The Banking Buzz: Unlock the Power of AI Classification: Smarter Document Management with FASTdocs

Hello, and thank you all for joining us today. My name is Cara Talcott. I am the Marketing Communications Coordinator here at Alligent, and I'll be your moderator for today's webinar. In this session, we're gonna explore the power of AI classification and smarter document management in FastDocs. With us today, we have two wonderful speakers, Cameron Marks, the director of product management for FastDocs and Mark Rind, a Senior Manager of Software Engineering for FastDocs. If any questions come up at all during the webinar, feel free to type them into the chat box. We'll have a designated Q and A segment at the end. And with that, I'm gonna hand things over to Cameron to kick it off. Alright. Thank you, Kara. So I'm I'm super excited to have this webinar today, especially with my my co presenter, Mark Rind. I work with Mark on a day to day basis, and, I'm really excited for him to share, what we're bringing to market, in our FastDoc solution and and talk about some of the things beyond that. With respect to our agenda today, we'll talk about our document classification strategy. And, really, this is about exploring the why as we embark on this journey and identifying, some of the problems that we're trying to solve, by using a tool. And that tool being, the the deployment of AI. And, and this really comes as we address some of the goals that we have around, AI classification. And then from there, we'll dig into the implementation, and and Mark, will guide us through how we've implemented this technology. And really excited to to walk through that component. And then lastly, we'll explore some future steps, some of the doors that this opens and the possibilities, in front of us. So talking about our document strategy goals, really, it focuses on automation. And, you know, we want to take the work, out of making documents accessible and available. A lot of the controls that we traditionally put in place, in FastDocs around, around getting those documents there, revolve around doing so efficient efficiently with repeatability and enabling, access and retention guidelines. Accessibility is a key piece. We want to eliminate exception cues, documents that fail to enter the system for whatever reason, and ensure timely accessibility both for your financial institution's knowledge workers, as well as the end consumer. And, really, this centers on, process and and policy and applying your compliance guidelines, document retention life cycle, and doing that in a seamless and integrated way. So document classification using AI, it provides the ability to automatically, without user intervention, identify types of documents that are entering the system, classify them, categorize the document, and also extract data and search criteria. So this automated process, has the goal of reducing the costs and errors associated with the manual classification and indexing. And it also increases accuracy and the robustness of, of existing, document import templates. So from here, we're gonna switch over to a poll and explore, we'll we'll we'll post that. And, really, the question here is, has your financial institution, developed an AI strategy? So if folks can take a look at that, chime in and give an answer, we'll leave the poll up for a few moments here for folks to answer. And then we'll also be reaching out to understand, what your specific strategy is. So we'll take that as a follow-up item on on our side. So just give it another moment for folks to answer. Okay. And then we'll start by then we'll continue on talking about, talking about the the common workflow, for document classification as we've as we've brought it to market. So really, this centers on, our file import studio. And so documents are coming in through third party sources. They enter file import studio. And file import studio, has some some mechanics to it, and that we've added that allow calling out now to, an Alligent hosted large language model. And then this will allow us to more dynamically classify those documents coming in traditionally through through the system, classifying them, extracting indexes, and then ultimately store them, in FastDocs. So we also wanna define, document classification. So this this centers on, what are the what are the mechanisms behind classifying a document? So identifying a document type, being able to extract metadata from those documents, and then managing exceptions when things, don't match a template or fall outside of the normal bounds, being able to manage that efficiently and effectively. And then the legacy, classification methods in FastDocs really center on File Import Studio and what we have in the form of report and document templates. Now the types of documents, have provided, some some challenges for the traditional processing for file import studio. And document types, can come in a couple of different flavors. So so structured documents, they'll have a fixed format, and layout changes to that format occur very rarely. They have defined fields, maybe even specific character boxes. And some examples of those are tax forms or loan applications. When it comes to semi structured documents, they can have a bit of a more flexible format, but they have some reliable markers or anchors that provide consistency, to that document format. But some of the content can also be dynamic. So some examples of these include account statements, core reports, or emails. And then unstructured documents, they they'll lack a predefined format or organization, and using legacy processing methods. This would be very difficult to automatically process, without an accompanying index file. And that really leads us to our goals in embarking on an AI classification journey. So, traditionally, a trained model has been used to, to tackle the the challenge of, classifying, data information and documents. So when elaborating Allegiance goals and it comes when it comes to leveraging AI, we've approached it in a couple of different ways and a couple of different goals in mind. We want to add, incremental value to the solution you already use, and we'll see this, when Mark presents. But the approach we took with FastDocs, builds on existing infrastructure. An additional pillar to the to those goals is a low cost to implement. So by opting to use a large language model instead of a trained model, we're able to eliminate, that data training overhead, that's traditionally, included as part of a trained model approach, and yet still produce accurate document classification and index extraction results. So we're embracing the use of of document management best practices when it comes to accessibility for knowledge workers, for members and customers, as well as fulfilling regulatory and compliance requirements. And then beyond that, we know there's more that we want to do with AI. So this really lays the groundwork, for our future efforts in, you know, leveraging and gaining value from the use of of of AI. So at this point, I'm gonna transition over to Mark, to talk about, how we've implemented AI classification. Hello, everyone. Good afternoon. I'm really excited to be able to speak today about our new AI classification implementation. Our new AI document classification feature involved using file import studio as our initial integration point. This was chosen as our first integration point because it has many existing features that work well with AI document classification. The path we chose was to leverage AI large language models, commonly referred to as LLMs, to extract indexes and to perform document classification. Once the indexes are extracted and the document is classified, our new integration submits the documents to file import studio for storage and folder. The advantage of using file import studio and its report templates is that many customers already have FIS reports set up that contain the foldering, indexing, and security for many documents that they already import. The change we've introduced is that now AI will perform the index extraction and pick which FIS report to use for processing the document. Using the technique of prompt engineering, we're able to instruct the LLM to classify a document according to existing report templates set up in File Import Studio. Prompt engineering is simply the ability to instruct the LLM on how it should behave. In this case, we've instructed the AI that it will be used for classifying documents and extracting indexes. These same LLMs are used for all types of different operations out in the business world, chatbots, you know, all kinds of shopping assistance, medical assistance. There's there's a lot of things that you can use these LLMs for. In this case, we've given it very clear instructions through prompt engineering that it's going to work on just classifying documents and extracting indexes. Each report template in FIS is used as a possible classification target with the indexing category and indexing labels of that report being the data points that the AI uses to make its decision. So if you have many different FIS reports already set up, the new Allergan AI classification process will evaluate those categories and labels and attempt to find the best match to the document being evaluated. Using this approach can lead to less maintenance of the FIS report templates as those documents change and evolve over time. For example, a new marketing message on a statement may be added and it moves the member name to a new row in the document, changing its physical location in the document. This may involve making a change to the FIS report template to account for that new location. AI classification should automatically adjust to this change in the document structure without any need to manually modify the FIS report template. So we should see a a pretty big reduction in, as your documents change, having to make changes in the FIS report template. With our first release of AI document classification, we offer several options for which LLM is used and how it is hosted. We support some cloud hosted models, such as the OpenAI or Azure platform, and can easily switch the model with minimum configuration changes. We can also easily switch the platform as well. We also offer the ability to have private hosted LLM models, allowing for more controlled security and ownership of the data in the model. For example, if a new public chat GPT model is released, we can update the private hosted model and start using that immediately, or we can switch to an entirely different LLM to achieve better results or better performance. As Cameron mentioned, in this release, we are not training the LLM on your data. We are using existing LLMs that have been pretrained on existing datasets. But not all publicly available LLMs are created equal. Some will have better datasets for documents and index extraction as well as better performance and better results. We have worked with many of the currently available public LLMs and have evaluated the results and performance for accuracy. New LLMs are released constantly. And as this technology stack evolves over time, our ability to easily change models gives us flexibility as new versions are introduced to the marketplace. These new versions will likely improve accuracy and results as new advancements are made with LLMs. Now let's see a visual example of how AI document classification works with File Import Studio and FastDocs. Let's start with a simple statement as our document. Here you can see we have a sample document. It's a pretty standard statement. We've got the member name and their address on the top left corner. We've got their primary account number in the top right. We can see their ending balance on the middle right, and we also have, where is it, the statement date as well. So there there's several data points on here. One thing I wanna call out is you'll notice there are two addresses here. One is the address of the bank at the top, and another one is the address of the customer below. That's gonna be important for us to look at a little bit later. Next, we'll use a file import studio UNiDEX report template to set up both the indexing and folder location of the document in FastDocs. On the left, you can see this is a pretty standard UNIDEX setup where we draw boxes for each of the indexes being captured. And on the right, you can see this is a foldering, setup. So we're gonna be foldering this document into a specific category using a specific security group and foldering it to the statements folder in the accounts tree. This screen, also from file import studio, shows the index fields that are being captured as well as specifying the name of the document that will appear in FastDocs. We see fields like the account number, customer name, customer address. You'll notice it says customer address as well as the ending balance being captured. Up to this point, this is all existing file import studio setup and functionality that users are familiar with, so nothing's really different to this point. At this point, the file is dropped into an input folder that our AI classification module is monitoring. Here, we see a dashboard that logs the communication with the LLM. Highlighted in the red box, we can observe there is a classified document response, which shows the details of the indexes that the LLM extracted from the document. Once this classification is performed, the documents and its indexes are sent to file import studio for indexing and storing. As mentioned earlier, this leverages the existing File Import Studio infrastructure and applications to work with AI classified documents. Here, we see the document has been successfully processed by File Import Studio as displayed in our admin tool UI. This screen shows the documents stored in the appropriate account tree in the FastDocs web client. You can see the statements folder on the left was the same folder that we had set up in the file import studio report template. The web client offers quick access to see a preview of the document as well as review the indexes. So we see a a preview of page one of the document as well as the indexes that were captured. This slide shows us opening the document in our FastDocs document viewer, again in the web client, showing a full size view of the document and displaying the indexes extracted by the LLM. This illustrates the entire process from a document being dropped into the monitored classify AI classification folder all the way through indexing and storage in FastDocs. One thing I wanna point out that was a a really good result that we experienced during this implementation and building this tooling is that I called out earlier that there were two different addresses on the, document. You'll notice on the right, the customer address has one two three four Anywhere Street in the indexes. And on the left, we see that's the same address as Jane customer. How did the AI know that that was the in index that we wanted, the value? Because it could have chosen the first address because both of them are addresses. So one one aspect is the label names. As as I mentioned earlier, it uses the categories and labels to figure out the best match of how to classify and extract the indexes. Any anything that we can do to make those labels more explicit, and that's why I called out that we were that the index label was the customer address because the AI is able to determine from the index label, instead of saying address saying customer address, it tries to find the address that's contextually related to the customer. So the AI gives you a little bit of that intelligence to be able to try and pull that data, but we can help it along with the names of our index labels and our categories in FastDocs and File Import Studio. So there are things we can do to help the AI give us better results as well within our software that doesn't involve really any report setup. So hopefully, this illustrated the entire process of us going from having a document being dropped into a folder just like you would with a standard file import studio process today, but now leveraging AI in the middle to do the classification and extraction of indexes leading to less file import studio report template maintenance. Great. Thank you, Mark. You are it's a it's a really exciting exciting time for for Aligent and embarking on this journey in AI classification. We talked about some of the some of the the behind the scenes, configuration. So we wanted to put up a poll to you as well to to explore that a little bit and understand, what LLM deployment your financial institution would would consider, whether it would be a public cloud, a private LLM, or if you haven't, you know, got down that far into into exploring it. And, again, this this will serve, as follow-up to to the webinar for us to explore and and and build on this conversation of deploying and and making use of of AI. So we'll leave that poll up for a minute here just so folks can can get their responses in, and then we'll continue on. Alright. Thank you. So now, what we really wanna dig into is the the value proposition for, embarking on this journey of classification. And, really, what this comes down to is it'll streamline, the setup and configuration of getting third party information into your system. And with respect to integrating with third parties, traditionally, a requirement there is that that third party may need to produce an index file. By going to to this mechanism, that that would eliminate, would potentially eliminate the need to produce that index file and really rely on the data, in the document and the AI classification engine, to to both classify the document and extract the data. It'll also drive down ongoing maintenance of your report templates in in file import studio and streamline that process. It'll increase accuracy, with with documents being classified, and so you will have fewer exceptions, to track down, outside of the system. And then it it supports additional document types. With file import studio, the the limitation is really around, text based documents, or documents accompanied by an index file. So by by going to to AI classification, we're able to really broaden the the net that we cast, with file import studio and the documents that can be, brought into the system in an automated fashion. Other areas that that we will find savings in are around exception rates, so driving down, those exception rates, the remediation costs, finding finding documents that have fallen through and getting them folded into the proper place, and then really going to a more automated process versus manual processing of of documents in FastDocs. And so as far as future steps go, you know, we talked a lot about file import studio, but, really, the goal is to drive, the AI classification, really to the end user's perspective. So enabling file import studio, with AI, is a first step. And then enabling document templates within the actual user interface and leveraging AI, sending documents, not necessarily having to drill down to a specific folder to manually file a document, but submitting the document instead to a classification engine and then it being folded appropriately. So extending to the FastDocs UI. Also, for our Accu account customers, being able to, instead of tracking down a specific exception that you want to fulfill a document for, being able to submit it to AI and automatically check that exception, fulfill the exception, have the document boldered specifically for, the exception that it's intended to track. And then a couple of other things, really around exception tracking, in future steps. Being able to identify, if there's a missing signature or a missing document as part of a business process, providing robust universal search, doing contextual, semantic, or sentiment content searching within your within your document repository, triggering workflows, and analytics and trending analysis. They're really a bright horizon of things that we can explore, as we as we embark on this journey with AI and our content management solutions. So we're really excited to to bring this to market. Should you have, further interest in in seeing a demo, please reach out to your sales rep. They will they will gladly schedule a demo, and I'll I'll be, side by side with those folks, to help you embark on this journey. So thank you all for for participating today. Mark, thank you again for presenting with me. Really excited to bring this to market and and embark on this journey with our with our customers. So at this time, I'm gonna turn it back over to Cara, and we'll see if we have any any questions from our attendees. Hey, Cameron. Thank you so much. Yeah. It looks like we have a couple questions, here in the chat. Let's see. First one, are documents sent to the LLM transmitted anywhere else? Great question, Kara. Thank you. No. The answer is no. The document is sent to the LLM model for classification and index extraction and is not shared or saved to other systems or sent to any other third parties, the file is stored permanently in our FAST Doc system once it has been classified. Thank you. Let's see. Another one. What document types are supported by the AI classification engine? With this release, we support PDFs, Microsoft Word documents, as well as most common image formats such as JPEG, PNG, TIFF, GIF, as well as many less common image formats such as BMP or bitmaps, ICO, HEIF. There there's a few other ones, but those aren't used too often in document processing. We do expect to be able to expand the supported document types with future releases as well based on customer needs. Awesome. That's great to know. Let's see. If I have an existing, Azure LLM subscription, can I point Fastdocs to it? That's a I'll take this one. This this is a really good question. So we've created, a robust infrastructure with the the Fastdocs AI classification implementation. If you have a specific, existing subscription, it's likely that we can point FastDocs to it. I would encourage you to reach out to your to your sales team, and and we can explore that directly. But but, yeah, I think I think that's something we could certainly entertain. Great. And it looks like we just have one more. Can FastDocs AI classification be used, in both on prem and cloud deployments? Yeah. I'll I'll take this one too, Kara. Absolutely, it can be used in both on premise and cloud deployments. Really, it comes down to, which LLM, your financial institution is comfortable pointing to. So if it's a cloud and Azure implementation, for example, we could, potentially explore, your you having a subscription to that, the financial institution having that subscription, or Alligent, having that subscription and administering it for you. Or, if you would like to rely on, an Alligent cloud hosted private LLM, we'll have that available as well. And that would simply require, coordination between IT staff and the Alligent data center team, to enable that connection from an on premise to a data center, data center implementation of the LOM. Thank you so much. Looks like that's the last of our questions. Thank you everybody for joining us today. If you have any additional questions, feel free to reach out to marketingallegent dot com or fill out a form through the website. Please check out our events page for more additional webinars like this on our calendar. And thank you so much. Have a great day. Thanks, everyone. Thank you.