The Bots Are Coming - Don’t Be Afraid!

So first off, as Catherine mentioned, I am a lawyer too. I have 22 years of experience, and I also founded LawDroid, this legal tech company, approximately 4 years ago, and really enjoy it, really am excited about what's going on right now. In the industry. And so the title of today's presentation is "The Bots Are Coming, Don't Be Afraid: Automating Your Law Practice with Intelligent Automation." So that phrase intelligent automation I use to reflect that we could automate, but we could also use AI, or at least classical AI, not necessarily machine learning. And we're going to talk a little bit about what the difference is between those two in a minute. So what is artificial intelligence? Without going down the rabbit hole, it's basically a way of making machines smarter, right? And so, you know, like the title of the presentation conveys, there's this a little bit of anxiety about how smart we can make machines. And so the bottom line for everyone is we're not there yet. There's really no worries about you know, robots taking over jobs or anything like that. I also want to give a shout out to Baron Henley. That presentation that he had was amazing in terms of the detail. And so the question about whether or not lawyers should code, I think that if you want to, you can, but if you don't, the fact that you come to things from a lawyer's perspective really casts a lot of interesting light on AI and automation. So one new development that I wanted to point out to everyone that is super exciting in terms of AI, not that we're applying this specifically in legal, but this past week I read an article which originally this was coming out at the beginning of the year. But, um, AlphaFold, which is run by— well, supported by Google, uh, basically they were able to crack how proteins fold themselves. And so this doesn't sound like it really applies to legal right now, but what it has to do with is allowing machine intelligence to predict how proteins are going to fold. And so it's pretty abstract, but when you look at this You could see, you know, green and blue, and they kind of twist around each other, and it's a little confusing here, but the green one is what the scientists were able to derive from experimentation. So think of it as kind of like the green part is the actual physical representation, like if you were to look at bones in an X-ray. You know, you would see that that's where those things actually are. The blue part is actually the computational prediction. So this AI that has been developed by Google has gotten so good that it's in the 90th percentile of being able to basically predict the structure of these proteins. Now You know, what does that mean? What's the money value of that? It means that things will be much easier in the future to predict how proteins will fold, and that allows us to develop drugs better, possible vaccinations better, and also to roll out gene therapies and things that would help a lot of people get well or maybe prevent disease. So this is actually a challenge for science over the past 50 years, where it's been mostly done by experimentation. And what is happening here is that the computers are going through various examples of how proteins fold and then being able to infer and derive from that how things are how things are folding, and then understand how they might fold differently in the future. So, that type of inference is something that could be applied to the law, and looking at different examples of, let's say, cases where negligence is found, or cases where negligence is not found, and then to infer or derive some meaning from that. Okay, sorry about that. Um, so the next slide here, machine learning, which we hear a lot of talk about, that is broken into 3 different categories. And so I wanted to just dive into it very quickly to give a little meat to that. There's supervised learning, there's unsupervised learning, and then there's what's called reinforcement learning. Which is how machines learn how to simulate human intelligence. This first one here is probably a slide that you've seen before. It has, you know, blueberry muffins intermixed with chihuahuas, and it's a— it's an example of supervised learning. So what would happen is there would be thousands of pictures of blueberry muffins and thousands of pictures of chihuahuas. And then someone would go in there and painstakingly label them as either a chihuahua or blueberry muffin. And then by doing that, just kind of like teaching a baby that this is this and this and that over there is something different, you know, that's a train, that's a truck. It would teach the computer model to distinguish between these two. And so by supervising the learning of the machine by classifying different items, the machine would be able to infer from that how to classify. This here is an example of reinforcement learning. And you might be wondering, what is this that I'm showing you? And basically what it is is a game that the machine learning developers created. It's a game where the blue team is attempting to hide from the red team. And so it's a video game, but they, uh, they sprinkled some, you know, AI into this game so that over time the machine would repeat the game over and over again with the blue team trying to hide from the red team. And the red team learned some tricks over time about how to find the blue team. So for example, on the left side, you can see the little red guy running up what he's built. He basically put a wedge to create a ramp so that he could run, run up and get into the room where the blue people are hiding to catch them. And you can see that the blue team has gotten 2 blocks and they blocked entry into that room. So what does this all mean? It means that machines could actually learn from playing games and repeating them, from finding the errors that they make. Where at the very beginning of this game, these two teams were probably tripping over each other and it wasn't making a lot of sense in what, what a hide-and-seek game would look like for us. But over time, after repeating the game over and over and over again, it gets quite adept at it. So that's called reinforcement learning. Um, and you might have heard of GPT-3, which is a type of AI that is used to predict words by looking at the context of words. So that sounds a bit abstract, and this whole presentation is a bit more abstract than Baron Henley was talking about. But GPT-3, it looks at, it reads text, and then it kind of, it's intended to predict, okay, what's the next text that comes after that? And it's gotten pretty good at doing that. And in fact, what it can do is it can take an input like a legal clause written in legalese, and then it could take that text and transform it into plain English. And it's able to do that by going over, again, thousands and thousands of examples, which it has built into it. So just keep in mind, this kind of, you know, artificial intelligence, it doesn't like answer complex questions for you. It's not like you could ask it, Okay, how do I build a rocket ship? And it's gonna give you step-by-step instructions. It's not there yet. But what it does do is that you can give it some more complex wording and then it could boil it down to something that is more intelligible at a certain grade level reading. It could also do some other things like, for example, you could give it several, different plays by Shakespeare, and then it could kind of create, extrapolate its own writing based on that. And it's also written some philosophical papers, some of which are kind of interesting to read. So that's a pretty exciting development in technology. So how is this used, boiling this down to something practical today? Well, Ross is using something that uses machine learning where It takes a natural language inquiry, like what is the standard of negligence in New York in 2004, after 2004, breaks it down in the timeframe, breaks it down in the jurisdiction, the concept being standard of negligence. It could understand that, retrieve information that's relative to that, and then rank it according to its likelihood to answer the question that you're asking. Kira is another technology, a legal tech company that works in a similar way, but the input in this case is a document. You know, so like let's say you upload a contract and Kira has built into it numbers of provisions. So again, you have example, example, example, example. The machine knows these examples, then looks at your contract and then identifies provisions that are similar. To these examples and then maybe highlights things that are different. Like, why is this different? This does— this looks non-standard. The implication of it being non-standard means X, Y, and Z, you know. So these are, these are ways to throw up before a lawyer just some, um, I guess, easy reference points so that The lawyer can focus on the more sophisticated issues, but in identifying the, the route or, um, you know, standard provisions, you could use the AI to identify that for you and give you some insight as to whether or not you want to negotiate them differently. So there's a number of legal AI companies. These are just a a handful of them. I believe that this, uh, map was created a couple of years ago. Things have exploded since. Investment in this area has gotten dramatically increased. And, um, as, as everyone knows, because of COVID things have gotten even more online and people are relying even more on, um, on computers and AI-type systems. So speaking of AI, there's the machine learning that we were talking about before, but there's also classical AI. Now classical AI is the kind of stuff that we were walking through with, uh, with Baron Henley. You know, that's an expert system or one that uses conditional logic, the if-then statements that Baron was describing within Word, which that, by the way, is amazing because you don't have to spend a ton of money to use Word. We all have it already. But some of the legal tech companies that provide a similar kind of expert system or form-filling ability is— Neotologic is one example, and they have more of a visual approach where you can create blocks that lead to each other, and they have different formulas. Like, one can ask a question. Another one can take input and then run it through a formula to arrive at a certain answer. Documate is another example of a technology that makes form filling a little bit easier, a little more visual and easy to, to do. My company, LawDroid, We have a bot builder where you can create a bot that asks questions. When it asks a question, it saves that answer in an attribute, and then it could also create documents. But we're here to talk about chatbots. So what is a chatbot? It's a computer program that can automate a conversation or task. Now, that blew me away, uh, several years ago when I first I encountered the idea of a chatbot. All of you probably have heard at this point about Joshua Browder and his Do Not Pay bot. And what he did at the time was allow people to challenge parking tickets by simply going through a question and answer that was automated. So the big deal that I think chatbots make is that all of that form-filling automation that we were looking at when, when Baron Henley was describing it. What chatbots do is it provides a friendly interface, provides a friendly conversational interface that can work with clients or lawyers to gather information. So how do you make this consumer-friendly? And here's some examples of it that my company has created with some of our clients. This one is called PapBot, and I created this bot for Patrick Pallas. He does workers' compensation law in Washington. And so what this bot does is it goes through a series of questions where it essentially identifies different legal issues. So like, is there an issue with the panel of doctors that was presented to the, the client? Uh, is there an issue with them not being paid while they, while they've been injured? Is, you know, all of these different questions are asked to try to identify issues that Palace Law might be able to help them with. But all of this is actually provided free of charge. There's no obligation. To the consumer to go through this interview, and it takes about, about 10 minutes at most. It asks over 2 dozen questions, and at the end of it, they get a report that, by using conditional logic, is customized to them. You might have this issue. Here's some more information about this. Here's a form that you might be able to use to file a claim. And so it could be a way of automating this interaction with a client or potential client and then providing them with useful information. This bot is on Erin Levine's website. She does family law, and, uh, it essentially reaches out to website visitors and automates that conversation with them to see if they might want to Schedule a consultation. So that's another way of applying chatbot technology is for that outreach to automate reaching out to potential clients. And you don't have to have any staff dedicated to it. You don't have to be paying a third-party provider to have some, um, you know, call center tasked with answering these questions. You could basically build the script with us and then automate it so that it's your best representation of your firm. Another example is this divorce bot that we created with West Tennessee Legal Services where someone who qualifies for legal aid, but maybe there aren't enough legal aid attorneys available at that moment to help them. What would normally happen is they would just get turned away because the supply of lawyers is tapped out at the legal aid organization. Well, if they happen to have an uncontested divorce, they can use this system, which actually you could interact with text, you know, typing in text, or you could actually speak to it and have it speak back to you. And it could walk you through all the steps that you need. I'm sorry, this is the workers' compensation bot we created. We also created a divorce one, but it walks you through all the steps you need to fill out your workers' compensation form in this case. And this one is actually, um, we coordinated with the Bureau at the state level in Tennessee for rolling out this workers' compensation bot. So some concerns that might come to mind when you hear about bots and you hear about it being applied in legal is, you know, do people really want this at all? And don't people still value the personal touch and handshake and speaking face to face, which that sounds pretty dated in COVID times, right? But we all naturally feel that, you know, the human touch is important, and it is. But when people were actually surveyed and there was data obtained for checking out these questions, the most important thing for everyone was reaching the desired outcome by 68%. It was overwhelming. And I think we all feel the same way. If you have a problem with your cell phone company and you want to get answers, does it really matter that you get a person on the phone and you have that personal touch? It might be nice, but the most important thing is getting the outcome that you want, the result. And then followed that by that was of course the speed, convenience, and the ease of the experience. And the number one reason why people would consider using a chatbot over a human being is to get an instantaneous answer if possible. And that's not to say that there always be an answer. There might be a question that has not yet been asked. Or might be asked in a way that the machine doesn't understand yet. But what you can do is have a fallback where it captures that question and then emails it out to somebody who can answer it. And what's even better is once you get these new questions and you provide answers to them, you could build that into the system to make it even smarter, right? So it saves us all time, and time is money. So why to automate? Well, Clio, um, back in 2017 really, uh, lit this up for a lot of people, um, that there's this missing 6 hours in the day, non-productive, non-billable hours. And it essentially broke down to a lot of administrative tasks that that 6 hours was taken by. And so how do you make that something that maybe you're not wasting your own time on, put your own time towards more billable hours and take the administrative stuff and task it for technology if you can. So that's one of the ways that bots comes into this. You could have a bot, um, collecting information from clients. It's kind of a virtual paralegal. You email the client a link to, to the bot and it could interview them for your intake. You could have it on the front end doing the marketing to try to get more clients and things like that. So when we're creating a bot, what are the different things that we want to consider? The number one step, the first step, is why are you doing this? You know, what is the bot's purpose? And it's really important to think about what the, what the the purpose of the bot is, because it's going to guide everything that follows, including like metrics and how you measure success. One type of bot is an informational bot. Um, this is another one that we created. This one is for Tennessee Alliance for Legal Services, Renter Defender, that provides information to tenants about their, their rights as tenants. So here there are various categories of information that the bot can provide to the tenant. If they're in a situation where they need help paying the rent, click on that, takes you down a, a conversation about how they can pay the rent. I believe it provides them resources where they could click through and get possible possible support if they qualify. If they need to fill out the CDC declaration letter, which is one that allows them to be protected from eviction through the end of the year, they click on that button and go down that conversation tree. Essentially, you could automate the provision of a lot of information. So again, this is kind of mimicking how people in— how the public interacts with legal aid. They have tons of questions. There aren't thousands of people to answer all those questions all the time, so you could have something like this tasked with providing those answers. Another type of bot is a task bot. It's one where you can have it laser-focused on one particular type of task. It could be booking appointments, taking orders, drafting forms. This is the Do Not Pay bot that I mentioned before that Joshua Browder created, kind of the granddaddy of them all. And earlier on he had it so when there was the Equifax breach, you could click on that and get more information about suing Equifax for the data breach. And it actually auto-created a complaint that you could file. Another example is VisaBot, which actually has been acquired by DoNotPay, and it was a similar type of approach but for immigration, and you could answer a bunch of different questions and find out what your probability was of qualifying for certain visas. So again, getting at those questions that consumers have for lawyers, which is, What are they interested in? How likely is it that I'm going to get this visa? Right. That's the result that they're looking for. There's a lot of details that go into it, but the result they're looking for is what are my chances of getting this visa? And that doesn't mean to exclude lawyers. This could actually drum up a lot of business because if they could qualify for it, well, then that means that they would be a fantastic lead for you as a lawyer. And then you can assist them in completing the legal services they need to get that visa. So when you're thinking about the purpose, you want to define metrics. Uh, one measure, like let's say we were creating the, the lead bot to create additional leads for your firm, we would want to look at, okay, what are the total number of leads, uh, divided by the number of website visitors multiplied by 100. That would give us a sense of how, how people are converting from visitors to leads. And then we might take a look at how they're converting to clients. And the idea is that you keep these metrics and you track them based on how the bot is performing, and then you gauge your changes to the system. Like maybe you want to change a question from being articulated one way to a different way, and that could improve your conversion rate. But the important through line here is you want to measure things and make decisions based on data and measurements rather than just, you know, seat-of-the-pants types of judgments. So to review, you know, in step 1, what is your chatbot's purpose? What are we trying to achieve with the chatbot? And what are the metrics of success that are tied to that purpose? The second step to think about is this bot that we're creating. Where is it going to live? There's a lot of different options. Some of these don't exist anymore, uh, because things change quickly. But let's say it could live on Facebook. It could be tied to your Messenger in Facebook. And so that might be maybe family law. A lot of people hang out on Facebook. You could identify through Facebook where they live, what kind of work they do, their income levels, things like that, and you can advertise to them specifically. So that might be ideal messenger for divorce clients. Another example is Slack. Slack, a lot of people use it for business. I know I use it every single day. And so if you want to reach people where they're at, well, that might be a way of generating business with business clients, is that you have a bot which maybe it automates the creation of a non-disclosure agreement and you make that available to people. Another is WhatsApp. It's very highly used by people outside the United States. And so that might be a way of drumming up business for immigration, where you could, you could task it with answering people's questions about different types of visas. So, uh, where's the chatbot going to call home? Who's your intended audience? I think that's the most important thing of those questions before. And you kind of work backwards. If your intended audience is business, well, then it might be something like Slack or maybe Skype for Business. And where does your audience live? Because you can now target everyone down to not only geography but income and occupation and all kinds of other stats. So step number 3 is, what is your bot's personality going to be like? And so, you know, you could go crazy You can make it kind of a silly bot. This is an example of a silly one that doesn't exist anymore called Poncho. And Poncho, it was an app on your phone and its purpose was to give you kind of like, you know, what's going to go on with the weather. And you would ask questions about the weather, but it would also have this dialogue where, you know, it would play around with you. And be silly. And so I'm not suggesting that you have a legal app or bot that does this. This is an example of more of an extreme, but it's not just being silly. It's about the things that make up personality. So it could be the name of your bot, the visual style of your bot, the backstory, and its conversational tone. So Erin Levine that I spoke of before, she called her, she calls her chatbot Bam Bam, right? And so Bam Bam literally has a heart on its sleeve, as you can see, and is a cute looking chatbot and is meant to kind of break the ice and make things a little friendlier. I mean, people when they're dealing with divorce and family law issues, understandably are a bit stressed out, and so having a bot that is a little more inviting is a good idea. This one, BillyBot, this came out a few years ago, but basically these, uh, this company in, in the UK, they not only created the bot which you could interact with about I think getting referrals to solicitors in the UK, but they also created a personality on LinkedIn, like it had its own LinkedIn profile and it would tweet sometimes about, hey, things are so stressful. I'm a new junior clerk and these people are driving me nuts. I'm trying to handle these senior solicitors, but they're being mean to me or whatever, and try to make it feel more human and more interesting that way. And PatBot, as you can see here, modeled, you know, with a kind of cartoon after Patrick Pallas. A lot of people that come to Pallas Law, they've interacted with Patrick either through listening to him on the radio or through some videos that are online on YouTube, so they kind of, you know, connect a face with a name. So for step 3, what are you going to name the chatbot? And it doesn't have to be, you know, Law Firm Chatbot or the Smith Law Chatbot. It could be something more fanciful like Fam Bam, but it doesn't have to be like that. It could just be somebody's first name, but you want to think about that. Makes a first impression. What does it look like? Does it match the branding of your law firm? But beyond matching the branding of your law firm, you know, does it look friendly, inviting? Because legal issues for consumers is a bit off-putting. So how do we invite them in, into the conversation? Does your chatbot have a backstory? Doesn't have to. Doesn't have to be as involved as, or silly as Billy Bot. But, um, maybe, maybe it does. And the conversational tone, so how the bot actually speaks to people, is an important choice as well. And so there's also like different structures of how the bot talks to people. Vertical is a very guided interview, question, question, question, question, then maybe it asks yes or no, and then it branches it off based off yes or no. And leads you down different trees. The horizontal type of conversation is more open-ended. So it's like you might be going down a vertical path, but then somebody asks a random question. The consumer says, well, what about this? With natural language processing, we can now like capture the intent, hopefully, of that question and then serve them up the answer. And so this is a bit of a diagram of how we do this. We have a question, we run it through a natural language processor like Google, and then it structures that unstructured data and runs it through a database, which is basically our database of questions and answers that we built into it, and provides an answer. And again, like I mentioned before, it's not always perfect. We might not have the answer. But the beauty of it is that for all those unanswered questions, you feed that back into the system so that it learns these new questions and the— excuse me, learns the new answers to the questions, and then it just gets smarter and smarter. So for step 4, when we're looking at the structure of the conversation, we want to structure it so that we have a disclosure, uh, disclaimer, If somebody's talking to a bot, "Hey, this is a bot. It's not a real person." There might be, you know, you don't want to have that expectation built that they're speaking with a human being. In fact, in California, it's the law that you have to disclose that it is a bot. Have you provided deliverable to the user at the end of the conversation? And created a path to them getting that deliverable that is easy for them to, to experience. So there's different ways that you could create a bot, um, different platforms. There's one called Flow XO that is a pretty visual type of platform to, um, you know, create these different modules for questions and answers and building in some modicum of AI where it uses at least the natural language processing to try to understand questions that are asked in natural language. ManyChat is another great one. MobileMonkey is pretty good. HubSpot, if you're using HubSpot for your marketing, actually has its own marketing bot you can build into it. Lodroid, we also have one. Which, by the way, if you email me, and I'll show you my email at the end, I could give you advanced beta access to our platform so you could test it out. And within the platform, it looks like this. You create these different steps, you know, how are they doing, options for which options, options they want to choose. Maybe one of the options is to pick a joke, This is a kind of a fanciful bot that I created with Yvonne Naff, where they could pick a joke and it would try to lighten the mood during COVID and make them smile. And so we would also like ask them at the end of it, did we make you smile? Did you like the joke that we ran by you? Here's some other parts of it where we would capture their name, and we could also set like the, the branding and look of the bot, and you could customize your bot to however you like it. Um, oh, one other aspect here is that you could also monitor the conversations, and if you want to jump in and have a human-to-human conversation, you could actually take over the conversation. And speak directly to the people that are visiting your website. But getting back to, uh, Cheerbot, we were able to break down, and you could also gauge these analytics. So the way people ask— answer questions, you could visualize that and you could gauge what the actual percentages are of different types of responses, which again, for tracking those metrics, is very important. To know if your bot is accomplishing its purpose. You could also, you know, track visitors and see where there's peaks and valleys. At the height of it though, this was a limited release for CheerBot, there was 206 people that tried it out, and thankfully the majority of them we made, we made smile, so that was good. So for step 5, building your bot, do you prefer to do it You know, with code, kind of like Baron's approach, or do you prefer to do it visually? Well, visually there's a few different platforms to use, including Laundroids. And, uh, will your chatbot need to integrate with third-party services? Uh, you want to look for one that can integrate with Zapier. I'm sure all of you are familiar with that, and it basically opens the door to a lot of very cool integrations to integrate your chatbot with. Zapier, and then, uh, you know, what language support might you need? Do you also need to have bilingual in Spanish, for example? And so we are going to have a breakout session in just a minute, but, uh, save your questions for that. I'd be happy to answer them. And, uh, this is my contact information. You can follow me on Twitter @lawdroid, and my email is tom@lawdroid.com. And if you do want advanced beta access to the chatbot builder, I'm happy to provide that.