- Barry Saunders AI Short
SPEAKERS
Barry Saunders, Will Bachman
Will Bachman 00:03
Hello and welcome to Unleashed. I’m your host will Bachman and this is part of our mini series of case studies on artificial intelligence consultants who are using an artificial intelligence in their practice or doing AI related projects. I’m very excited to have on the show today, Barry Saunders and McKinsey LOM based in Sydney, Australia. Barry, welcome to the show.
Barry Saunders 00:25
Hi,
Will Bachman 00:28
Barry, why don’t you start by giving us just a snippet of your background at McKinsey, you were brought in on the expert track, tell us a bit about what you did at the firm.
Barry Saunders 00:37
So at McKinsey, I was a I originally joined as a digital expert, focused on design, customer experience and product. And I worked there for four and a half years. And a lot of what I did was in the LEAP practice, which is McKinsey is kind of new business building practice, built everything from video streaming platforms to preventative maintenance platforms for refineries, Goldmine, optimization tools, and I got to work on a project where they were trying to move the capital city of an Asian country to a different city. So quite a variety of experience. I left McKinsey went became Chief Product Officer at an Australian fashion company, did a turnaround there. And then recently joined another McKinsey alum to run MSA, which is a strategic kind of digital technology company in Australia that does a lot of work for regulated industry and government.
Will Bachman 01:45
Amazing, and as So, you mentioned when we were chatting before I hit record, that you could provide a bit of a typology classification scheme for AI related projects. I’d love to hear that. Yeah,
Barry Saunders 01:58
I mean, I just want to preface this, this is just how I think about it, it’s, it’s probably not the most accurate, but I like to kind of think of a bit of a two by two. And one of those is a you trying to understand or you try you kind of trying to predict so a looking backwards to understand the pattern or what’s happened? Or are you looking to the future to try and create or try and predict what people are doing? And then the other lens is, are you looking for a? is the material working in? Kind of more text and creative? Or is it mathematical? And if you think about those kinds of quadrants, you’ve got the kind of mathematical backwards looking sort of stuff, which is understanding patterns of behavior. So what’s actually happened in the system you’re looking at? The forward looking mathematical stuff is next best action that’s kind of like predictive behavioral stuff. It’s modeling, what is the system going to do under new constraints? And then the stuff that’s more text oriented, backwards looking, I was thinking that is like classic sentiment analysis. How do you how do you dig into a body of text and understand what’s being said, what’s the meaning. And then obviously, the forward looking stuff is much more regenerative AI, and you’re kind of content creation stuff. It’s not a perfect quadro, but a lot of the projects I see sit kind of within those four boxes, or the really interesting stuff is when you when you stick two of those things together. But if you kind of take the next best action marketing tool, and you build it connected to some generative AI, then you kind of got an interesting engine to do personalized content for different customers. The backwards looking texts, that’s really interesting, because then you start to get really useful tools for digging into our 300 page policy document. How do you understand the drift and meaning of an insurance policy over time, things like that kind of start to unlock some really interesting value.
Will Bachman 03:54
I liked that framework. Tell us a little bit about the type of work that you do now.
Barry Saunders 03:58
So I should say when I started working in MSA with Roshni, we had actually been working on an AI tool to automate a lot of the, the really kind of time consuming work that management consultants do. So we have been working on a startup called first thing which was using Gen AI to take a whole lot of unstructured data and create the kind of classic McKinsey storyline. So do a big tax break down. I’ve actually, here’s a 300 page annual report, build out in clear statements, what’s the logical tree of assertions in this thing? And that I found really helped me in my job as an executive to actually work through here’s us, here’s the strategy. What are we trying to say or here’s like, the annual report, what’s actually the outcome of this? We were building this obviously last year, there wasn’t a lot of VC funding around so instead of trying to monetize it, we took it and went, well, let’s do consulting with this toolkit we’ve built and that has has worked really well for us, we’ve taken tools like that, and then worked with clients who are on the AI journey themselves, especially regulated industries. They’re pretty conservative by nature. And in many respects, I’ve missed, you know, a few of the waves have kind of digitization that a lot of tech forward companies have have gone through. But it’s kind of helpful in a way because they’re now ready to leapfrog. They didn’t spend all that time doing robotic process automation. But some of that stuff can now be done much more effectively, with AI. So one of the one of the projects that we did, which was actually really straightforward, we did it, I think over the course of a week, we did some work with an insurance company. So a company that insures large scale kind of agricultural businesses. So they had a number of policies that over time had grown naturally, some of these are 300 pages long. They, they tend to drift in meaning as as the language is updated, because you’re trying to deal with something that’s enormous, right. And it’s rare that somebody’s actually read that thing into end and actually understood and follow the chain of logic. And as the insurance market changes, you will have, you’ll find that your policies, you’ve got a policy over here for a small, small farm policy over here, for large farm, the legal definition of what a small or a large farm is might change, or the risk profile may change. So those two policies are now actually crossing over. They’re both covering some of the same stuff, and actually pulling apart the chain of logic in something but big is really challenging. And especially if you want to then consolidate those two policies into a single policy with sub clauses, that’s just an enormous piece of work. But what we built was a tool that would actually analyze the policies and actually extract the semantic meaning. And so you could actually take the two of them and say, well, within this section, they both speak to the same event, they use different language, but is the meaning the same as the kind of underlying semantic meaning the same. And this gave the client a really useful way of looking at two policies and actually kind of bringing them up. And in the same way that in something like Word or a code editor, you can see the the versioning, over time, the text change, what we had created was something where you could see the semantic drift. So where has the meaning change? And how do we kind of align these documents, that allows you to then get to a point where you say, Well, if you wanted to take this other policy and just create a single policy, is the meaning the same? If you go through and actually understand what is being said, at each step, is it the same meaning? And that that kind of allowed us to help them simplify the policies? We then but one of the things we’re looking at at the moment is taking the application of that work? And saying, Well, if you can do that with an insurance policy, what can you do you want to do a lending policy. So if you go to in Australia, a lot of lenders do home lending as as kind of their primary, the primary base for that lending. But homeland is very complicated. But a lot of the technical platforms that banks and non bank lenders use are ancient, they’ve taken months to kind of get through the process. And one of the things that’s happened over time is they’ve had the same thing, they have these enormous lending policies that are just very difficult to navigate to the point where it’s actually hard to give someone a home loan for some of these banks. And so what we were trying to what we were working on with some of our clients on that is, well, how do you just take your policies and simplify them? What can you remove, where the meaning of the policy is still the same? And allow you to kind of get to a point where you can offer a home loan much more simply.
Will Bachman 09:13
Talk to me a bit about what’s involved in doing that kind of analysis or, or building that kind of tool that you did it for a consultant that’s listening and says, Oh, yeah, my client could use something like that. If you want to someone wants to build a tool like that. What, what’s the level of effort? Is this a? One week is this a month is this a year is this $10 million to build with the tools that are coming out? Now? Tell me a little bit about that. It’s,
Barry Saunders 09:44
it’s definitely not limited to build. The thing that’s amazing about this space is if you want to build a tool that you use yourself, a lot of this stuff’s free. I mean, obviously the cost of doing the analysis using like opening are one of these other language models, there’s a bit of a compute cost. But you certainly don’t have to build a massive kind of enterprise tool. And there’s a lot of off the shelf for software as a service tools out there that are already starting to do a lot of this stuff we’ve seen, particularly in the legal space, quite a few tools for like generating contracts or analyzing contracts. I think if you were to do something like this, you’d want to be familiar with? How, firstly, just be really clear on what you’re trying to do. So are you trying to simplify a policy trying to compare two different policies, your what what are the kind of the input data that you’ve got, then you’d want to be quite familiar with the large language model, or the tool that you’re using to understand how to actually query it accurately. So there’s a fair bit of work to be done and actually understanding the prompting. And some of these, some of these tools are a little. There’s a lot of variation in how to get the best out of them. That’s it, there’s a bit of an art to actually kind of tuning them to get the response you want. So there’s, there’s ensuring you know, what you’re trying to get it to do? What is the kind of chain of logic of what you’re actually asking this thing to do? So one of them is to actually kind of say, well, if you’ve got two documents, and you want to compare them, how do you want the machine to actually break down the document in the first place to do the analysis? So are you asking it to take the document, read through it actually assemble a chain of logic, so read through every paragraph, cat capture the meaning in that paragraph, structure a tree where the flow of logic as it goes through the tree, then store that, do that with the other document that assemble the chain of logic, and then match one to one is the semantic meaning here and hear the same. And then ask what you’re building to kind of render a view where you’ve got the two documents, I have to be clear, I’m not the software engineer on this stuff. So I’d need to actually sit down with the teams who are actually writing the code to really get an accurate, but in terms of the kind of logic of how you’d work with somebody who’s used to these systems, just being clear about what you’re asking it to do and what the outputs are. And what kind of view you want to see at the end is kind of how this stuff would work. There’s a lot of stuff online, you can read about. Like the different kinds of workflows that people have built. There’s one called Reg, which I think is like, I forget what it stands for. But it’s like, it’s how you assemble the chain of logic to use AI to do tasks for you. So there’s a lot of kind of easy to read guides on how do you actually build this kind of task flow for this work. And I’d have a look at those kinds of things. If you want to build something is much more production oriented, there’s a lot of thinking around what people are calling LLM ops, which is large language model operations, which is kind of how you would integrate these tools in kind of a broader software architecture or broader workflow. I think Google, and AWS and Microsoft all have kind of guides of how you would implement that in your software stack.
Will Bachman 13:30
What other applications have you worked on? That are, you know, sort of analogous? I think, I think I saw that you were doing some work around RF cues, or potentially analyzing other sorts of large documents in addition to the contracts.
Barry Saunders 13:46
Hmm, yeah. So we, we built a bit of a proof of concept for this, actually, just using the initial one that I built was actually using a tool called mem.ai, which is it’s, it’s like notion or Rome, or one of those, you know, kind of note taking things that we did this is a proof of concept, we now want to kind of take and build our own system. But what we had done is built kind of a template, where you could take the questions that are asked in an RFQ. Government RFQ is great for this, because a lot of them are very templated, they’re going to ask a lot of the same questions because a lot of it is consistent across government. And just have it structured so that you’ve built out your source documents, which is, you know, your about us, sort of stuff, your capabilities, your case studies, there’s a lot of work to clean that data to make sure the case studies are well written and kind of structured in a way where stuff is marked up. So what was your approach that you used here during like a lot of tagging and stuff? But then what the system is quite good at is saying well, the Here is the section in the RFQ. They’re asking for a boilerplate thing about how we handle operational health and safety. Okay, that’s a very easy query, prompt the the AI disable using the text in your database about MSA. Please generate for me the paragraph about MX A’s operational health and safety approach. And it’ll give you kind of a nice little paragraph of this is how we do this. The case studies stuff is really helpful, because often those questions will be quite, quite open ended. And broad is like what demonstrate where you handled a digital transformation for government agency that had focus on on these key areas and an architecture that looked like this. And it’ll give you a couple of paragraphs that will be like, Okay, well, we did this work for this commission, in this way with these outcomes. Now, the text is obviously a little generic, and it always feels a little bit like written by a robot. But as a first pass where you go, Okay, this RFQ is 50 questions. And I’ve got a million case that needs to read through, give me the skeleton of this thing, so I can actually read through it, and then start to build it out and go, Okay, here’s, well, these are the five case studies, I need to reread and refresh myself and now build it out. But that’s for an RFQ, that might take a week or two, you can kind of get that first draft on and off the day. And then you’ve kind of set the tone for what you want to do next. That’s amazing.
Will Bachman 16:40
This was such cool, so cool to hear about this example that you’ve done and how you’ve been using AI. Very, for listeners that want to learn more about your firm, where would you point them online?
Barry Saunders 16:51
So our website is just an msa.com.au. Obviously, we’re in Australia, so that to you. And then obviously, we’re on LinkedIn as well. So Roshni Ratnam. And myself very Saunders.
Will Bachman 17:05
Yeah. So again, that’s m x a consulting is,
Barry Saunders 17:12
yeah, that’s the name of the company and our website is just MX a.com.au MX a.com.au.
Will Bachman 17:19
Okay, we will put that link in the shownotes will also include your LinkedIn URL in the show notes. People that want to follow up. Barry, thank you so much for joining today.
Barry Saunders 17:31
Thank you
