Paving the way for AI in the Enterprise

Kevin Craine: All right, let’s move on to our next presenter, who is Greg Hauenstein. Greg, welcome aboard. Are you with us?

Greg Hauenstein: Oh yeah. How are you doing?

Kevin Craine: I’m good, Greg, where are you calling in from?

Greg Hauenstein: I am actually in Tampa Bay, Florida, so I apologize if you hear any thunder in the background. It seems like we might have a storm rolling in right now.

Kevin Craine: I see. Well, all right, well let’s welcome aboard Greg Hauenstein. Greg is an enterprise account executive with Hyperscience, and Greg is here talking about paving the way for AI in the enterprise. So Greg, please tell us more.

Greg Hauenstein: Thank you very much. The biggest thing that we’re trying to help people understand at Hyperscience is really the realities of what’s going on with artificial intelligence and really more specifically machine learning. Artificial intelligence, when you boil down the definition from all the times that it’s been the hot topic in science fiction all the way up until today, really all starts with just the simple concept of a computer making a decision. Any program that you’ve created, any business rule you’ve put out there, any of the more complex things you might have done really boil down to it. They’re all artificial intelligence. They’re all machines trying to replicate the human decision making or the cognition that we normally apply are only available to team members, but it’s all a matter of degrees and a matter of what those different layers and different tool sets can accomplish and how they’re going to be used inside of organizations.

Greg Hauenstein: When we look at the overall curve of innovation and the different revolutions that have happened over the intervening decades, we of course saw where all of a sudden the internet changed everything and then mobile changed everything and then cloud changed everything and now AI is changing everything and really is that next beachhead into how we can drive real innovation and real change inside of our organizations. One of the statistics I look at is this one of AI itself and the investments in it growing two and a half times faster than the entire software market in general, just further proving the interest, but also really the attention that it’s getting at many different areas.

Greg Hauenstein: One of the big things that people think about a lot is: why are we here now? What is it that has made AI so relevant and made it the thing that really is gathering so much attention? It kind of goes back to the history of computing itself. When you look at where all of the large mainframes, all of the large actual machines and these huge multi-story devices that literally were taking up entire floors of buildings, they all started with where it was worth it to apply the massive investment to create those devices to then actually get outcomes in large governments, large insurance, really big problems that were trying to be addressed.

Greg Hauenstein: As more and more of those computers or computing devices were created, the machines actually needed to then start storing data. That’s where you all of a sudden had decades ago the revolution of all the databases where everything went and how can we store everything. That then led into how can we devise more access and the internet and the actual interchange of information that ultimately then started creating more and more and more data all while at the same time the cost of computing was being driven down by all of these investments. What the current deployment and current interest in AI is enabled by is this decades-long evolution that has shifted the point of computing and the point of these systems being created from being systems that processed data and output something to systems actually where the data itself that had once been output is becoming the informer or the controller of these algorithms.

Greg Hauenstein: That’s where you’ll hear the term ground truth inside of data science. It’s where you’re actually using known truths or known things where it’s the actual end results that you would’ve gotten out of previous tries to make rules-based processes to input into systems to create predictive algorithms on unseen data or unseen documents. That is what is so exciting because that’s what allows us to actually start having the ability to say, how can we automate on what we haven’t seen before? How can we create systems that don’t have to be trained on every edge case, that don’t have to be looking backwards at all the situations I’ve already experienced in order to know how to handle this one? It’s actually about how to create understanding inside of the AIs so that they can actually make those assumptions all within the bounds of the accuracy or the risk ratio that you want to apply to what they’re able to do.

Greg Hauenstein: That really underpins a lot of what all of this is about, is really understanding how your strategy on deploying AI in the enterprise is going to evolve and where it makes sense to deploy tools that are going to be able to take part of the work and do it in a more automated fashion. The reality is, though, when we talk to executives, a lot of them are really under ideas of how hard it must be to do this. A lot of them are coming from the point of view of how to do it from scratch or doing it from a real techno-centric mindset. There’s a belief that in order to start getting the results or being able to start leveraging the tools that you have to staff up data scientists or you have to have a large AI practice inside of your organizations or that you need to actually understand the end-state of an entire process to be able to articulate how this big staff up and how this big process is going to translate into a meaningful business case.

Greg Hauenstein: And then back to that ground truth that to do any of this, you have to have significance amounts of labeled, normalized, stratified data in order to be able to train or build any kind of AI that you need to then deploy. And the reality is that’s true if you’re going to be creating an entire data science pipeline from scratch or trying to use one of the more generalized AI models. When you don’t have direct control over what’s going in or over how you’re deploying these systems, you really don’t have the governance function in-house. You don’t have the ability to know what am I doing to account for bias? What am I doing to interact with these systems? You’re relying on others which then requires you to build a lot more to be able to do that.

Greg Hauenstein: Platforms like Hyperscience and many others in the market, though, what we are focusing on is how do you abstract these layers of data science away and these layers of reporting and supervised machine learning loops so that you can focus on having a business user tool. We like to say having a point-and-click experience to interact with, fine-tune, and retain model accuracy. It’s something that is built upon this idea that you shouldn’t be trying to create an expertise in all of these extremely complex topics at every organization. You should be being able to rely on fundamental tool sets to be able then to build on top of so that you can have your business users directly interacting with these platforms and really directly being able to start seeing the benefits inside of organizations.

Greg Hauenstein: So when we think about this first you have to actually have the choice of what tool set or what pipeline or what interaction methodology you’re going to want to use and leveraging the tools that are provided along with the business information inside of your organization. Hyperscience has a fundamental belief that when you’re able to use your own documents, your own data, your own ground truth as the foundations for the models that are going to be deployed inside of your organization, you get so much farther than when you’re trying to interpret or interject an external model. You not only have better control over the ground truth or the realities of what’s the foundation of your model, you have better explainability of being able to know exactly what went into it in order to create the predictive natures of it.

Greg Hauenstein: It also means that you’re able to leverage the preexisting tool sets and having your business users able to fine-tune the interactions. Fine-tuning is another data science term that is really around some of those final revisions of an underlying model, being able to come in and get feedback from the model itself and then say, “Hey, this is where you were right and this is where you were wrong.” Take those answers, take those points of clarification just like when you’re training a new employee on the job that you’re aiming for them to tackle on how they should have behaved or how they should have interacted or what the response should have been. When you don’t have those loops and you don’t have those points where you can actually drive that learning back into the underlying model, you’re going to constantly be dealing with an error handling exercise rather than actually having your business teams able to influence and further improve the models itself that are really gonna be the underlying engine for automation inside of the organization.

Greg Hauenstein: And the point of all of this is to say that we have a belief that we really can democratize AI, that we can make it something that isn’t constrained to the realm of a technology project, that it’s something more that really the front lines of the organization can get involved in and then they can really, from a change management perspective also understand how it’s going to be deployed, where it can help and where it’s going to be able to be leveraged to make their lives easier when they’re doing processing and not just seen as something that’s coming into take away jobs. When they can really be a part of the training, can be a part of that fine-tuning and understand the interaction between human and machine, they really start understanding the fact that the only way for AIs and machines to learn is for humans to teach it what to do and not to just blindly intercept an idea it might have.

Greg Hauenstein: When we look at all of this and look at the things that we’ve been focusing on, our core belief has been on creating this enterprise AI platform, but the first place we’ve been deploying it inside of organizations is in intelligent document processing. And that’s backed up by Neil Ward-Dutton’s real belief that IDP or these really highly transactional document centric processes inside of organizations are ripe for these kinds of tool sets and ripe for real innovation using AI modeling to be able to actually look at how you can drive more automation without sacrificing accuracy. And that’s a core belief of many people in this space is that you should not be trying to use these new tool sets to just blindly strive for straight-through processing. The goal of all of this is to really create a better than human result while automating so that you can always have that partnership when it’s not confident and always have the human in the loop to be able to actually aid and assist whenever there is a question, but being able to automate everything that there is confidence on.

Greg Hauenstein: When you look at the rules of engagement around things like this, I love the thing that’s come up a few times of how do you do a proof of concept, how do you prove what’s going on? And the other thing I think about too is how do you not only think about how to prove that first use case? How do you think about how to actually create the pipeline inside of the organization of what are the different projects that each iteratively prove something so that they can be built off of to do that next project.

Greg Hauenstein: A customer like Mars, customer responsiveness is crucial and one of the places that they saw it is that really in the total financial processes in the organization, there was a huge focus on how do we actually do finance transformation globally? And instead of trying to say how do we tackle all of it and try and deploy one solution to do everything, the biggest thing that they said is how do we actually go market by market and use the wins in one market to then prove to the next market what needs to be done? And one of the key ways that they did this in the invoice to pay process was not only focusing on the true results of the platform, but also on the human in the loop interfaces.

Greg Hauenstein: Our team at Hyperscience and many other organizations are trying to do a significant amount of work focusing when we’re really working on AI on not only how to deliver a better outcome, but how to make the tools our teams are working in more pleasant and more efficient by focusing on the user experiences rather than just on the outcomes those team members are delivering. When you focus on how to actually make the tools that your teams work on a better tool for them to be in day in and day out, not only do you get the outcomes of the actual tool delivering, you also then are able to deliver significantly more actual decreases in processing time just by them being able to work better on the manual tasks.

Greg Hauenstein: It really becomes something where that human and machine partnership needs to be a critical, real foundational mindset. You have that it’s just as important to think about the tools you’re gonna be giving to your team members and how that they’re gonna be interacting with them, what the change management is going to be as it is going to be, “What is the actual best AI platform? What is the system that we’re able to control the data going in where we know what is actually building the models, where we are involved in that process?” So that you actually are able then to really know this is how this model was built, this is how it’s going to interact. I was a part of the testing, I know where there are edge cases, I know how that we can deploy it with grace so that we actually are able to make our teams better, we’re able to actually deliver a better service to our customers and ultimately gonna be able to drive the value to the organization that all of us ultimately are trying to do inside of these processes.

Greg Hauenstein: At the end of the day, really all of these concepts boil down to this comparison many organizations are making. On the left side we have the behemoths and the most exclusive AI club that’s ever been created. These are some of the largest and most well capitalized organizations in the world that either already own massive data centers or have very easy access to them and very easy access to capital to be able to go take these gigantic public or semi-public data sources and create models that they say are best. I heard the statistic earlier today about the decline in accuracy of ChatGPT over not a long period of time. It’s really amazing how so much money can be spent to create so many things, but when it’s done on a finite data set that none of us as the fastest adopters of any AI tool we’re directly involved with can actually then go and say, “Well this isn’t right anymore, but now we can’t do anything about it.” We can’t be involved, we can’t then say, “Hey, I wanna take that subset and use it over here to do something.” It truly is another incarnation of the black box of, “Okay, here’s a tool that has functionality but I can’t interact.”

Greg Hauenstein: While on the right side, it’s platforms like Hyperscience where we take the approach that we think you need to be directly involved. You need to be using your own data. You need to be actually having a simple point and click interface to be able to go through and deliver bespoke AI models that are created specifically for your use case with your documents at heart so that you can actually deliver a more meaningful and specific result for your organization rather than an application of a generalized result. All of this kind of built off the fact that it shouldn’t be that hard to do all of this if you’re working with platforms that are trying to abstract the hardest parts away and trying to make it something that you can run on a laptop so that you can actually deliver differentiated results to your organization.

Greg Hauenstein: AI and the enterprise, it’s a dance that’s been going for years and it’s one that’s going to evolve even more over time. We have a strong belief that there is always going to be the need for partnership between team members and machines for that human and machine collaboration and even more than just human in the loop. It is something that the point of all of this is to elevate the work of our teams and being able to do it in a way that before regulation we know what is going on and after regulation we can prove why things happened and what the point of all of this was. We have a strong belief that there’s only going to be more investment and more interest in these topics.

Kevin Craine: Very good. That is Greg Hauenstein from Hyperscience. Thank you so much, a really great session today because I am always trying to understand how we really use technology, in this case AI, in real ways that really move the needle in terms of digital transformation, in terms of organizational performance. Tell me a little bit more about the input and training of AI. You’re all about using our own data to train and deploy AI models. How can we best do that effectively and can you give us an example of how that is applied in a process that we may know that is common among most of our organizations?

Greg Hauenstein: Yeah, definitely. So really the biggest reason we really have this point of view is because we also know based on real world experience that every use case is going to be different. Every organization might have their own discrete data points or entities if you want to get into a real kind of data science term that they’re trying to get out of documents. And the purpose of this process is to create a model to get your specific result as quickly and easily as possible without having to do all the extra work. So if you look at an example of account opening statements or driver’s licenses or checks, every organization when they’re processing these might have their own set of fields or their own schema or their own metadata points that they’re trying to extract that we believe it’s easy to just say, well drop in a hundred documents and click on those where those data points are and train that model to be able to understand the actual meaning behind the relationships between those fields and the actual ability to find them inside of those documents. On the unseen.

Greg Hauenstein: The things Hyperscience is doing is trying to say how can we make that process of training as easy and as future proof as possible? So we’re actually doing development and releasing products as recently as our latest release that actually will curate the documents inside of your data sets to be able to say not only, “Hey, here are different heuristic groups of documents to use in training,” but “Here’s the specific high important ones you should use in your ground truth so that you don’t introduce bias, you don’t over sample, and you’re able to get as much predictive automation as quickly as possible.”

Kevin Craine: Now Greg, you guys at Hyperscience are all about democratizing AI. You mentioned that the perception may be that it’s somewhat complicated and you need a team of scientists to do this, but you say it’s not necessarily so. How do I do all of this without hiring expensive AI engineers?

Greg Hauenstein: Well, luckily, since all of us are in information management, I’m sure we’ve all created a virtual filing cabinet or a repository or something in a document management system before, and you know, like any application information management, it all starts with the actual metadata or the taxonomy that you’re trying to get out of a document. So with Hyperscience, it literally you begin with the end in mind, what are you trying to get out of the documents? Create some fields, you give them data types to say what is the way we might want to use natural language processing to normalize data? And then you train on as small of a sample set as possible to create predictive automation without sacrificing the accuracy. And that’s where the magic of Hyperscience comes in because we are able to actually help you see as you turn that dial of accuracy required, what automation rates you might see as well as where you might need to actually sample differently to create more specific and better ground truth to create more automation. It really is as simple as defining your fields and clicking on the samples to be able then to create something that any business user can be trained on how to do in a few minutes.

Kevin Craine: We are here talking about using AI in the enterprise. If you have a question for Greg, now would be the time to pop it into the Q&A feature and we’ll try to get it into the few remaining moments that we have with Greg. This question coming in, Greg: how can we use AI within our organization for our learning and development such as certification training?

Greg Hauenstein: Wow, that’s an interesting topic. I’d say that inside of learning and development there’s some topics that you might be able to say like maybe I’ve got large data sets that I want to use more like sentiment analysis or summarization topics to be able to actually kind of condense the amount of information we’re asking team members to interact with. Also on the other end of that I could say that you could also use different tools like named entity recognition to be able to say, “Hey, maybe I want to actually mine the responses I’m getting in more open form content, try and see if I can see patterns in how people are responding,” if I might see that there’s maybe suboptimal results in the certifications we’re asking people to do or in the internal quizzes that we’re having people actually try and fill out. Ultimately, I think somewhere like training and certification, it comes down to: are you looking more for something that’s on the generative side of the coin or on the processing? What’s happening that will have dramatically different recommendations on the options you have available.

Kevin Craine: Now, Greg, I often think, I often wonder, yeah, we hear, we see a lot of examples of using AI in process performance, intelligent document processing. How can we apply these tools in information governance itself?

Greg Hauenstein: So governance itself I think is something that comes down to the actual processes that you’re gonna be having where you’re interjecting the tool into. It’s something that the topic becomes one that now when we’re using AI and we’re gonna be using tools that are built on top of data that we have, that we then need to start saying, “Well what are we doing to one govern the underlying data that’s gonna be inside of these models, the things that they’re built on top of?” And then how do we make that a part of not only the AI ethics that I also recommend to everyone start actually thinking about creating an AI ethics committee or working group, but also then how do we manage that data? That’s a whole other repository now that we’ve never had to deal with before. This is now something that a lot of times will live on a lot longer than like operational documents. They’re things that are living and breathing intermediate repositories that we need to create specific policies and specific procedures around so that we can really adequately folded in to what we’re doing broadly in the organization.

Kevin Craine: All right, Greg, it’s been great speaking with you today. Before I let you go, one last question from our audience. What if the information doesn’t have a lot of metadata associated with it, how can the information be trained with a lack of rich metadata?

Greg Hauenstein: That’s a really great question and that really gets to the heart of different tools being used for different steps or different outcomes you’re trying to achieve inside of real processing of these kinds of data. The artifacts, I like to say, a lot of times the nouns are gonna change, but the verbs will stay the same in what we’re doing. And when you have situations where maybe it’s more long form or it’s a sentence out of an email, it’s things that it’s not a form-based document or something where there’s a real schema that can be applied to it. Really, that’s where more text classification comes into play. You have to get to data, but once you have data in some form, then it’s really about saying, “Well, given this whole string of text, was that positive or negative? Was that a specific identifier? What was that?” That then the modeling tools use to be able to create that predictive model to label that in the future. And a lot of times this is used for being able to do kind of real broad form classification exercises where you can think of like the triaging of incoming information. What if we could automatically label some of that and just get it to the right place to be able then to start using? And so a different kind of tool would be used for things like that.

Kevin Craine: Very good. That is Greg Hauenstein from Hyperscience. Find him at hyperscience.com. Greg, thank you so much for being a part of our session.