Latest / Technically Sound Strategy / Episode 48 - AI, Security, and Sustainability With Andrew Crosley

Transcript

Episode 48 - AI, Security, and Sustainability With Andrew Crosley · Technically Sound Strategy

• 0:01
  Welcome to Legal Marketing Radio, a production of LaFleur marketing.
• 0:05
  This week I'm joined by Andrew Crosley from Crosley Law Firm.
• 0:08
  Andrew is an intake coordinator and legal AI implementer who's bridging
• 0:12
  the gap between traditional legal workflows and emerging technology.
• 0:17
  In this episode, we talk about how to build secure AI tools for intake and law
• 0:20
  firm optimization, what sustainability looks like in a fast changing environment,
• 0:25
  and how operations leaders can approach AI thoughtfully and responsibly.
• 0:31
  Andrew, it's great having you here with us.
• 0:33
  You and I have talked a lot about different AI projects.
• 0:36
  I think we're working in the same, within the same tools to some degree at least.
• 0:41
  But you've done some really interesting things.
• 0:43
  I really appreciate you taking the time to, to talk to us here.
• 0:45
  So it.
• 0:47
  Tell me a little bit about some of the tools that you built, some of the services
• 0:51
  that you built in-house for Crosley Law.
• 0:55
  Yeah, my name's Andrew.
• 0:55
  I work at Crosley Law.
• 0:57
  I am an intake coordinator here.
• 0:59
  But for the past year or so, I've been working on how to integrate AI
• 1:05
  into our firm to make our processes easier and more automated so that
• 1:12
  people spend less time doing what they don't need to spend time doing.
• 1:17
  So some of the projects I've been working on, one of them is
• 1:19
  like a call intelligence system.
• 1:21
  So it takes our call, audio transcribes it, and then it pulls all
• 1:26
  of the information um, to do lists.
• 1:29
  It'll populate things in our case management software in the future.
• 1:35
  And gives us items that we can use to follow up.
• 1:40
  Another cool thing that it helps us do is get an email out to a potential
• 1:44
  new client as soon as possible.
• 1:45
  So instead of maybe a couple hours, we have the draft ready
• 1:49
  to go in a couple minutes.
• 1:51
  Another project that I've been working on here is taking all of Texas legal
• 1:56
  codes and statutes and all of that and turning that into a system that
• 2:01
  AI can reference to provide answers to attorneys about specific law questions
• 2:08
  that they have, and it'll come back with the exact code and where it found it.
• 2:13
  That's just two of a handful of projects that I've been working on,
• 2:16
  but I'd love to get into any of them.
• 2:20
  I think that's awesome.
• 2:20
  I mean, it, It seems like your approach has really been to first
• 2:26
  see where there's a pain point and then when you see that there's a pain
• 2:29
  point, you start to build something.
• 2:32
  I think that one, for a lot of people listening to that, it's a little scary.
• 2:36
  Because in a law firm, you're dealing with medical records and
• 2:39
  you're dealing with all sorts of, very private personally identifying
• 2:42
  information potentially HIPAA content.
• 2:46
  How are you approaching security around, introducing these elements to an LLM?
• 2:53
  Yeah that's A really good point.
• 2:54
  The security issue, personally identifiable information that
• 2:58
  it's pretty sensitive subject.
• 3:00
  So one of the things that I've done is use, and so LLMs AI uses context, right?
• 3:07
  One of the things that I've done is scrubbing whatever context we
• 3:11
  want to give to an AI model, like chat, GPT using some local models
• 3:17
  and local means that it runs on your computer and only on your computer.
• 3:22
  So what that can do is it can pull names, dates of births social
• 3:26
  security numbers, anything like that.
• 3:28
  It'll scrub all of that out.
• 3:30
  So that the AI is only getting information relevant to the case, but not containing
• 3:35
  any personal information of our clients which we've seen some success in.
• 3:41
  And we can get more into that if you'd like.
• 3:43
  Yeah.
• 3:44
  What's Required to build that out?
• 3:46
  That's not something I could just run on a laptop, right?
• 3:48
  Our firm invested $4,000 in a a computer that can run the,
• 3:54
  some of the smaller models.
• 3:56
  There are hundreds of models that you can download online.
• 3:58
  Most of them require pretty big infrastructure a lot of compute to run.
• 4:05
  But.
• 4:06
  There are smaller models that can run on a $4,000 machine.
• 4:10
  That can essentially, if you tell it what to do, and if you have the framework
• 4:15
  it can do what you want it to do.
• 4:17
  Which in our case is essentially scrubbing any information that we
• 4:22
  don't want to feed to the cloud.
• 4:25
  Oh, that's sweet.
• 4:26
  Yeah.
• 4:26
  And so the other project that you talked about is less sensitive, right?
• 4:29
  So
• 4:29
  Texas code manuals for machinery, stuff like that.
• 4:33
  Like that on the other hand is not sensitive.
• 4:34
  It's a large body of information, so walk me through the advantages
• 4:38
  of one iteration versus the other.
• 4:41
  And ultimately we want to get to the point where we're talking about how
• 4:43
  accessible is this to a firm, right?
• 4:45
  Is this like crazy unaccessible stuff that, requires a bunch
• 4:48
  of data scientists to build?
• 4:49
  Or is this more within reach?
• 4:52
  Yeah.
• 4:52
  I know every firm varies.
• 4:54
  Some firms have slightly different workflows.
• 4:59
  But this technology is accessible to everyone if you have the
• 5:02
  right tools to implement it.
• 5:03
  So for me, I've, I spend maybe I would start off a couple hours a week working on
• 5:09
  this stuff and, that was enough at first.
• 5:13
  Advantages to local, it's all running a local machine runs entirely in-house.
• 5:19
  So none of the information that you feed a local machine goes anywhere else.
• 5:23
  It doesn't get uploaded to some massive server where all of the
• 5:29
  high level AI compute takes place.
• 5:31
  But one of the disadvantages to that is that those models are.
• 5:35
  Dumb, if you will.
• 5:37
  They're a lot less smart.
• 5:39
  So you have to keep that in mind when approaching the local side of things.
• 5:44
  And then the cloud computes and models like Gemini, jet, GPT, what we know that.
• 5:51
  Infrastructure is it's more on like a pay per use side of things.
• 5:55
  So local ai, once you have the hardware, you can run it unlimited
• 6:00
  times for free and offline.
• 6:02
  Like I could unplug myself from the internet and still run these programs,
• 6:06
  but when we were sending things to the cloud, you can't do that.
• 6:10
  So I know there are, security measures that you can do with some
• 6:15
  enterprise solutions with Chat GPT specifically that can protect some
• 6:21
  information, but it's never 100% secure.
• 6:25
  Which is why I've tried to lean towards local AI when we can and when we should.
• 6:31
  That makes sense to me.
• 6:32
  And I think like mitigating the risk.
• 6:37
  Is such an important factor for a law firm in determining
• 6:41
  which tools they're gonna use.
• 6:42
  I think Cloud just came out with a medical HIPAA compliant offering.
• 6:47
  I think that Chat GPT either just came out with one or has one in the works.
• 6:51
  And I'm sure Gemini is in that ballpark too.
• 6:53
  But like you said.
• 6:55
  As soon as that information leaves your network and it touches the internet,
• 6:59
  you've just increased the risk.
• 7:01
  And even if they're, their models are tight and if their security
• 7:05
  profile is great, it's still.
• 7:09
  Outta your system.
• 7:10
  And the difference being for someone to access the private information that you
• 7:14
  have on your local machine, not only would they have to break in, but they'd
• 7:18
  have to break through your login and they'd probably have to break through the
• 7:21
  encryption that you have on your drives.
• 7:22
  And so the difference in those risk profiles is so dramatic.
• 7:26
  It makes sense.
• 7:27
  How you're doing it makes sense.
• 7:28
  And it's, from my perspective, I go to a lot of conferences
• 7:30
  and spend a lot of time.
• 7:32
  Perusing through the AI offerings of different places and I hear what the
• 7:35
  offerings are and see what they are.
• 7:36
  And.
• 7:37
  They almost feel like they're outta reach for the average firm.
• 7:40
  That the technology is so crazy that they couldn't pull it off.
• 7:43
  But looking at the things that you've built and seeing how you've
• 7:46
  built it now, one, I don't think that just anybody can build it.
• 7:49
  I think that you are a special case and that you are able to both
• 7:52
  understand the business side of it and also able to understand the AI
• 7:56
  side and how to build things in ai.
• 7:58
  I don't want to I don't wanna overstep that, but the fact that you were able
• 8:02
  to jump in and start building this stuff right away and then have it be usable for
• 8:05
  the firm I think is really outstanding.
• 8:07
  You touched on one thing.
• 8:08
  Talking about your voice tool that you built and how that will plug into.
• 8:14
  Case management.
• 8:15
  Just wanna clarify that the only reason that it doesn't now is because
• 8:18
  the because the case management isn't capable of taking it Correct.
• 8:22
  In the future.
• 8:22
  In the future, I think that'll be a different story.
• 8:24
  So the software is already built in a way that extracts all of the information,
• 8:30
  whether it's a phone number from the metadata of a call so like the call
• 8:35
  details or it's somebody's name.
• 8:38
  Anything like that, it's getting pulled and structured in a way
• 8:42
  that will be very easy to plug into a case management software.
• 8:47
  So the goal is that intake would get off the phone and everything would be
• 8:52
  in the right place and ready to go.
• 8:55
  There wouldn't be any manual entry of those preliminary details
• 9:00
  that you're getting on the call.
• 9:01
  So it would, will make notes.
• 9:03
  It's already making notes, but we, intake still has to, at least at our firm plug
• 9:08
  those into our server based platform.
• 9:11
  And so once we have and I know most firms are working in the cloud specific, with
• 9:16
  a cloud-based system using something like we have here, you would be able to
• 9:22
  easily plug in all of the intake data straight into the case management software
• 9:26
  without any entry from the intake person.
• 9:30
  It would be all automatic and almost instant.
• 9:33
  That's crazy.
• 9:34
  Now, I think that I know for myself, and I'm sure a lot of people who
• 9:37
  would listen to this have had experience with voice to text.
• 9:40
  Tell me how that works, because the accuracy of voice to text in
• 9:43
  our experience for many years is a little bit different than ai.
• 9:48
  Voice to text is now.
• 9:49
  Yeah, so I know in the past year there have been some crazy
• 9:53
  improvements with the voice to text.
• 9:56
  We currently use Gemini and we use Whisper as a backup, which Whisper is
• 10:02
  an open source software developed by OpenAI, which is the Chat GPT company.
• 10:07
  And.
• 10:09
  It's much better than what you're used to on your phone.
• 10:13
  So if you get the apple transcript of a voicemail, you'll notice that the word
• 10:17
  error rate is the term, the word error rate is much higher on those systems and.
• 10:24
  At least recently, I know last week Nvidia released a new
• 10:27
  automatic speech recognition model which I haven't tested out yet.
• 10:33
  I plan on doing it, but almost every week there's, there are new updates on
• 10:38
  the not just speech to text every field, but speech to text has seen a lot of
• 10:42
  crazy improvements in the past year.
• 10:44
  It's crazy.
• 10:44
  So it's, and it is funny, like I use Whisper as an app.
• 10:47
  Like the Whisper app on iPhone and on and on your Mac.
• 10:52
  And it is such a game changer.
• 10:53
  It figures out my punctuation, it figures out my writing style.
• 10:57
  It figures out unusual names that I'm, of people that I'm talking to.
• 11:01
  It's crazy.
• 11:03
  And I find more off, more and more, I'm not even using the keyboard.
• 11:07
  And that's such a dramatic change from six months ago.
• 11:10
  Because I even then, if I'm driving or doing something where I don't
• 11:13
  wanna look at my phone, sometimes I'll hit that little microphone and
• 11:16
  the results are disproportionate.
• 11:18
  It's faster too to just talking to your phone than type.
• 11:21
  One, one thing about calls is that, they can sometimes be muddy like
• 11:25
  somebody's calling in and it's pretty difficult to understand.
• 11:29
  But we've gotten to a place now where, i'll be reading through a
• 11:33
  transcript and I'll see a word that maybe doesn't quite make sense.
• 11:38
  It's pretty rare, but it happens.
• 11:40
  And then I'll go back into the call and I'll listen to it and I won't know what
• 11:44
  it means, and the, I'll listen to it over and over again and I'll come to the
• 11:48
  same conclusion that the AI's coming to.
• 11:50
  So it's pretty, pretty phenomenal technology that we're able to use now.
• 11:55
  That's wild.
• 11:56
  So you've built the things, you've seen the needs.
• 11:58
  You've identified the need, you've built the tools to address it.
• 12:02
  And then I think the next challenge is, building something that is going to be
• 12:06
  incorporated into the process of the firm.
• 12:08
  So like, how do you.
• 12:10
  Decide when to stick a fork in it and call it done and call it a product.
• 12:14
  Versus, another mistake I often make is to continue to work on it and then,
• 12:19
  and I'm just like, but it could be better and it always could be better.
• 12:22
  So how do you
• 12:23
  Yeah, and that's something that I also struggle with.
• 12:25
  I. Over optimizing is what I would call it.
• 12:29
  So what I do is I have people at our firm test things out.
• 12:33
  So as I'm rolling out updates to a new software people will be testing it out
• 12:39
  and I'll regularly ask for feedback on it.
• 12:43
  And once I feel like it's at a place where maybe it could be better, but it is.
• 12:48
  Serving its role, then I'll move on to the next thing.
• 12:51
  I find the next thing to optimize.
• 12:53
  And that's been a pretty recurring cycle there.
• 12:56
  Nice.
• 12:57
  And you can always come back, right?
• 12:59
  If someone says, Hey, this isn't working how I expected,
• 13:01
  or Can we make this improvement?
• 13:02
  So you're using full like version management and all of that
• 13:06
  stuff, and that's incredible.
• 13:08
  What advice would you have then for, you're not in every firm and,
• 13:12
  you're a little bit of a unicorn, but there's a lot of firms out there and
• 13:16
  there's a lot of very smart people who are capable with technology.
• 13:19
  What advice would you give to intake teams, to operations managers,
• 13:24
  COOs to guide them on this journey?
• 13:26
  Yeah.
• 13:27
  So I think having someone who has dedicated time to work on projects
• 13:32
  like this is really important.
• 13:33
  It's, it won't just happen.
• 13:35
  So whether that's hiring someone or giving somebody a couple hours
• 13:38
  of free time, if they're familiar with the technology, that's one
• 13:42
  piece of advice that I would have.
• 13:43
  Another piece of advice is that most employees don't have or employees have
• 13:49
  different ranges of understanding.
• 13:52
  So as you're implementing something, some people.
• 13:55
  Might be super familiar with AI and some people may have never used it.
• 13:59
  So teaching your employees each person's a little different.
• 14:03
  But getting everybody familiar with it is one of the big steps.
• 14:08
  And then I think also having somebody dedicated to working on
• 14:11
  projects like that is another step.
• 14:14
  Yeah, I think so too.
• 14:15
  I think one of the things that I've run across is this, again this kind
• 14:19
  of fear of something new, right?
• 14:21
  And then there's a little bit of obfuscating, obfuscating the landscape.
• 14:25
  In AI where there's all of these AI companies that say that they're AI
• 14:30
  companies and they're really just putting a wrapper on chat, GPT on
• 14:34
  Gemini, on some of the big models.
• 14:36
  And I think that's also really important to keep in mind is that like, when it
• 14:41
  comes to LLMs, I mean there's a ton of models to choose from, but everything
• 14:45
  is built on those models, right?
• 14:47
  So all the functions and then those functions, they're only
• 14:51
  limited by your imagination.
• 14:53
  But at the core, it's still those models that are doing it.
• 14:55
  And I think, I know for me and I'd love to hear what you think, but for me, one
• 15:00
  of the biggest unlocks was starting to use an integrated development environment.
• 15:04
  And you actually introduced me to Cursor was using VS code and Tray before that.
• 15:10
  Tell me a little bit about that process for you, because this isn't
• 15:12
  something you went to school for.
• 15:13
  This isn't something that you started using.
• 15:15
  Like how'd you get plugged in there?
• 15:16
  And then what changed when you started to use those sorts of tools?
• 15:20
  I was introduced to it just by looking up on Google, the how do I do this?
• 15:25
  And then it's that's how I got introduced to Cursor, but now that
• 15:28
  I'm using Cursor's, really cool.
• 15:29
  So Cursor is an a, a platform that can look through your files
• 15:35
  and make changes to your files.
• 15:37
  And it also has a chat bot that you can talk to.
• 15:40
  So it becomes familiar with the project that you're working on and it knows
• 15:45
  how to do most of what you want to do.
• 15:48
  So I don't really have much coding experience.
• 15:50
  I took CS one in college and I was a finance major.
• 15:55
  And it, the learning curve, there's definitely a learning curve but
• 15:59
  it's not as steep as it once was.
• 16:01
  Not nearly as steep.
• 16:02
  Anybody can get familiar with it in a matter of weeks.
• 16:06
  That's awesome.
• 16:06
  The other thing that I found, from my time working with Cursor and the
• 16:09
  unique thing that I have that I bring to the table is just my comfort on
• 16:13
  the DevOps side, on the backend.
• 16:15
  And without that, I think it would've.
• 16:17
  Had a different path but you could learn so many things by using it.
• 16:22
  You go in there and you say, I wanna build this thing, and it will
• 16:24
  help you build it and it will give you the instructions to build it.
• 16:27
  I'm pretty comfortable on the DevOps side, so those things are
• 16:30
  not scary for me to get in touch.
• 16:32
  I think you brought a comfort level.
• 16:33
  Building, building a PC and building the functionality within it and
• 16:37
  work and finding your way around, setting up the OS and all that stuff.
• 16:40
  But I find that because of your experience, that makes me more excited
• 16:43
  to try to play around with a local model.
• 16:45
  And then also just, I don't know, I guess my point being.
• 16:49
  Using some of these tools to dip your toes in and then recognize
• 16:53
  the tools are actually gonna help you become more familiar.
• 16:55
  And I think part of that is just a decision to do it.
• 16:58
  I know for me, like I had a month or two of reservation to say I was
• 17:01
  really starting to hit the road.
• 17:03
  Do I have to learn how to do everything?
• 17:05
  New again.
• 17:06
  And then a couple months after I thought of course I, 'cause I,
• 17:09
  'cause why would we not want to?
• 17:10
  And then if it's gonna help us get there, what a journey.
• 17:13
  What a right.
• 17:14
  And I would rather be in the front row of that journey than in the back.
• 17:17
  And I think that all of us are, all of us are boosted in that process by
• 17:22
  having these kind of conversations.
• 17:23
  Talking about what we're building and the things we were talking about
• 17:26
  before we got in what did you build?
• 17:27
  What did you build?
• 17:28
  How'd you do it?
• 17:29
  Which models?
• 17:30
  I love that.
• 17:30
  And that's just been something I've appreciated about our conversation and
• 17:33
  relationship too, is you've exposed me to some amazing technologies
• 17:38
  and tools given me a window into, Hey, try building this way.
• 17:42
  It's incredible to have that within a firm within the staff
• 17:46
  and within the team there.
• 17:47
  I love what you're doing, man.
• 17:48
  I think it's awesome.
• 17:49
  Yeah, likewise.
• 17:51
  You've shown me how to use a render, which is an, a platform
• 17:55
  that we can deploy apps onto.
• 17:57
  You've shown me how to integrate all of these different technologies
• 18:01
  into the projects that I'm working on which has been super valuable.
• 18:05
  It's I love having these conversations.
• 18:07
  Likewise, man.
• 18:08
  I appreciate you coming on here.
• 18:10
  I hope we shared something that the listeners will be able
• 18:12
  to wrap their heads around.
• 18:13
  And play with.
• 18:14
  But thank you so much, man.
• 18:15
  I really appreciate you coming on, Andrew.
• 18:16
  This is awesome.
• 18:17
  Yeah, chip and thanks for having me.
• 18:19
  It is been a blast.
• 18:20
  I love talking about this stuff and I'm always open to having
• 18:22
  these conversations for sure.
• 18:24
  A big thank you to Andrew Croley for joining us today.
• 18:26
  You can connect with him on LinkedIn to learn more about his work in legal
• 18:29
  technology and AI implementation.
• 18:32
  Legal Marketing Radio is a production of LaFleur Marketing.
• 18:35
  You can find us at LaFleur.Marketing and follow us on LinkedIn,
• 18:38
  Facebook, and Instagram.