This post includes the full transcript from the Intelligent Automation Exchange held on September 23, 2020 and and the YouTube video of the entire educational webinar: Intelligent Automation Exchange 9/23/2020.
An interview presentation on intelligent automation with Marshall Sied, Co-founder of Ashling Partners and Tom Wilde, CEO of IndicoAbout the Event: This one on one interview that will cover all topics regarding intelligent automation and how it is utilized within each company. Topics Covered will include: Intelligent Automation, Machine Learning, Unstructured Content, Transfer Learning, RPA, Citizen Development.
Transcript: Introduction
Marshall Sied, Co-Founder of Ashling Partners and Tom Wilde, CEO of Indico
Ted
You’ll have another chance to ask some questions in this session, which actually resonates with some of the themes we just heard about. We heard about some companies, like Southwest, looking at document processing, and we have the pleasure of having Tom Wilde, who’s the CEO of Indico, and has 25 years of experience. Actually on your list Tom: I once worked for Lycos also, so a lot of the audience we know will have to connect at some point to exchange stories like us.
And then Marshall Sied (I’m sorry if I’m pronouncing your name incorrectly). He’s one of the founders of Ashling Partners and is also leading a community on AI for Ashling. We’re gonna get an interactive document processing presentation, followed by Q&A, so make sure you use that Q&A button at the bottom. Then, we’ll go into some roundtable discussions after that, so we’ll give you some structure at that point. But Tom, take it away!
Tom Wilde
Okay, excellent. Thanks so much. Ted and Marshall, welcome here. For the fireside chat, we’d very much like to talk about, from a solution standpoint, this notion of document intake and understanding, but also from a deployment standpoint. How do you work with a vendor and automation services providers? There’s obviously Indico, on the vendor side, Ashling on the service provider side. Maybe to get started, Marshall, if you could just introduce Ashling in a little bit more detail, what kind of work you guys do and, and the markets you focus on? That’ll be great, and then I’ll walk through just a couple orientation slides, and we’ll get into the conversation.
Marshall Sied
Sure, yeah! Happy to, Tom, and good to be here today with everybody. Ted, thanks for the intro. I am one of two co-founders of Ashling Partners. Ashling Partners is really an intelligent automation service provider. We really cut our teeth in this space, more in a monolithic enterprise application, process re-engineering, and implementation side of the house. There’s many years of ERP doing fit gap analysis across current state future vision. And, we classified some of that as automation back then, but, we really evolved more into the advisory side of the hyper automation space and a natural momentum pulled us into the build and run side. So if you think about the full lifecycle (traditional life cycle, plan, build run), we really covered that. And because of that, there’s a myriad of technologies, including Indico, that we leverage pretty heavily in our workflows.
Tom Wilde
Excellent. Well, why don’t I do that? Why don’t I set the table here a little bit around this question of Intelligent Automation, specifically around unstructured content, which, for the most part, we talked about documents and emails. Let me just share my screen here for a second, and we’ll go through just a couple orienting slides, and then dig into the conversation.
Well, great, so just two minutes on Indico. Indico is an intelligent process automation solution for unstructured content. By that, we typically refer to documents, emails, and increasingly, images. So for up-and-comers, especially around insurance, things like claims on the mortgage side and things like underwriting where image based content also play in here. What we see in the market is that there’s just a tremendous amount of unstructured. Unstructured just hasn’t had the same moment of innovation in the last decade or so that structured data has had. If you have a structured data problem, you have hundreds of solutions to choose from. Probably the brightest stars around that have been the explosive growth in RPA and its skill in handling structured data type automation challenges. But inside this unstructured, there’s this anxiety that there’s risk that customers aren’t able to quantify. There’s opportunities they’re not able to take advantage of, and if they can make documents as accessible as their structured data, it opens up just an ocean of opportunity.
By and large, I think we’ll chat with Marshall about this too, in terms of what Ashling sees with customers. Generally, you have some form of a manual process when you’re dealing with documents. It’s a “see it key it” swivel chair type behavior, and that means it can be quite inconsistent. One thing that shines through when we talk to customers is their desire to codify these workflows or make them more software defined. As opposed to this individual understanding of a workflow, in insurance, especially, where you have people with decades of experience doing things like compliance and other functions and claims, there’s a strong desire to make sure there’s continuity as new people come into those departments. How do you do that knowledge transfer? It’s obviously much easier if it’s codified into some kind of process. The other problem is, there’s no way to scale these sort of people driven “see it key it” processes. You simply have to add more people. There’s never this moment of economies of scale, there’s simply no ‘add more people to accomplish more work.’ And customers are really looking to find how they can get more throughput, and more efficiency out of their current resources. So more often than not, what we hear is: how do we grow our business but not have to continue to invest so aggressively on the input side? So I think this is a very common framing. At the end of the day, we want documents to become data. If documents are data, then they can take advantage of all of those in those investments. We’ve made an RPA and CRM and predictive analytics, and all those good things. So it’s all about providing this high quality input to drive those processes we’ve already begun to invest in or, in some cases, have invested in for a long time. And so what are the challenges when it comes to documents. If you really scratched at this document problem, that document live on sort of a spectrum, right? On the far right, you have very structured documents. These are things like company forms, W-2, ahead of time, the structure of the information you’re going to get. And those are often accomplished with templates or some kind of rule engine. In the middle you have, invoices and remittances and purchase orders. They feel like they have some structure. But in reality, because there’s so many variants of those, they really look to the computer, much more unstructured. They really challenged the template and rule based approach, and it becomes very brittle. And then on the far left, you have totally unstructured contracts, compliance documents, filings, things like that. Titles and deeds, where the information location varies wildly within the document, the vocabulary varies wildly within the document terms of what you’re trying to extract. And you simply can’t process that kind of information without some kind of artificial intelligence based approach. And so the whole spectrum here requires a suite of techniques and approaches to really deliver a comprehensive document intake and understanding. That’s kind of the baseline Marshall, that I wanted to set here. Let me stop sharing so we can see the video more completely here.
Great. So when Ashling’s out talking to customers, let’s start with just automation broadly. We’ll set aside the specific problems within automation. But what are the top two or three things you see customers having to overcome when they decide that they want to make automation a priority in the business?
Marshall Sied
There’s several things, Tom, and so even sitting in on that last session, I heard kind of a couple recurring themes. It kind of goes back to distilling the people process technology challenges of any emerging tech program at any large organization. So on the people and process side, change is a critical challenge right now. Change to process and change to how we intersect with our work. We’re going to be doing less data manipulation, data collection, data entry, and more data analysis and critical thinking in this future of work that everybody’s kind of defined here. And I think the change piece on the people and process side is probably the biggest challenge. I heard, I think it was Tamra in that last session, talk about how people think they know their process until you sit down and you start to flow it out from a process discovery process understanding of perspective. We need to do that still, right? We still need to do that to capture number one business outcomes of current state and future vision. But also because we don’t want to just lift and shift automations. So redesign is a part of that. I kind of view things in bottlenecks because we know the tech is there. We know it works now from an orchestration perspective. One bottleneck is the process discovery process design side. And on the tech side, it’s orchestration across multiple capabilities, multiple platforms, truly embracing the best of breed that’s really kind of enabled with modern tech in the API economy and all that goes with that. And then it’s also how we run these workstream cycles. In automation, previously, or what I would call Intelligent Automation. When you went live with an enterprise application, anything that went bad after a ‘go live” was considered very bad, right? This type of technology is more of a CICD approach. We expect things to go wrong. Automations are a digital workforce. Your knowledge workers don’t know everything on day one, they’re learning on the job. And it’s just important to change that culture and that mindset in regards to it being okay for little bumps to happen after production moves.
Tom Wilde
Yeah, it’s always interesting. More generally on the topic of being able to define what success looks like at the start specifically with unstructured and with artificial intelligence, it’s vital that you’re able to precisely define success, and then work backwards. What we find often in this sort of awkward moment, is when in a lot of these deployments, the customer feels like they have sort of a ground truth on what the gold standard outcome needs to be. And then when that is presented back, and they start to try to turn that into an AI driven approach, you find that, well, wait a minute, here, there’s five people on this particular task, and they all five have a slightly different interpretation of what success looks like.That comes to the surface very quickly when deploying an AI solution. There’s always this moment where they have to huddle up and say, “Okay, we got to tighten up what success looks like, before we can proceed further.” It happens almost, I would say, 100% of the time, in my experience when you get started.
Marshall Sied
Yeah, absolutely agree with that. Absolutely.
Tom Wilde
When you think about the customers who rely on Ashling, an automation services provider in general, what’s a good reason to engage with a service provider on automation? And what are some bad reasons? Are there bad reasons, or are there poorly thought through reasons to lean on a service provider?
Marshall Sied
Yes. I’m going to give you an unbiased opinion, Tom. When I look at any automation kind of program, bigger than an individual workstream, and individual automation script, if you will, it’s hard to decouple, plan, build and run, because I’ve seen it happen a lot. Part of this is because of the RPA Renaissance. It’s really put a spotlight, or put a real emphasis on, Intelligent Automation. A lot of folks will get excited about the business case at first, so that’s kind of the plan piece. They might need a little help prioritizing, but they don’t actually understand the complexity of the build, and the fact that we might need an RPA Platform Plus. Something that can capture unstructured content, like Indico, or RPA Plus can recognize an invoice, a prebuilt invoice recognition model as an example. So really, their business case is not. It’s no invoice, it’s not accurate at that point. Make sure you’re working with somebody that can do, plan, build and run. So increasingly, that’s where we’re seeing a lot of demand in the marketplace because we are building our business case around an end to end automation. We’re basically component sizing those automations into smaller task automations. We need to get all those done in order to fully recognize the business case. In that business case, we’re not always thinking about the run. The fact that we still need to be supporting monitoring. This is still enterprise technology, interfacing with enterprise applications. On top of that, you don’t want to pull your build teams away from what they’re supposed to be doing, which is recognizing more value by doing more builds. We’re increasingly seeing a demand on the run side. Just in general, regardless of bashing or not, it’s important to engage with somebody that understands that full lifecycle, because if you miss one, it becomes the bottleneck of your program. And that’s kind of the path to stolled programs, in my opinion.
Tom Wilde
Yep. In terms of this sort of alphabet soup that customers have to deal with, how do you distinguish IPA from RPA from hyper automation? How do you think about that, or at least maybe echo? Or think about that, and how they’re going about, evaluating finding solutions?
Marshall Sied
I mean, great question. I think the reality is, your true north, and it sounds almost cliche at this point, should be to start with your business outcomes and reverse engineer from there. The reality is, as long as you have this myriad of technology in your tool belt, you will be able to figure it out. I think the biggest lessons learned is don’t be single threaded, don’t rely just on RPA, don’t rely just on traditional OCR platforms, don’t rely on traditional BPM suites, right? You’re going to need to have several of these capabilities in order to achieve higher ROI and business outcomes. But to us, that is the truenorth. From there, you can start your spectrum of content. It’s a perfect example, right? There’s different ways, different techniques, different methodologies to start to call out components of a process flow and understand where we might need some orchestration of some of these Intelligent Automation platforms. I think that’s a very helpful area to start. Just distill it down to why we’re doing emerging tech and from the get go, it’s to actually achieve results. It’s not to be a sound project, right? That would be my kind of biggest lesson learned, I guess, not just to have a shiny coin with a hammer going to look for a nail that fits, right?
Tom Wilde
When you think about unstructured use cases, why are they so difficult? What’s different from the traditional OCR approaches or even some of the features that RPA brings to the table? What is it about unstructured that you hear from customers that they’re frustrated with and are trying to solve when they take on those types of use cases?
Marshall Sied
Yeah, I think there’s several factors, Tom. I think number one is traditional OCR engines, it’s more template based, right? I have to build a template for every vendor. If I’m talking about accounts payable automation, I have to build templates for every bank. If I’m doing bank reconciliations, depending on the use case, it’s traditionally been template based. It’s not to say that a more machine learning model based approach hasn’t been there for a little bit. But organizations typically revert back to the templates, just because they know what’s going to work. It’s also been very cost prohibitive over the last 10 to 20 years. You typically tackled your top 20 or top 40 vendors, but you didn’t take everybody, so you really couldn’t get to straight through processing. I think that’s been a challenge for an OCR perspective. RPA has been a challenge too because it’s more in its native form. Not saying there’s not a convergence and an expansion of capabilities of these RPA platforms happening right now, but it’s business rule driven, right? Basically, if you can write it on a piece of paper and script it out, that’s what RPA is very strong at. When you start looking at going upstream, which most organizations do at a certain point because they realize there’s more business opportunity to go upstream and downstream as well. That’s why, they always say, RPA is the gateway to a lot of these discussions, to longer running workflows. Unstructured becomes challenging in the sense that it’s really the three V’s of big data, in my opinion. Two of those V’s are actually very much an enabler for something like an Indico. One of them can be a challenge for any machine learning project. When you look at velocity, that’s good for algorithms, right? That’s good for transfer learning, because that means we have more volume to train our models on. The quicker we can train our models, the higher competence rating, the quicker we can actually start to invoke this into a longer running workflow of automation. The challenge that I’ve seen for unstructured is on the variety side. So let’s say we have 1000 legal documents, so we think we have a representative sample to train our models on, but you started going through in the data dictionary, the data attributes are not representative of everything we need to be capturing across those 1000 documents. Before you know it, these training iterations are taking longer and longer. And so, that’s been a major challenge for us for unstructured content. But it’s also a major opportunity. I mean, you said it before with your slide: 80-90% of the content in an enterprise today is in an unstructured format. And it’s not just documents, right? It’s voice analytics, emails, so on and so forth. It’s a huge challenge, but it’s also been an increasingly huge opportunity, as we’re able to create these handshakes between these technologies.
Tom Wilde
Yeah, AI certainly has many compelling traits to try to get after these use cases. When you start talking to customers about using AI, what’s your counsel on? What are some of the new realities the enterprise has to face when they think about deploying AI in a production environment? What’s different versus an RPA bot?
Marshall Sied
Yeah, I mean just to be transparent Tom, I think some organizations are still figuring that out based on their industries. But I think an in-flight project is one way to respond. And then, post-project, there’s another way to respond to that question. In flight, what I will tell you is that it’s important that the business owns the results of some of this technology. We still like to have humans in the loop as an example on some of these because at the end of the day, we want an approver. Even if you could automate that fully, you should ask the question, right? It’s got to be a problem that’s worth solving, number one, but it’s also got to be a problem we’re solving with something like AI. I think when you get into projects, most business executives come to us at a certain point and say, “Man, we didn’t really know our data and the poor quality of it. And that data labeling isn’t as sexy as I thought it was going to be.” I mean, you guys have a great user interface, but still, you still have to do that, everybody still has to do data labeling that that’s it. That’s the blood and sweat equity we need to put into making these programs and projects successful, right?
Tom Wilde
Absolutely. I think that one of the key things that we like to do when we talk to customers about AI is reset this expectation, or this characterization that’s been in the press, which is sort of fun to think about, but not very realistic. It’s not like a bunch of robots are going to come in and sit down and do the work, that’s the wrong way to think about it. It’s much more that you’re going to basically fit your knowledge workers for a bionic arm, right? So all of a sudden, they can lift 10,000 pounds instead of 100 pounds, as opposed to robots coming in and just sitting down and replacing those people, which, while there’s great progress in the industry, we’re a long way from that literally happening. From an analogy, we always try to kind of reset people’s thinking around that metaphor.
Marshall Sied
It’s the Tony Stark Iron Man analogy that a lot of people use nowadays, right?
Tom
Yeah. Tony, is the human in the loop, right? I mean, quite literally.
Yes. So on that topic, as a services provider, how have you started to talk to customers about the emerging trends with citizen developers and citizen data scientists? And where does the service provider intersect with, a lot of the growth in on the vendor side, like Indico, of trying to make these tools more friendly? You know, for a subject matter expert, how’s that conversation evolving?
Marshall Sied
Yeah, I mean, I think citizen development is not a new topic, it just has a kind of a new veneer to it. I think a lot of people that have been in the industry for a little bit sometimes cringe when they hear the term and I think rightfully so. In previous areas, citizen development truly wasn’t to the empowerment level that it is today. To us at Ashling, at least, citizen development is about empowerment of the knowledge workers. It doesn’t mean that everybody in tax or in accounts payable or in the supply chain is going to become a professional developer. I think that’s kind of dispelling myth number one. That’s not the reality. What it does mean is that they’re likely to be consumers, or, kind of providers of input, to retrain an algorithm as an example. I think citizen development is really about empowerment. They could consume something a professional developer built in script, and the low code movement is suggesting that you can drag and drop these into workflows. It’s all about governance at that point. When you ask about service providers, there’s going to be complex builds that an organization and citizen development does not want, nor should they do. So it’s a build one-use mentality to me. So that’s where a service provider comes into play in the citizen development side. And it’s based on complexity levels. There’s complexity definitions and complexity calculators out there that a lot of people can use now. I would certainly recommend you to make sure you have refined complexity guidelines. Just because a citizen developer has the tools to build something, should we let them? And then automating that code review process, too? So that’s really where a provider can come into place, making sure you have the foundation to start to plug in places developers are into, but there’s an esoteric conversation around citizen development and low code in general, right?
Tom Wilde
Yeah, absolutely.
I’ll put you on the spot here. For 2021. If you had to list three, either predictions or trends that Ashling is paying attention to that’s driving your strategic planning, what would you pose here in that we’re getting close to the end of the year? Hard to imagine?
Marshall Sied
Well, I mean, you kind of hit you kind of hit on one of them, you kind of stole the thunder a little bit. I do think the term citizen development and low code is only going to grow. Once again, I know some people have a visceral reaction to that, sometimes negative, sometimes positive. So I do think that’s a 2021 trend. I think a lot of the platform providers are trying to solve this problem that we’re trying to help. Enablement and empowerment of the knowledge worker is something that’s important and critical for the future of work. I think another one is where we started the conversation on the change side and kind of the challenge side of automation programs. Process discovery process understanding, I mean, I’ve built more Visio process flows than I care to share with anybody, and I’m sure everybody has had that experience. We’re really excited about that, but how do we get future state design quicker, right? Because we’re spending all this time setting up governance and standards around development and intake and prioritization and getting more into a true agile DevOps mentality to continuous integration, continuous deployment. But unless you understand your processes, all that’s null and void (so process mining, task mining, even using process modelers) to some degree, that is an area where I think is only going to increase. We’ve already seen process mining requests from our clients come up in a big way. I think they’re viewing that as bigger than just “Hey, what are the best opportunities to automate?” It’s also “Hey, we can continuously improve and monitor our processes, because business is not static, it’s dynamic, it’s going to change, right?” The last piece, you asked for three right, so maybe one more. Transfer learning in general, we get called into a lot of. It’s a pretty natural trajectory, you’ll run a couple waves of automations, usually using RPA to start when you set up a program for an organization, over 50% of those are typically finance and accounting. Then you get a call and the CFO wants to talk about machine learning and AI. It’s very important just to do definition setting in those scenarios, and transfer learning to me is something that we highlight, because it’s really where the markets are heading. It’s how we leverage economies of scale, something we’ve already built, and transfer that over to a relevant use case. So it’s something we’ve already been componentizing. Python models, as an example, we use where we can. And that’s transfer learning in a nutshell. In my eyes, I know I’m oversimplifying it, and I think the sub, like the concentrics, sub circles of AI, I think transfer learning is one you’re going to hear a lot more about in 2021.
Tom Wilde
Cool, interesting. Well, great. So I thought we’d bring Ted back on and help us moderate the Q&A here. It looks like there’s a good stack of questions. So Ted we’ll let you sort of screen and direct those to either Marshall or myself. And we’ll take those off and the balance off the ten minutes here.
Ted
Sure, Tom, thanks. And great session. I mean, I just think this is such an important topic. Maybe, Marshall, we’ll start with you. Can you give our attendees practical advice on where they should get started? What’s the sort of first unstructured document type that they ought to kind of tackle and, you know, how to build confidence with this approach?
Marshall Sied
Sure, I think in general, Ted, I think it’s going to depend on the industry. The organizations in the big boulders are the ones that are going to be in your operations, whether that be like commercial lease agreements, or, legal documentation, if you’re a law firm. Those tend to be a lot more complex, so I always like to parallel pathways. Can we have one, that’s a semi structured use case? The example that everybody always uses is an invoice, right? Doesn’t need to be an invoice, could be a bill of lading, or whatever. But I would start there, while you start to also work on starting to move to big boulders in a parallel path. But a semi structured document is going to give you a feel for leveraging a machine learning model, even if it’s pre built, even if it’s not something we’re building for that because you can standardize. Those are more horizontal use cases, right? Then move to the big boulders that are the verticals or start that in parallel, because the collection of that documentation takes longer than I think a lot of people realize. To really have confidence in something like a model that we build for Indico, you need about 1000 documents. Some people will say 200. But, you know, you really need 1000 to get closer to that STP metric that is a straight through processing metric that a lot of people are looking for.
Ted
Okay, that’s good, that’s very helpful. And you mentioned the topic of pre built models. And, Tom, I would pick on you now, as the guy who spent 25 years working in this industry and evolving, you’re thinking about the technology that we can use to structure data. Where are we in terms of these pre built models? Is this a practical approach that people can take today? And if not, now, when?
Tom Wilde
Yeah, it’s a fascinating question that comes up all the time. I think that the first wave of using ML and unstructured did focus on this concept of pre built skills. Whether it was entity extraction, sentiment analysis, or summarization, these were all sort of pre built skills. When we got started, what became really clear is that intellectually, or conceptually, we all gravitate towards this notion of something that’s pre built, but in every case, the customer would say, “Shoot, we really want it tailored to precisely our documents, our data, our workflow.” And so we decided to step back and kind of flip over the problem and say, “Well, what if you could always build custom machine learning models and you could do that in a day or two?” That really is the game changer. Now that said, and that’s kind of how we approach the market, we don’t come to work with pre-built models, we make it dead simple for you to build a custom model. That said, I think the markets are evolving again, where there are going to be sort of intermediate starter models, not quite prebuilt, that sort of starter models. If you think about the transfer learning concept, we have a very generalized model. You’re going to see sort of these intermediate starter models, and then your custom models, so you kind of keep slicing the problem to further accelerate deployment. So that’s kind of where I see the market. A big trend, I think, for 2021 is the emergence of these starter models that aren’t quite all the way to the point where you call them pre built? That’s super helpful. Well, unfortunately, we’re gonna run out of time here, so let me just end with one last question. And maybe Marshall and Tommy can each weigh in on this? Because I think one of the earlier questions in the Q&A was around the evolution of RPA into Intelligent Automation. And certainly, we see a lot of the mainstream RPA vendors now adding, document understanding and document processing as a part of their product suites, as well as lots of independent countries like Indico with specialized technology that’s going even deeper. Maybe I’ll take Marshall first, and then Tom, you can close this out. Give us your perspective on the next couple of years, and what companies should be expecting to be able to achieve? What should the aspiration be with these technologies. Wanna start us off, Marshall?
Marshall Sied
Sure, that is a big question, Ted. So yes, I don’t think there’s any question there’s a convergence, kind of this extension of the platform. And I think everybody knows the Gartner term. Now, hyper automation, they invented that term, because this expansion of these RPA vendors has already started. Some of the bot process mining companies have their own OCR intelligence and OCR capability. Now, I think those trends will continue, but the old saying is that, you have to be a master of something at some point. You’re going to need to go deep into certain pillars of that hyper automation platform. With unstructured content, I think there’s a lot of opportunity for specialists, if you will, still, because that is a big problem to solve. There’s going to be a lot of opportunity for people to come in and solve it. But I just think a lot of people are gonna. I kind of see how ERP or the CRM era was a decade or two decades ago. You’re going to have a lot of modules, if you will, and you’re going to need to, to figure out if you’re going to stick with your core platform, or if you need to differentiate because of the problems we’re solving with the specialist.
Tom Wilde
Yeah, I agree, I’ll build on that. When you think about it, why didn’t the VPN vendors own the RPA space? They didn’t, and it’s because the RPA folks were focused on a different bolt technology stack and expertise in a set of problems to solve. I think Intelligent Process Automation is the same, it’s sort of the third wave, it’s a distinct set of use cases and distinct technology stack required, and a distinct focus on the kind of problems to solve. So I think you’re going to continue to see that. History shows that no one new vertical owns the whole vert, the whole span into the future. We’re seeing this kind of third wave as cognitive approaches enter the automation space.
Ted
Terrific. Well, I feel like we can have a lot longer conversation about all this, but this has been a terrific contribution to our collective understanding. Thank you, Marshall. Thank you, Tom. And, unfortunately, we’ll have to keep our pace up here as we have a lot more to cover over the next day and a half.
This blog post includes the full transcript from the Intelligent Automation Exchange held on September 23, 2020 and and the YouTube video of the entire educational webinar: Intelligent Automation Exchange 9/23/2020.
