EPISODE 193 – YOU CAN REACT TO CONTRACTS & AGREEMENTS AT LIGHTNING SPEED WITH LEGAL ROBOT’S NEW AI TOOL

Legal Robot’s Artificial Intelligence engine can scan a long form contract, and provide you with a 1 page summary of all the important issues contained in the contract, as well as red flags that may be present.

Legal Robot substitutes traditional contract review (and drafting as well) with an automated, intelligent natural language processing AI system. Using their legal language model, the intelligent assistant flags issues & proposes enhancements by considering best practices, risk factors, and jurisdictional differences.

Legal Robot gives you an immediate error check for your contracts helping you produce better legal documents, but more importantly, avoiding contracts that would be detrimental to your wishes, unfair agreements, and agreements with hidden clauses that can hurt you.

Listen to the complete interview below. Make sure to subscribe and post a review.

 

Podcast: Play in new window | Download

Richard Jacobs: Okay. This is Richard Jacobs. Today I’m talking with Dan Reubens, co-founder of Legal Robot. Dan How are you doing?

Dan Reubens: Good. Thanks.

Richard Jacobs: Thanks for coming on the show. I really appreciate it. It would be interesting. We are talking about Artificial Intelligence and how your firm is using it in the legal field. So can you give a quick background on what it is that Legal Robot does and what areas of the legal industry you work in?

Dan Reubens: Sure. We at Legal Robot analyze contracts with machine learning. We try to make it more understandable and help to free companies from some of the friction that contracts cause in their legal processes.

Richard Jacobs: Are you using natural language processing to interpret what’s in a contract? Is that how it works?

Dan Reubens: Yes, we use natural language processing and some other forms of machine learning. There is one specific field of machine learning called deep learning. We use that to understand how people use legal language in practice and then use that to understand how it’s being used in a particular target document.

Richard Jacobs: Can you give me an example of a couple of legal terms that are pretty common and that your AI system can understand them?

Dan Reubens: We don’t think of them as individual legal terms. We really look at the entire sentence or the entire paragraph in context to the whole document. So we can throw out a word like indemnification and we see a lot of that in contracts naturally but the meaning of that really changes based on the words around it. So we really like to look at the entire context that the word is used in. That might be a trigger word for a lawyer that is reading contracts, they want to look at that section in particular and look for something but we really try to look at the entire sentence in a way that it’s being used.

Richard Jacobs: Okay, so a document will be analyzed and then what’s next? An attorney will be given the analysis of the document along with the original? What’s the process?

Dan Reubens: Sometimes we work with attorneys but we’ve actually found that usually we are working directly with people in the business. We are working with a head of sales or a sales person himself or we are working with someone in HR. These are all sort of the end users of law and I think what’s funny is that lawyers and by the way, I’m not a lawyer. I’m a machine learning guy myself, that I think lawyers sometimes tend to forget that the end users of law actually aren’t lawyers. Lawyers are sort of the mechanics of law. They fix problems and that’s great. They are very necessary in that but they are not really the end users of the law. The people that are actually operating businesses. There are individuals that are actually trying to resolve legal issues. So we really are making more products for the end users of law in trying to make the law more friendly for them.

Richard Jacobs: Makes sense. So as an end user, let’s say I’m going to rent a commercial building and I get an agreement from the landlord’s attorney and it’s like a 30 page nightmare and my eyes are crossing. What will Legal Robot do for me in that example? What will I get from it as an output? Will it translate the gobbledygook into English? What will it do?

Dan Reubens: You get a few different things. For one thing, you will get a certain one page summary for the document. Some things that help you understand the key terms in the document and understand the point of the document. It’s basically like an automatically generating term sheet. We also pair that with some deeper level analysis. If you are trying to proof read the document, it hasn’t been executed yet and there is still some time for changes, we’ll help you spot legal issues in there and maybe used the definition wrong or maybe there is some other sort of problem in there. Maybe you really probably should put something like a reasonable effort instead of best effort. Maybe that’s sort of the standard for that type of agreement. We’ll look for those sorts of errors and we’ll provide some statistics on how often certain terms are used in that sort of agreement and then you can make your own decision if you want to go ahead and negotiate on that point. We won’t necessarily tell you the way but we’ll give you some more data to serve and inform that decision.

Richard Jacobs: So, it sounds like, for a particular agreement, you’ve had to have trained your system on similar agreements in the same niche or genre of the agreement, right?

Dan Reubens: Right. Yes, we’ve trained very generally on legal language and how people use legal language. That was a very large data set that we trained on. We’ve also trained on specific legal agreements that you might have more targeted questions about. So if you have for example like a stock purchase agreement or a commercial lease or something like that. Those are some things that are very well represented in private data sources. That’s something that we have trained on. Other data sets, if we don’t have a particular type of contract that we’ve seen a lot on maybe your own niche or industry, you can actually go and train your own algorithms on that through our web interface.

Richard Jacobs: Are you saying that you have an API where you can take your AI system and train it on my data, for instance?

Dan Reubens: Yes. You can actually do that yourself without any coding. Let’s say, you work in oil and gas and you have some really specific type of patented agreement that you are looking for and we haven’t trained on it yet. Right out of the box, you will get some use out of our product. We understand how people generally use legal language but for something really specific, you can train your own classifiers, your own algorithms and you can run your own checks on documents that come through the door.

Richard Jacobs: So the results; let’s say it’s a commercial lease. So I’ll get a one page summary on what the lease is about in plain English. Will the system also point out, hey this provision right here is not customary. In only 3% of leases reviewed have we seen this term or what else do you gather that’s useful in how I want to use it?

Dan Reubens: We do a number of things. We’ll acquire things like Can Adams manual self-contract drafting plotters. But yeah we also do the kind of statistical analysis that says, in this type of document which we’ve reviewed, a statistically significant number of these documents, we don’t really see this clause a lot. Let’s say you are looking at an NDA and for some reason, someone copied and pasted something and they got a royalty clause in there. As a lawyer, you can review that and just be like, what’s this doing in here? Get this out of here. This is clearly an error but a nonprofessional who is reading that might not spot that. So it will help you spot that and then we’ll also go a little deeper. We’ll actually help you to identify areas of ambiguity. Sometimes that’s strategic ambiguity, sometimes it’s not intentional and it’s sort of the way the document is structured that can result in significant changes in meaning.

Richard Jacobs: So if I use Legal Robot on a contract, do you still recommend that I get an attorney to review it and your analysis? Will you be able to make a recommendation like that?

Dan Reubens: We definitely think that people still need them. Attorneys aren’t going anywhere. This is definitely not going to put any attorneys out of business. This is still a parlance of proper legal analysis. We just think that for certain individuals, this is something that can help them have more meaningful and productive consultation with their attorney and when we are thinking about using a more appropriate environment, this is something that allows the lines of business or the people that are actually operating and working with these contracts. It lets them sort of have a pre-check before a legal department review. They sort of know that this is going to get through legal and that none of the stupid errors that I usually send to legal, they are not going to get flagged in here or I just need to go in and fix this one thing that is not compliant with my in house legal guidelines or something like that. That is sort of one of the cases that we work with but on the consumer side where it’s sort of a pre-check or prima facie review for some of those lawyers out there and then on the corporate side, it’s again something to make sure that you are compliant with in house standards.

Richard Jacobs: Are any of your client’s attorneys or the firms themselves? What about a firm that has to review thousands of contracts or high volume stuff? Do you have a use for law firms themselves?

Dan Reubens: We definitely do have law firms that are interested. Quite a few law firms have signed up for our early data. We’ve actually got 8 of the Top 10 global law firms signed up for our early data. I think a lot of them are interested in it for some due diligence and Boyer scale document review cases. How that ends up playing out in the market, we’ll see. Honestly, they are not our primary clients. We’d really rather sell to the in house legal department in a business. Selling to law firms is notoriously difficult and we find that a lot of them are already adopting technology that is quite helpful, so they may not have a need for it. If you’re in a large law firm and you are using associate time to do large scale document review, you are not really using your associates well. So a lot of them in recent years, have made certain advances and invested in their own technology for this.

Richard Jacobs: Technology that uses AI or technology that just uses character recognition like mass scanning into a database?

Dan Reubens: Sort of the basic level of the document understanding and the natural language processing. There are a lot of good e-discovery tools that can be adapted to this purpose, so we don’t really see a huge market there, honestly.

Richard Jacobs: You see a big market in the end user. The person that wants to rent a commercial property or is negotiating a NDA or other agreement with the company, that kind of thing?

Dan Reubens: Actually yes. Going back to the sort of car example. If the end user of the car is the driver. We all need to drive a car to get to work that’s really our end user and not the mechanics. We’d really much prefer to service all of the end users in the world rather than just the mechanics.

Richard Jacobs: It seems like a big imbalance of power especially with the terms of use or somebody saying, this is web related. Could this be used as a tool for people to apply when faced suddenly with a 30 page terms of use document? Are there any landmines if I agree to this?

Dan Reubens: Well, actually, recently I just got an email from Uber and they just updated their terms of service and I was like, I would like to know the summary of what they’ve changed but I’m kind of curious. What did they really change? What are they trying to slip in there that’s needed to try and fly under the radar. So I mentioned some changes to the arbitration procedures. Turns out those changes might not be totally in the consumer’s favor as you would expect. It’s kind of interesting. We do see some things like that and we can pretty quickly analyze those major changes to those major terms and privacy policies. We are actually going to make a lot of those public. Changes to major services like Apple and Google and Uber and sort of the big companies that tend to make a lot of news, we are actually going to make those changes public.

Richard Jacobs: Oh really? So you are going to analyze them yourself and you are going to make them public?

Dan Reubens: Yes.

Richard Jacobs: Okay, very interesting. Can you talk about some of the mechanics that you deal with? Do you use programming languages? If so, what do you use? How does a piece of code do what your program does? How does it analyze a document or multiple documents and find concepts or understand them?

Dan Reubens: If you think of sort of the traditional programming of a computer, you are writing out rules and you are writing out the logic of particular rules and you compose them into functions and higher level objects. That’s certainly one way to write a program. But I think what has changed in the computer science world in the last few years and what made this possible is we are now evaluating the source document rather than passing it out through a big set of handwritten rules. What we are doing is we are looking at the internal relationships in that document. The way we start is we actually build what’s called a code trans matrix or we count how many times a word shows up alongside every other word. That’s a lot of computational work to do that which is why we’ve only been able to do it recently. So we start out and we build that code trans matrix. That basically gets changed into a long set of numbers that represent a single word or a single sense. We now have basically a set of numbers that represent an indemnification clause. We can test that long string of numbers and pass that through a bunch of old style statistic algorithms and we can analyze the relationship between that set of numbers and other sets of numbers, so we can see how close that indemnification clause is as represented by those numbers. We can see how close that is to other indemnification clauses. Then we can say, Okay, it’s this far apart and we can measure that distance. That’s something which is really interesting that’s become possible in the last 5 years that has enabled our technology.

Richard Jacobs: Alright. So you are looking for occurrence of certain words in the same sentence or not. Obviously, you’ve disclosed as much as you can but what else goes into making your program work besides the frequency of certain words appearing together? How does it even understand a concept? How does it work?

Dan Reubens: What we do is we take the tested long strings of numbers and we build what’s called a classifier on top of it and when we build enough classifiers, we get to the point where we can understand what legal concepts are being used in a particular sentence or a particular paragraph. We actually send that to phonontology. So if you remember back to legal theory and in drawing out the different sort of exchanges in the contract, you can actually draw those out in between two parties if you have a circle on one side as Party A and a circle on the other side which is Party B. You draw a line between them. That’s an exchange. So we’ll actually build out each of those components using a classifier. We build a whole lot of those, so we are actually able to model a contract in that way. We are able to draw all the different lines and concepts of that contract. We call it the conceptual draft of the contract and then we prepare that conceptual draft to other contracts.

Richard Jacobs: How do you define the universe of all the contractual exchanges and elements that are possible and then your system is looking for when any one of those elements occurs in an agreement that it’s reviewing?

Dan Reubens: Right. We can look at an agreement and say that it’s got a 100 different exchanges in the agreement or 12 different changes in the agreement. We can look at the nature of those exchanges and how they transfer between different parties and different characteristics than they have. That provides us with really rich information that we can then display to the user of the software. Suffice to say, how the complicated stuff behind the scenes is made but on the surface what happens is you upload a document and within a few seconds, you get very rich summary of the legal language inside and what’s really happening in the legal language behind of it. It’s a much more visual way to work with a legal language than most people are used to. I mean, you could flip through all 60 pages of the contract but if you can also look in there and on one screen, find the one particular instruction error that you care about. That’s a much more efficient way to review a contract.

Richard Jacobs: Makes sense. I’ve seen out there on the web, Google has a natural language processing API, Watson, IBM also has one. Did you guys have to build your own from scratch or did you piggyback off those? How good is your versus theirs?

Dan Reubens: We actually started out with what suits you best thinking. We can get this into the market real quick if we just used some off the shelf tools. There are some really great natural language tools out there. The field has just been taking leaps and bounds in recent years. What we found out very quickly is that a lot of them are trained on sort of traditional prose, so they are trained on normal sentences and there is really nothing natural about legal language. We really couldn’t use those off the shelf tools. We had to build our own. We started off trying to use those and we tried Watson and we tried the Stanford Pore and LT Toolkit and we tried some other tools that are relatively well known and they provided some insight to be sure but they really didn’t give us the kind of level of accuracy that we were looking for on legal sentences. The reason is that legal sentences are so much longer, they have what’s called the long range dependency problem where you have a sentence that might be 15 words in normal text and in legal text, it ends up being something that is 40 or 50 words long in a single sentence. So it’s a really different kind of problem. The nice thing is, we get a lot more hierarchy in legal language, so we can actually explore that hierarchy in unique ways, so we ended up doing our own algorithms but it ended up being something really good for us as far as accuracy.

Richard Jacobs: Where are you at with your product? Is it available to the public and for what types of agreements or is it still in data? Where are you at in the process?

Dan Reubens: So it’s not available to the public yet. We are working with 6 companies that are in the hyper data stage right now. We are really trying to get a sense of how they use it day to day in their business before we actually launch it publicly. This is something where people are going to be depending on this for understanding legal language. They could be making business decisions based off this. So we want to make sure that this is something really solid and we also need to make sure that it’s incredibly secure. So we are only working with a small set of customers at this point just to make sure that we get everything right early on. We haven’t launched it publicly yet but in just a couple of weeks, we are actually going to be launching just a small product that is just a part of it to do some legal automation work.

Richard Jacobs: What’s that going to be?

Dan Reubens: Just recently, the US Copyright office kind of screwed up the entire digital millennium copyright act, Safe Harbor for everyone. So they decided to wipe out everyone’s Safe Harbor rights and sort of replace them with a new online registry which is good in a way but there is going to be a scramble over the next year to re-register for Safe Harbor for any website based in the US that hosts user created content. We are actually providing a product that launches in a couple of weeks that automates that process for website owners. So they may not be legal professionals, they may not be people that are super-technical and familiar with this but we will give them this really important legal protection with a minimum amount of effort. We’ve provided essentially a DMCA Safe Harbor bot that will go ahead and do the registrations for them and keep them up to date.

Richard Jacobs: It sounds like your tool could actually help with compliance. Let’s say, crafting a compliant terms of service policy, privacy policy, those kinds of things. I would guess multiple uses of legislation need to be addressed in one of those. Do you see a use case for a product like that?

Dan Reubens: Yes, we do. There is actually some other folks out there in the legal tech industry doing some interesting things with that. Naturally, on a large scale there are products like Contract Express that are quite useful. There are also tools like iubenda. So those things are interesting and certainly useful but we were trying to stay away from automatically generating legal language because that gets much more into the realm of actual legal practice. When you are generating a particular document for someone, it might straddle the line between sort of providing them something that they are going to use and actually I guess, it really depends upon as a legal document.

Richard Jacobs: Where do you see your biggest opportunity in the upcoming years?

Dan Reubens: I think you nailed it with the compliance and the automation there. Really, the way that we think of our product right now was contracting and letters. This is our sort of product. We are eventually moving more into the compliance space as we develop this product more. It’s really compliance with contracts and the automation of that compliance. I think that’s where we are headed as a company. As far as the overall industry, I think there is going to be more of a push towards smaller units of work. When we hire a lawyer, we typically go through a service called out counsel that when you think of it as the word for legal services. So we go to them and we hire a partner level attorney for a fraction of the cost and we get a fixed fee. So I think you are going to have a lot of small businesses going toward that kind of model because it fits with their consumption model.

Richard Jacobs: The nature of AI and machine learning seems to be harmless improvement on a pretty expansive and vast scale. Do you see that happening with your products? How “good” will they get and what will it mean when they become better than humans or exceptional. What do you see as happening with your products?

Dan Reubens: There is always this question that at what point will we reach human level intelligence with a particular product. I think that’s really sort of asking the wrong question and sort of looking at that in the wrong way. Computers are already much better than humans in a lot of ways. They are certainly much faster and so if you are trying to review 10,000 documents, you can do that in a few minutes with a computer or with some AI. I really think that kind of question is interesting but I’d really think of it more in what sort of things do humans do today which will no longer need to be done. If we think of filing a brief or something, the human is probably going to come up with legal concepts on that and whether they actually write the whole document or not, I think that will change. I think enforcing compliance will change but as far as the creative aspects, I think we’re a ways away from computers managing the creative aspects of certain intelligence for quite a while.

Richard Jacobs: Okay and the last question. To get some use out of this interview or this podcast, who would make a good client for you at this stage? Is it an individual or a company or a law firm? How can they get in contact with you?

Dan Reubens: The bulk of our interest right now is acquiring customers and large companies. These are typically heads of sales or heads of HR or heads of procurement that are having problems with their in house legal team. Or if the in house legal team is really feeling the pain of deficiency in their own department. Those are our ideal customers right now. They can get in touch with us through our website legalrobot.com or just drop us a line at hello@legalrobot.com.

Richard Jacobs: Okay. It’s been really great and informative and very interesting. I think the work that you are doing is really fascinating. It has a lot of applications and I’m glad it’s more of a contract side and seemingly at the consumer side as well. So thanks for taking the time for this interview. I really appreciate it.

Dan Reubens: Sure. Thanks for having me.

Podcast: Play in new window | Download