3.3 How do I develop an AI strategy? - Video Tutorials & Practice Problems
Video duration:
15m
Play a video:
Video transcript
<v ->All right. So how do you develop an AI strategy?</v> How can you use the process that we've introduced already to be able to define the strategy? Which I've already showed you these things. You're gonna use data, key performance indicators to describe where you are, and you want to also use them to assess your progress as you execute, just like we showed in introducing it into the previous process. But once you define your plan, the next thing you're gonna do is to start executing it. And when you execute it. then you start to refine your strategy. Because guess what happens? When you execute your strategy, you suddenly start throwing off more data. So one of the things that's really important is, when you describe where you are, to use data. Don't just make it subjective. Don't make it opinion. Use real data because as you're doing that, what's happening is that when you execute your strategy and you see where those KPIs showed up, now what happens is that you can then compare it and go back and redescribe where you are. You can execute this process over and over again. So you can redescribe where you are to see, "Hey, have you made the progress you're looking for? Did you actually reach that nirvana goal that you envisioned?" If you didn't, then you can start asking yourself, "Hey, have other things gotten in the way? Are there other obstacles that we need to really assess for what can stop us?" Now, sometimes you might not even need to do that. Sometimes you might realize, "You know what? This fairy tale land that we said we wanted to get to, actually, that's not really the right place to go." You might have to reinvision where you want to be. So whether it's because you reinvision where you want to be, or just because you've learned more, you might then find that you have new obstacles that you've identified that maybe you didn't identify the first time around. That might cause you to come up with a new plan to overcome those obstacles. And so you use this process of first defining your strategy. Then as you execute your strategy, refining your strategy. And by doing that, strategy doesn't become something that you do every three years, whether you need it or not, it's not something that you do and you put on the shelf, it's something you're doing over and over and over again. You're constantly pinpointing and improving and getting better and better and better. And so one of the things to focus on is that as you're doing this, it really all starts with money. So we already talked about how return on investment is involved, but what you really start with is a budgeting process. It's really about where do you get the money to do AI? So if someone says AI is a priority, well, what exactly does that mean? Unless there's a budget for it, or unless you can produce some type of business case that shows that people should put a budget aside for it, you're not really gonna get started to solve any problem at all. Now, here's something that is really important because we've emphasized so far all this type of work that you're gonna have to do, how you need to work with a data scientist to really deliver your AI. And you know what? That was actually an over-simplification. What I want to show you now is, is there's actually several different flavors of AI. And what you wanna understand is that not every problem that you're going to identify for AI needs the same flavor. It doesn't all need to be something you build from scratch. And so yes, there will be problems where you need to have a completely custom model. You need to develop some type of new technique in AI. Yes, that's gonna happen, but what I want to talk to you about is other ways for you to take advantage of AI and get value in your organization that are much faster and cheaper and might be a lot easier as the place to start. So the first type we want to talk to you about, the first flavor, is called "embedded." Now, what embedded does is it really is incorporating some type of AI into software you're already buying. So for example, if you're using Salesforce and it's using AI to do lead scoring, well, just keep using Salesforce, and you're gonna get the benefits of AI. And so that's an example of embedded. So we talk about commercial off-the-shelf software, or COTS. C-O-T-S. And so it's using generic data across all of the clients. So it's got one model, it knows what the model is, and you can use this wherever possible because it's the cheapest. Another example of embedded would be the email software that you use and that it identifies spam. So when you use your email client, your email probably automatically puts some email into a spam folder. Well, how's it doing that? Well, it's doing that using the same model across every person using that email system. So it's not using any specific data from anybody. It's just using a model that was trained on all sorts of data across many, many, many people from all sorts of different companies. And it just has this one generic model. That's what embedded is. One step up from embedded is data-driven. So it has a model, but it's going to train on your unique data. So that's a little bit different from embedded. So it might require a little more work on your part because you might have to make available data that you have available to you. And so it could be that the lead scoring mechanism might actually be, not an embedded model, but a data-driven model. Maybe it's doing lead scoring based on data where you're training the model based on what your outcome data is. So you're not just taking a model that exists already and using that, although it could work that way, but it could be a smarter lead scoring mechanism is actually helping you to train the model based on your own data. That would be a data-driven model. "Specialized" means that it's your data, just like it is for data-driven, but it's also your model. So here is where you would need a data scientist to work using tools to create the model so that it's uniquely fitted to your problem and your constraints. Now, understand that now you've suddenly taken a leap up in cost. Each of these costs a little bit more, right? So even with embedded, you have to buy the software. So it's not like it costs nothing. And you might have to train people to use the software, to use that AI feature that maybe they weren't using before. With data-driven, it's definitely gonna cost a little more because you have to somehow enable your data to be poured into the software. And so that's gonna require some type of IT integration. With specialized, now it's taken a big leap up because now you have to pay a data scientist. Now you definitely have to get someone involved. It's not just about pulling your data in, it's about the data scientist creating the model. It's not a model that just needs to be trained that's embedded in some software. Now it's a model that your AI person is going to invent. So now we're much more expensive. And custom is the most expensive of all because what custom might say is that there's actually some new type of technique that you might need to use in AI. And so there's something that you don't even have a tool for. And now you have to go get a new tool to be able to use that. And so now that's a very expensive proposition. And so you can see here that if you could get a lot of the benefits of AI, the lower you are in this model, the better off you are because it keeps your costs really low, it makes it very fast to be able to get the value. And so you only want to do these specialized and custom models when you really have no other choice, there's no other way to solve the problem, and it's a problem it's so important that you it's really worth it for you to spend this extra money on it. So let's look at some examples. So let's look at embedded. So here's a marketing example for embedded where it's actually fixing content errors. And so if you look at the web pages on your website, the value for this is it can fix lots of content errors at one time. And so it might be hard to fix all these problems manually in your content management system. If you have a large website with thousands of pages, it may have propagated the same error across thousands and thousands of pages. Like, so for example, suppose you put your new website redesign, all the pages, you put them in your staging server. And because you wanted to make sure that Google wasn't going to crawl them, you put in a robots tag that's told Google, "Don't index any of these things." And all those tags are in there on all those pages. Well, suppose you make a big mistake. And then what happened is you promoted to production and you didn't change all the tags. Well, now you have to quickly go in and change all of those tags on all those production pages. That might actually take you some time to do. It might take you a few days to do manually. It might take you a day to write an automated script to do it. And maybe there are actually some pages that really shouldn't be indexed. Like, for example, the cart page on your e-commerce platform. You don't want Google doing that. So you can't just go in there and blindly go in and say, "Hey, let's take all the ones that say no index and make them index." And so this could be a big problem. And so instead, maybe you have an AI technique, and what it does is it actually can detect when you made that mistake and when that page should be fixed. And so this actually allows you to go in and fix content errors. Well, how do you know this is embedded? Well, it's a piece of software that isn't special for your site. This piece of software could run on any website. And it doesn't need any of your data to train its model. Yes, it needs to look at your web pages, but it's not training the model using those webpages, it's just using those webpages to figure out how to execute the model. And so there's no special training involved to create a new model. That's an embedded approach. What about a data-driven approach? A data-driven approach, as you remember, is different because a data-driven approach actually creates the model with your data. So the model isn't embedded in the software, the model has to be created with your data. So what's an example of that? Well, suppose you were trying to suggest keywords to people in your site search engine that actually find the answers to the questions. So they're not just popular because you don't want, if people are popularly typing in a keyword that gets no results, well, they're not finding the answers so it's not enough just to be popular. You want to know if people are finding what they're looking for. Well, how would you do that? Well, maybe you would have a piece of software that's using the activity data on your website to detect when people are finding what they're looking for. So maybe they can tell, "Hey, they spent a long time on that page. They clicked on it from the search results. They spent a long time there. That's an indication that they found what they're looking for." And so they're gonna use your data from your website to be able to identify when people find what they're looking for. Well, how are they gonna do that? You're gonna have to do some kind of an integration. You're either going to figure out how you're going to pull your web analytics data into this model, or you're gonna maybe embed a JavaScript from this program so that it's collecting the data directly. But there's gonna be some kind of work that you have to do to integrate this so that they're collecting the data to train the model, to know when your searchers are finding what they're looking for. And they can use that to suggest successful keywords. They might even be able to use that data and that model to rerank successful pages to the top of the list. And both of those things are an example of data-driven, where the model exists already, but you need your data to train it. And so you don't need a data scientist because they already have all the features, they know how to extract all this stuff, but you do need to retrain that model using your data. So there's some type of integration involved. What's an example of a specialized model? Well, maybe optimizing your buyer journey. So you would still use webpages from your website, but you would maybe use your historic navigation and search activity. But how would you personalize this? Well, you might need to use some type of visitor identification technology. So visitor identification technology actually looks at things like what company do people come from? What geography are they from? Maybe you might even be able to understand what job role they're in based on content they've looked at in other places. And so this is all ways of you knowing more about the visitor. Well, different companies are gonna have different interesting information, different characteristic, different features that they need in their model. And so now all of a sudden you have to have a data scientist 'cause you can't use the model for some other company and just put your data in it because you might need different features. If you're a B2B company, you might care what industry a person's in. But if you're a B2C company, you don't. And so you're gonna use different features. You might care about what income they have. And so because you need different features, you don't just need to train the model using your own data, you need to create a completely different model that uses the features that you think are the ones that are really going to optimize your buyer journey. So how are they going to recommend the right pages as a buyer goes through your site? Well, you might need to create your own model. And so now we're suddenly in a place where you need a data scientist, where you've got a specialized type of AI that's being implemented. So ask yourself if you really need custom. So the truth is that most of you will probably never need custom AI. Maybe if you were a company that was trying to create AI software, maybe you'd need to use custom AI then, but most of you probably never need it. You probably will never get past specialized. And so these are really the focus items for you. Can you solve your problem with an embedded? If you can, that's great. But what if you can't? What if you need to train a model using your own data? Well, that's the data-driven approach. And that's the approach to take if you can, but again, you might have to have that data scientist involved. Maybe the problem you picked really is that specialized approach. That's gonna be more expensive, but it might be the only way to solve the problem. So as you're thinking about your problem, think about which of these forms of AI are really the one to solve the problem that you have.