Unlocking the Power of AI in Legacy Contract Migration with Samir Bhatia

Made with speech-to-text software, so a few words may be off.

0:00You're listening to the Contract Heroes podcast, your one-stop shop for all things contract management. And now here are your hosts, Marc and Pepe. Hello, everyone, and welcome to another episode of Contract Heroes. Today, joining us from Brightleaf is the CEO, Samir Bhatia. Samir, thank you so much for joining us. We're really excited to have you on the, the show today. I love to be here, Marc and Pepe. Looking forward to the questions that you have and answer them. Yeah, absolutely. And before we get into that, I think our listeners would be interested to, to kind of hear how you've gotten into, you know, into this space, right? You, you've been in the space for over 10 years, and we're gonna be talking about legacy migration today, right? That's kind of the focus. But how'd you start Brightleaf and kind of get into the legacy migration game? Okay. So there was a company called Brightleaf Corporation. My company is called Brightleaf Solutions.

0:54Brightleaf Corporation started sometime in 2007 or '08 or something, and it was well-funded by some VCs, and they were looking at assembly of contracts with clause libraries and typically stuff that the CLM would do now. And but they were trying to sell to law firms, and that wasn't very successful. They f- a friend of mine worked there, and they, he got me because I turn around companies, and he says, "Help to turn it around, but the VC says we are not funding it anymore." So while they were going to cor- first initially going to law firms, they decided to go to corporates. Corporate says, "I don't wanna assemble contracts. I wanna know what's within my contracts." So they started the semblance of an extraction, as I call it, extraction document, data extraction business, but there's nothing there except the logo and potential of two clients. So I took over the assets, I took over the logo, and the potential of two clients. British Telecom was the first client, and to date, we are still doing work for them and developed our AI software.

1:52In that time, it was not AI, but it was always been AI. I don't know when that term started. But extraction software in various ways about 10, 11 years ago landed BT and got a ton of clients, and that's all I've focused on as a company, hyper-focused on legacy contract data extraction, but as a service using AI software. And our AI software is everything that comes out. It's hard to keep up these days, but if you talk about gen AI, we have that agentic. So we keep enhancing the software, but we use it internally to provide the service because there's a whole layer, which you'll probably talk about, Marc, where you have to do validation after AI software finishes, and we'll talk about why, too. So that's the history of Brightleaf and me. And I think that was the perfect intro because you just summarized what we're going to talk about ... in this interview, Samir. Mm-hmm. So I would say that one of the biggest heavy lifting in a CLM implementations when you wanna start working with all of the contracts is the legacy documents migration.

2:51Because a lot of the time people look for a CLM because they have a mess with their contracts. Maybe if they have a SharePoint or a Drive with their contracts, that's good. But the thing is that a lot of the times they just, once they sign a contract, they just put it on the file, and then they may be opening it again when something wrong happened, right? And that's, like, one of the main common problems that peop- that company faces before they start looking for a CLM. So I know because we've been in this space for a couple of years already, that legacy documents migration is not just, "Okay, here's my couple thousand contracts. Just help me do all the legacy migration." You need to have a process before that. Samir, like, can you tell us a little bit when, in your experience, like, when a company is starting a legacy contract migration, what's the best first step to take? Okay, so you hit it on the head there, Pepe.

3:50So when you've got tons of contracts, and if you've got five, 10 contracts in a small company, the, the migration is very small, but you start getting to thousands, and we deal with sometimes hundreds of thousands of documents. Now, there's a difference between documents and contracts. Typically, they are all over the company. If it's geographically distributed in different departments, it's all over. So first thing you need to do is get all those into some central repository, okay? Just to, as a side note, Brightleaf has... I've written a white paper on sorting through and collating the contracts, and there are best practices around that, but the project manager shouldn't own that task themselves because different departments have their contracts and documents, I just call them, and in different geographical locations. So you've got to come up with a project plan. I'm gonna give you two months to get everything in a central repository. Design a folder structure.

4:43The folder structure has to be consistent just so that you get some amount of organization when you're gathering the contracts. So and you say, "Okay, it's your responsibility, the sourcing department from a geographical location of, let's say, US, that you guys are supposed to get them into this." The file folder structure is by vendor name. So under the vendor name, if you have a subfolder of agreement types, whether it's NDA, sourcing, order forms, whatever, that makes it better. The more organization you do, the less you have to pay some provider to do the work because somebody has to do the work, right? So if you're doing it while you're getting the contracts organized, if you de-duplicate them, remove the unsigned ones, remove the drafts, all that cleanup can be done while you're doing and gathering the contracts. But let's say you give a two, three-month project plan in a larger company to get everything into a central folder. Now, at least you've got some semblance of a, what you want to migrate into the system. That doesn't mean you migrate all the contracts into a system, and maybe that's a separate question Pepe will ask me, for what should you migrate, but I won't go there now.

5:45But at least you've got everything somewhere in a central location to start doing more work behind that. But that's the start of how you would do it. Yeah, I, and I think it's sometimes a heavier lift, right, than people are thinking it's going to be to, to go out there and, and find where all these contracts or documents are, are stored. And, you know, and then we look at kind of the next question here that think people should be asking themselves, and that's what metadata do I want to extract from these agreements, right? And, and, Samir, maybe you can give us a bit more information on, you know, the thought process there, right, helping clients to understand what data... 'Cause I'm sure they, you know, of course, they know what some of the data probably is that's going to be important. But what's some of the, the boilerplate items and maybe some things that folks don't think about that they should be looking at when they're extracting metadata? So before I answer that question, just let's stick to the document level, Marc, if you don't mind. Sure.

6:37So at the document level, now that the documents are centrally in some folders, you can use a deduplication software which you get off the shelf to at least remove some of the duplicates. You can remove the DOCX files, you can remove the MSG files, whatever you don't want, right? Remove all those stuff and try to clean it up as much as possible. You won't be able to do a whole cleanup because there's a lot of processes involved, uh, in there to do the cleanup, but clean it up as much as possible. And then you've got to decide, uh, what you say, wh- which ones are your priority. So are you gonna migrate all the documents? So even at the, like I told Pepe earlier, you can decide not to migrate all the contracts. Now, when we talked to Nasdaq, we said, "Why don't you go to your AR and AP systems and see which vendors you've done business with in the last five years? Anything you haven't done business in the last five years, why do you wanna migrate?

7:29Leave it in the file folders or just put it into the CLM or wherever you're putting all your data and your legacy contracts in, in an account says, 'All expired contracts' or something, you know. So at least it's there in a central repository." But you can decide what you want to migrate by looking at your AR and AP systems to see who we've done business with, right? That's one way to filter down and boil the ocean into smaller chunks. Now, coming to your question is now you decided that, then you have to decide what categorizations you want in the documents. Are you gonna call them, is it NDAs, MSAs, SOWs, order forms, vendor agreements? You decide those categorizations. Then you have to map the header that is there in either the top of the contract or in the preamble to those types. So whether it says, uh, Master Services Agreement, Professional Services Agreement, MSA, the type is always MSA because that's how it's configured from a CLM. So you have to now start thinking from the CLM recipient perspective, the system recipient perspective, how you're gonna use it in the future to do this categorization, and you do the categorization at the document level there.

8:33Then, so one of the attributes you have to extract is the document header, right? What is, is it an MSA or is an SOW, so the type of contract. That's one, one of the elements you have to extract. Then the other mandatory... Then you've got to look at, at the CLM and see what are the mandatory attributes it needs. It always needs a counterparty or an account name. So you've got to decide, okay, what are the contracts are drafted any way, shape, or form, which is why we'll come to this, but you always need a human oversight after AI software finishes. So what is the party? What is the counterparty? So you have to list, come up with a list of your party names so that the other one becomes a counterparty. So the company might have bought different companies in the past, so that becomes all the parties. But it could be said in different ways. It could be IBM, International Business Machines, and just the drafting is poor. So you always need that oversight. So you have to decide what is your party name so everything else becomes a counterparty. Now, extraction is never perfect because the drafting is not perfect.

9:32That's why you need to QC it. Then, so the counterparty is needed. You always need the effective date. You might need the initial term date. So if it, the initial term length. So if it's two years from initial term, you know the first expiration date is two years from then. So you might have the initial term expiration date as a calculated then. Then you want to know whether it's the, the renewal type. Is it evergreen? Is it perpetual? Is it auto-renew? So that you can calc- know at any given time whether a contract is current or not. So there's this set of attributes, about eight to 10, that you always should extract just so that you know the basic information, the, the contract title, the contract type, the effective date, the initial term length. You know, when you get to the date-based attributes to know if it's current or not, then you can start layering. You can say, "Yeah, I know it's a perpetual contract or auto-renew, and it's forever. Well, what are the cancellation terms? Is there convenience for cancellation?" So you can make that a yes/no attribute.

10:30See, now you start going a layer deeper, but these are some attributes across all of them, across all the contracts that are required, and then you start... And you do it in a phased approach. So don't try to say, "I want 100 attributes," in the first phase. Firstly, 100 attributes if you want, are you gonna, the new contracts coming in, put key in 100 attributes, you know? Don't expect that. So that's a totally overkill. All the clients have asked us for that because they want to know what's really deep in their contracts, but on a subset of contracts. So then you say, "By each contract type, I wanna add these attributes. From my vendor agreements, I want to know some other stuff more than just these basic attributes. From my N- MSAs, I want to know is there first measure. From my NDAs, I wanna know is there non-solicitation clause." So you start going deeper, but do it in a phased approach by starting with the basic set of attributes and then going deeper by contract. Yeah, that's a good one.

11:26And so, Samir, a lot of the times, I was actually, while you were speaking, thinking about there's this study by the World Commerce and Contracting. In more than a study, it was a maturity test, right, for contract management. And one of the questions where a lot of the companies, even big companies failed, was if there was a centralized repository for documents. And the thing is, and a lot of times even big companies that we think they are have enough technology to do that, but they don't. They have siloed repositories, and a lot of the times maybe by department, procurement has their own, sales has their own, IT has their own. So when you're facing that type of organizations, what would you say is the, a good repository structure or classification method to work best for this type of large Siloed repository for these companies Right. So you have to think backwards from a CLM perspective to then how you map the documents. So you wanna see...

12:22When you're getting a s- you're typically looking at a CLM s- because your contracts, like you said, Pepe, are all over the place. They're not organized. So you want at least a central repository. But having just a central repository is not enough because then you don't know what's in, in your contract. What are the expiration dates? So you're digging deeper to get those attributes out of the contracts, right? But getting them all in one place, you've gotta say, "Okay, now how is my business running?" And that's where Koho Consulting would come in more than we would, where you would advise them looking at their business processes, what they're doing. Are you going to divide it geographically? Are you gonna divide it geographically and then by department? So if it's... And even geography, is it by state in the US, or is it USA one and UK one and Germany one, or is there cross-functional stuff that is needed? But Pepe, to put the question back to you, it's more of an advisory role on Koho that can help to get there because you do that advisory to them knowing each one is individualized versus the other.

13:23But on the higher level, you at least want the categorization of is it sell side, buy side? Do you get deeper? And what are the, just the legal contracts like NDAs? So you categorize them. That's why I said at the initial start of, at the document level, what is the categorization of the documents that you want? And then on top of that you can layer do you wanna divide it geographically or whichever way you want, or de- department, et cetera. But yeah, the users, so you have to look at it from a user standpoint, work backwards what they do daily basis from a CLM perspective, and then come up with the categorization that you want. And now on the categorization, another of the issues is, of course, a lot of the companies don't have a way on how to connect companies. Let's... To connect, let's think about MSA to their contract to their amendment, right? So how do you handle those, especially when you have, like, everything in just one space? Is that part when you were, like, tagging, if it's an MSA and then if it's like a subcontract to that MSA and then the amendment, is that the way on how you structure the information while doing the import, how we start to make sense?

14:29Absolutely. And, and we've got patents around that too, Pepe. So on the software patents on how do you create the parent-child relationship or the hierarchy. So like you mentioned, there's an MSA. There's an amendment to the MSA, amendment one, amendment two, amendment three. Then under that you have an SOW. SOWs are a little weird because they are little contracts on their own, but they're still under the master services agreement. So in order to do that hierarchy creation, it is an involved task, and we figured it out and got some patents on that. So you've gotta extract the, some information from the amendments that match it up to the masters, and that needs to be configurable. That's what we have. And when you do that, otherwise you're taking... If you have 10,000 documents, you're opening one, you come across an amendment, and you're trying to search which from the 9,999 others, which is the master. That's very labori- laborious process. So we've got some best practices around that and software and patents underneath that, that we create everything and the whole hierarchy for the client.

15:27So little bit on Brightleaf, we typically get a mess of documents like you started, Pepe, and talking to me about, do a perfect spreadsheet of customized and extracted data, all the work including amendment creation, including renaming the documents. We do that for a living day in and day out. But yeah, you've gotta extract information that does the mapping of the amendment to the master. You will always, always get amendments that you cannot map to the masters for multiple reasons. One's they haven't, the clients don't, haven't given us the amendments of the masters, or the clients can't find the masters. But typically it's a drafting issue where it doesn't have enough information, or the party name has changed, or the effective date of the master agreement is not in the amendment, and there are multiple masters in there, in them. So we don't know which one to map it. But from Brightleaf's perspective, we'll do all the work and just the exception, such exception handling is done by the client.

16:22Then they go run and find the uh, the documents and say, "Oh, well, yeah, I know this party name changed. Yeah, so map it to this one." You see? Interesting. Okay. It's an involved process. It is... And you hit the nail on the head. Legacy contract migration is not trivial. Yeah. And architecting all that. No, absolutely. And I think, you know, it's interesting, right? And we, we talked about kind of that human element in this process, right? And obviously it's not cheap. It's an expensive task to have folks manually review contracts and go through it to make sure that, you know, you get above that 99% success rate when you're looking at, at things that AI has extracted. So what does that process look like, right? From taking the, the data that the AI has given you and, and then having folks, you know, take a look at it and start the review process. Okay. So good question, Marc. But be- before even you look at the extraction of data from the contracts, all the stuff I talked about, document level work needs to be done before you cut to the extraction of AI, using AI to extract that. But once you've cleaned up the document, you use software at all this time, but always have a human element. If the software says, "This is a duplicate," but somehow you've gotta just make sure it is really true or not.

17:31So duplicates, yes or no, with a checksum you can do it, but if you put the duplicate parameter as 99%, it might still be a few full duplicate because it just scanned in wrong or something like that, or it was not OCR'd correctly. Many reasons. So once you do all that document level stuff, then to answer your question, Marc, you have to think about what attributes you want, but also define the attributes in the way you interpret them, right? One of the examples I gave you was a party and counterparty. So you've got to know what is your party name to know the other party is a counterparty. Otherwise, the AI software will just, you know, take the first party as the, as the co- client and the other one as the counterparty or whatever, and that's wrong. So there's no real way except to get a list- And then it never match perfectly. So do you put it a percentage match? Like somebody has misspelled I- international business machines with, and just said machine. Is that yes, you know, how close it is. So you've got to make these judgmental calls or say, "Signal to me so I can verify and, and look at these exceptions." So then what is the start date?

18:33I mean, if you say extract the start date, I've never written an SOW that say, "This engagement will start on..." So do you put in the effective date? Do you put in the date of signature? Is it date of counter signature? Is the initial term start date? It could be multiple answers. So how do you come up with, with a answer that you want? You decide that as a group, and that's what Brightleaf helps facilitate, and document all that in a playbook. Then train your AI software to those definitions. Now the AI software will go through and crawl, and the only way to get perfection and 99% and above is to check every attribute. Now, AI vendors say that I will get 95% extraction. Two things, you don't know which 5% is wrong, but that 95% is a challenging number only because it is not the fault of the AI software. But since they are legacy contracts, the drafting is suspect, and there's not enough information, conflicting information.

19:33And I can show you so many examples that the accuracy, because of the documents, goes down as the document counts ma- ma- mount up. Which is why thorough process is check every attribute by qualified legal personnel, because it gets so interpretative at, at times. Whenever we bring up the exceptions to the clients, they are, "I'll get back to you," or they're talking for 10 minutes and trying to figure out what is the right answer. So if you just leave it to the AI software, when the document counts are in, you know, thousands, it does get very tricky for multiple levels. So the stringent process is to check every attribute to make sure it's perfect. Samir, I think that's... It's a super important step that I think people, organizations, right, especially if it's their first time migrating this data from whatever system they've probably been using, from a legacy system or, or SharePoint, you know, I think it's, it's not something that folks think about, right? They think that the, they're gonna push it through the, the AI system, and it's gonna spit out perfect information.

20:30And so now that we've had folks review it, right, and that you have that data on hand, what does that process look like when, when you're ready to import it into the CLM of choice? Okay. So I alluded to having a creation of the playbook for the rules of extraction. To create the pay playbook, you have to get a good representative set of your contracts, you know, third party papers, different way of saying the same thing. Very loosely put, if you know some contracts say for force majeure, says force majeure. Some say volcano, some say earthquake. Put that in the sampling. So all that has got to train the AI that it means the same thing, right, for you, and that's all dependent on how you want to interpret that attribute. So it is important to train the AI correctly so that you can extract this. And when you do the sampling, let's say you do, do 15 to 20 contracts of a good representative set, then you d- iterate on the playbook, iterate on the definition, do a second sample.

21:29Typically, we've noticed that after three samples, things subside. The input from whoever is configuring the CLM needs to be there. Have you configured it as a mandatory field? Let's say it's an end date, but it's a perpetual field, and it's the end date. It's a perpetual contract, auto-renew, so it doesn't have an end date. But you've configured it as mandatory. We, we can't give you back a blank. It'll burp when you wanna ingest it. So is it a pick list of, let's say it's states of jurisdiction, M-A, C-A, D-E, or is it free-flowing text of short text of Commonwealth of Massachusetts? If it's a pick list, everything has to ma- match exactly to the pick list. So look at it from a CLM perspective. Those inputs also need to go into the playbook. And after the third sample you've done, test ingest the batch of the third sample to make sure nothing is left so that you're not doing a ton of work at the tail end and re-extracting stuff in a different definition.

22:26And update the definition at the tail end within the software on a CLM perspective as a recipient system. So do that, and then the client and, and you can decide if you've got 10,000 contracts, do you want to do it in four, five, six batches, whatever. What is the priority? Is the vendor agreements first? If you... Do you have a list of counterparties that you do business with more you want or more spend that you want us, that you want to do first? Because it does take time to do the QC. Let's say that 10,000 contracts and on an average of 20 attributes, just to be easy on the math, somebody has to check 200,000 data points if you wanna do it correct. So that takes some, a, a team and effort to do that. And therefore, you extract it in batches and then decide, do you wanna ingest the batches as they come out or wait till the tail end? But the best, best practice is ingest in batches so that, A, you don't need to wait till the tail end. B, if anything that is needs to be changed in the definitions, you are not reprocessing 10,000 of them, just what you've done. No, I think that's great advice, right?

23:30And again, right, something maybe folks, folks aren't thinking of when they're going to be ingesting this information into their, their CLM. So Samir, this has been a fantastic conversation. I, I'm sure we could sit here for, you know, hours talking about this with our backgrounds, but I, I think we're gonna wrap the episode up here. So if folks wanna learn, you know, more about Brightleaf, or if they wanna connect with you, where's the best place to do that? So just send me a personal email. I'm fine with that, too. S... My name is Samir Bhatia. Email is sb, as in boy, h as in Harry, a as in apple, t as in Tom, i as in India, a as in apple @brightleaf.com. Or if you forget that, and it's hard to pronounce, just go for info@brightleaf.com, or there's a Contact Me form on our website, so multiple ways to send me something, and we will be in touch. And- Fantastic... we go through all the pricing and everything also, because just to allude to that, Amar, since Brightleaf does not charge for the AI software but charges for the checking, it's as granular as per attribute because we are cha- charging to check the and fix the attribute. Right. So it's as granular, and that, and the clients could save a lot of money. Yeah. Yeah. Makes total sense. Well, Samir, thank you so much for joining us. This has been a great episode. We appreciate you coming on.

24:38I know it's been a long time coming, and thanks everybody for listening to another episode of Contract Heroes. Thank you, Marc. Thanks, Pepe, for having me. This episode of the Contract Heroes is sponsored by Koho Consulting. Koho helps organizations of all shapes and sizes find, implement, and manage their chosen CLM tool. Whether your project is buy or sell side led, Koho can help you navigate the waters of the CLM space to make sure that you're getting a product that's going to fit like a glove years to come. If you'd like to learn more about Koho, visit their website at C-O-H-O consulting.com.