 

Hugh Seaton: Yeah I get it. And I just, invoices are so easy for people to understand. So what goes wrong with that? I can imagine it, but in what have you seeing what have you guys found like, goes wrong with [00:04:00] that?

Rishi Srivastava: Several different things can go wrong with it. First off let’s start with the approval process. So let’s say I am a project manager and I am proving this invoice I can say that this this invoice looks good and go and pay pay this person. I’m looking at a piece of paper, but I don’t really know if this invoice got coded correctly.

Was it coded to the right job? Even though I approved for the payment. So that coding not getting right into the system is, can be a big problem. Besides that let’s say I’m an accountable clerk and I am typing this information in into one of these ERPs. The opinion is data.

I can make mistake, right? I’m a human. I’m not perfect. So let’s say I get a due date wrong, right? And if you get a due date wrong, you can have problem with getting discounts. For an invoice, you may not be able to get a discount. And other problems besides that is [00:05:00] let’s say something gets recorded wrong in your accounting system, right?

And it may or may not get noticed and, but let’s say couple months from today the pm the project manager comes and says, Hey I wonder what that cost was. Now as a CFO or controller, you go and dig into your file cabinet and find that is and you look at it and you, the PM sees it and you as A CFO also see it.

And together you determine oh, it went to the wrong job. Fixing that error towards the very end, maybe a few months after it was made is also a big pain and it’s a very manual process. 

Hugh Seaton: Yeah, that makes a lot of sense. There’s some things in here that I wanna dig out. One of them is the coding thing’s right.

Both in terms of the cost code, but also, recording the right date. One of the things that’s really, I wouldn’t say unique, but definitely special about construction accounting is how complex it can be, right? Because you’ve got things that are, you get it done early, you get it done at this point, you get paid [00:06:00] x, so on and so forth.

The dates and coding and all that really matter, and I feel like when people look at a document. We have a tendency to focus on one or two things and maybe miss some of the peripheral things or some of the secondary things. So people are gonna, are like, think about the pm to your point, their number one concern is gonna be price.

Did I, is this the right amount? Is this what we agreed to? Is this what we should be paying? They may or may not have the time to really look at the date, or maybe they just assume the date or so on. If you do a hundred of these. How, what percentage of them are gonna have little things missed that have an impact?

So that’s really interesting is that to what you’re saying is, or at least what I’m hearing, is, to make the complexity of construction, billing, and finance work, the numbers have to be right, not just the quantity, the, not just the amounts of money and items, but also the timing and the coding and all that stuff.

And as soon as you start to introduce errors. You have all sorts of cleanup on the backend, right? So it’s not just how fast it takes to get this routed, which I’m assuming [00:07:00] we’ll get to in a sec, but it’s also all of the cleanup at the end that you’re not, you don’t have to do. Does that sound right?

Rishi Srivastava: Yeah. That’s exactly what it is. And one thing specifically we do is we make sure that the AP and the pm, they have the same visibility. They’re looking at the same invoice and extract information. So besides the price, they can. The PM can help even with the coding, just take a look if it’s the right coding.

Hugh Seaton: That’s a really good segue to maybe I’ll ask you to just introduce what the product does. ’cause we started with why you did it and what the problem is. And I wanted to dig in there, ’cause it’s nice to just illustrate how messy this is right now. But what do you guys do? How do you approach this from a product standpoint?

Rishi Srivastava: From a product standpoint, The biggest problem, is converting these unstructured documents, let’s say noises into a structured piece of information that these ERPs can consume and doing it with a high accuracy, right? And this is where we do not go with LLMs. [00:08:00] ’cause LLMs, the big issue there is that when you extract the information out of these invoices or other documents, they don’t get give you the location where this information is coming from.

And as a product designer for the solution, we knew that humans validation of this AI’s work is super critical. Even if our accuracy is in high nineties, we still want human to be able to validate the AI’s work. So that was unstructured to structured information con conversion.

That’s one main part we’re doing. Besides that, oh, there are various other things. Like one of them is coding we talked about, right? Approvals how the the invoice is gonna move from AP test to PM’s test. And at the very end of it, it’s also important that we are able to index these documents carefully in the ERP.

So when, let’s say a PM comes in and they wanna dig into their job cost coding. And like how different jobs performed, they [00:09:00] should be able to dig into all the way to these documents, through the ERP itself. They don’t have to go and find another system like ours to go and find that kind of document, right?

We wanna be able to so tightly integrate with the ERP that the end user it doesn’t, it seems like we don’t even exist. 

Hugh Seaton: I love that and I love the humility of saying, I, I don’t need to, everyone to be singing my praises. We just wanna work. We wanna make your life easier. Remove some of the errors and remove some of the annoy.

It’s, nobody wants to be typing data in nobody. So the fact that you can replace that, I think is huge. Yep. Rishi, I wanna dig in on the AI a little bit ’cause I really like some of what you said and I think it helps people to feel comfortable with a product like yours to know that.

Reliability, but also verifiability, if that’s a, if that’s a word. Let’s talk about those two things right now. So what do you do? So you don’t use large language models and I really like that because one of the problems with LLMs is they’re based on probability. So even if you didn’t [00:10:00] have the location.

Kind of issue to make you wanna look elsewhere. You also have the, how do we corral it and how do we make sure that it’s inherently probabilistic approach doesn’t create, issues. Now I, there are ways to do that. It’s just, you gotta put a lot of effort into it. What do you do? 

Rishi Srivastava: So we use this model called Doc Farmer, and that’s the base of what we do is.

Still probabilistic, but it’s got three modalities vision, which is how these in invoices look and then layout. So a lot of times one important piece with these documents is like encoding that the field, let’s say vendor name and the value correspondent to that field is gonna be close enough so that the second modality is the layout.

And the third one is language. All three together are part of this solution and. That’s how the accuracies can become so much better than a large language model. Besides the, the data extraction, the verifiability [00:11:00] part you said, right? Even if you know the liability is high, let’s say high accuracy, high nineties, how does the user verify that?

That’s where we do something called ensuring that the user. Is able to see the AI confidence. So these confidence numbers, we encode it in the human, in the loop screen we have. So let’s say the AI has high confidence, 99%. We code that as a green line, between the extracted information and the place in the invoice, and that user more relaxed.

Okay, this is high confidence stuff, but there’s also occasions when the AI is gonna have trouble having high confidence. And then we make it apparent to the user one major thing we are after is, people’s attention are limited. We wanna make sure that it’s properly utilized when people are using our platform and the confidence and these green, yellow, and red indicators, they help our users verify as work.

Hugh Seaton: I love that and people are really used to that good, really good choice. I other software has done that too. Not in your space, [00:12:00] but across construction tech. Red, yellow, green is just such a good way to go. So people, they don’t have to think hard about it.

It’s okay, good. I can skip through this. Focus on the red if I need to. So are they able to see. Your output and the original document next to each other so they can just eyeball it and say, look, I’m gonna go to the ones that are red and just eyeball it myself. Is that what, how it works?

Rishi Srivastava: Yeah. Yeah. So you have this AP voice maybe three fourth of the screen and one fourth of the screen is the extracted information, and they are side by side. At the same time, we are also connecting the extracted information, let’s say vendor name. To the exact place in the document. So you have a line helping you guide the extracted information to the actual place in the document.

Hugh Seaton: I love that the ability for folks not, and the point here is they don’t have to do that. You give them the ability, if they wanna eyeball it and check it themselves. ’cause it is numbers, it is, finance. But they don’t have to, they right. They don’t have to review all this. You give them the opportunity to, so they could verify it.

[00:13:00] But if they’re, if they’ve gotten to a place where they’re trusting you and so on, or maybe it’s just the same document type over and over again, ’cause of the same two or three customers, they don’t have to go and review everything by hand or do they have to, 

Rishi Srivastava: So they don’t have to, they can mark everything good, but like the high impact items, like the total dollars, maybe the vendor name, those things.

We definitely want people to be reviewing. Carefully, even if they are as confidential, as very high, but like this lower impact fields that people can be usually relaxed in verifying, at the end of the day, it’s a statistical model and I don’t wanna seem like saying that it’s gonna be right, always right.

That’s this like too much pressure on the AI and it doesn’t work like that. We humans make mistake too, right? For us. The AI model’s performance is the, like the data extraction of these, let’s say, documents. The thing that we are aiming for is human level performance or better, if humans.

96, 90 7% of these we wanna be there or better, but it’s very hard [00:14:00] to beat at human at this kind of task where, our system, we are very good at extracting unstructured information from unstructured, right? So the AI just has to be at least there or better. 

Hugh Seaton: That makes a lot of sense. And it speaks to the fact that a lot of, you know, finance folks aren’t, are gonna wanna see it with their own eyes anyway.

Especially the, to your point, the high impact numbers and so on. So just switching gears, you mentioned that you integrate with ERPs. Talk a little bit about how that works and what, how you see your role in kind of the flow of work from, invoice comes in and so on.

Rishi Srivastava: Yeah. Yeah. Type integration with the ERPs is critical because, we don’t want to be sitting on the side and not part of the natural workflow, so let’s say an ERP let’s focus on one, maybe we call cut out foundation. We look at what’s information is there, for code coding purposes, we need like things like vendor name.

We needed jobs and, gls, all of that stuff is already there in the ERP. We pulled that in [00:15:00] our system, and the invoice comes in and they, let’s say the data gets extracted out of the invoice. And now we present the extracted data along with the coding information to the user and the choices of coding, they are coming from the ERP directly.

And the user reviews it and they’re okay with it. The data extracted with AI as well as the coding, and maybe they pass it down for approval and then it even invoice eventually gets approved. And at the very back end of it, we also push the information back to the ERP, and once the information is pushed to the ERP, let’s say the AP invoice data, go and look at the EP and look at all their AP invoice and make sure that we sync it back.

So the complete and transparent communication in our system, what’s there in the ERP and what we are pushing, that’s very important for us. And without a seamless integration I don’t feel like this product can be as useful as it has been. 

Hugh Seaton: That’s great. So to recap what people get with with being [00:16:00] human is.

They get an a, a rapid automation of something that nobody wants to do, which is managing these pieces of paper and extracting the information. Then they get a really thoughtful user interface where they can go and just review it themselves and verify.

That everything is the way it should be, but you skipped that enormously painful step. So you’re going straight to the part where they can quickly eyeball something that might have taken hours to do, now takes minutes and they can make sure everything’s perfect and accurate, and you’re eyeballing things where, or you’re giving them indications where they should look.

So it’s not, of course they’re gonna look at every line, but you’re having them first go to places where the system felt like, for whatever reason, because the way that. The, I don’t know the document worked. That’s often a big deal. People underestimate how variable these documents can be, so you’re helping to draw their attention to where to look first, but they can look at everything.

Then once that’s done, they’re able to then go and digitally route and digitally deal with things in a highly integrated way. Does that sound like a good [00:17:00] description? Yeah. Yeah, that’s exactly what it is. Thank you. And what is the adoption process typically like? So someone listens to this and says, my gosh, I’d love to not have to do that.

These invoices are killing me. Or, and the rest of the documents, I keep acting like it’s only invoices, of course. It’s, a load of documents that go through the the accounting team. So let’s say they love this and they say, all right, I wanna do it. What happens? 

Rishi Srivastava: Usually it takes two weeks for us to get this thing fully into people’s hand, and most of it is users putting time with us.

When I say two weeks, I mean there’s not much training involved. The way we have designed the system, the UX is very intuitive where we don’t have to put much effort on training. It’s just making sure that, various people in the organization are aware of the new upcoming changes.

And after that people can just start using the system, and be saving times. 

Hugh Seaton: I love that. Do you find that there is need to review the documents that they, that the customer [00:18:00] gives you? Do you have to train it on the documents that they’re used to or by now, do you guys, you’ve covered so many of them that’s rarely the case.

Rishi Srivastava: That’s actually rarely the case, but the model’s improving every time the human is giving feedback to, as I said, the IES are in high nineties, but the times when it’s not right. The human corrects it and that feedback is routed to the the model dev team. And eventually, maybe it’s gonna be a couple months before the model is retrained it’s, it picks up the human feedback.

And the other day I was talking to a customer, he was exactly saying the same thing. Like maybe when you start, there’s gonna be some low confidence extraction, amber or red on a small frac fraction of the invoices you have. But as you give it the system feedback, it improves. 

Hugh Seaton: Also we are talking about documents that are intended to be pretty [00:19:00] standard.

This isn’t artwork. People want their invoices to be understood and they want their information to be communicated. So presumably by now, the number of the amount of learning, it’s not nothing, but it’s probably pretty, pretty small. Do you find that’s true?

It’s look man, people don’t want their invoices to be original works of art. They want them to be standard so everybody knows what they’re looking at. 

Rishi Srivastava: That’s a great point. But even if I want mine, which is to be, easily understood. But another person, let’s say you, when you create a voice they, it can be from a different template, And that’s where like when it, that your and my voice, let’s say they go to a construction company, the AI needs to be able to deal with both of them seamlessly. And that’s the kind of main problem we are solving. A lot of these old solutions in the market, they’re OCR, which is, just.

Not template agnostic, and the state we are in. Now these documents, they can be read regardless of the template they’re coming from, even though they’re not artwork. 

Hugh Seaton: Yeah, that makes a lot of [00:20:00] sense. So when you think about how you’d like to evolve this and where you see it going in the future, what are you guys looking at?

As you continue to develop, being human. 

Rishi Srivastava: We would like to, with a few more ERPs. It’s, we’ve always been a little bit strategic in where we go in terms of integrations, I’m a strong believer in, solving problems that are really painful and not when people are solving, instead of trying and going and competing with every single solution in the market.

So we’ve started with a couple of ERPs where there were no good solutions, not even OCR, and then we feel like we are gonna go to other ERPs a few ones that are in our pipeline. CMIC is one that we are looking at to go connect and on the, besides integration on the product side we are thinking of bringing some gen AI.

On part of the solution. So again, I’m a strong believer in only solving the problem that is painful. So in [00:21:00] our system, what happens is, let’s say an invoice is going to a pm and then the CFO, there’s a bunch of activity, right? Which is like coding happens. There’s, okay, the CF PM maybe says, oh, this this price doesn’t look like can you fix it? Coding doesn’t, right? There’s a lot of back and forth that we’re capturing in this documents lifecycle. Eventually we would like to summarize that with Gen ai. So let’s say a hundred back and forth, we just do it in three sentences, right? It’s so obvious to the people what happened to this invoice over its lifecycle.

So those are two major things I want to highlight. 

Hugh Seaton: That makes sense. I love that. And I think, yeah, there’s a lot of opportunity to think about, interacting with the data once it’s been, securely extracted.

 Rishi, I love this. Anytime we’re saving people, that kind of drudgery, that just winds up taking up tons of time and introduces errors and all that.

That’s exactly what. Software and AI are best at, so I’m really excited to hear that you’re doing that. If companies and people wanna reach out to you, what’s the right way to get started? 

Rishi Srivastava: So I’m very active at [00:22:00] LinkedIn. Besides LinkedIn we can always look at our website through their, you can schedule time with us also.

My email address is rishi@beinghuman.com. Being with two eyes. 

Hugh Seaton: Fantastic. In the show notes for those who are interested, I’ll have all that. I have the website and how you can follow Rishi and his email. Awesome. Rishi, thanks for being on the podcast. I love what you guys are doing.

Thank you so much here for having me.