AI Leadership Insights: Where AI and Workflows Merge

Mike Vizard: Hello, and welcome to the latest edition of the Techstrong.ai video series. I’m your host, Mike Vizard. Today, we’re with Michael Beckley, who’s CTO for Appian. And we’re talking about where AI and workflows come together in an interesting way, where maybe two plus two is seven or more, and we’ll see how things go from there. Michael, welcome to the show.

Michael Beckley: Thank you, Mike.

Mike Vizard: One of the things I’m starting to see is we’re all amazed and enthused and maybe equally scared of all things AI. But at the end of the day, it creates something, whether it’s a summary of something or an analysis of something or what it is, and I feel like increasingly people are saying, “Now, what? I have this thing, it does a thing, but how do I automate something off of that?” Because the AI platform itself doesn’t have an engine in it per se, so what will be the intersection of AI and the automation platforms such as what you guys do and how does this all come together?

Michael Beckley: Yeah, that’s a big question. I think the simple answer is what brings it together is a workflow. It is a process automation. Appian is a business process automation platform, as you said, and we work with businesses, largest organizations in the world on their missions that matter, on serving their clients, onboarding new customers. For government agencies. It’s regulating those businesses, making sure that the decisions made in financial services are fair, that business is done in a compliant way, and that customer cases are resolved quickly and to everyone’s satisfaction.
And so, when we look at AI, where’s it going to fit? How are we going to get real value from it? It’s about AI being a partner, not a substitute, for human work. And to do that, to make AI a true partner, you inject it into a workflow that already exists, and you have it work side by side with people, and you make sure that as a workflow participant, AI is able to access the right data.
But AI is strongly, in our opinion, something that will propose decisions and humans make them. AI will be a great member of your team, but ultimately routing work to and from AI, that’s where we get the value from it because we’re injecting it into a workflow where the humans are in control of those decisions.

Mike Vizard: How will that manifest itself? And I’m asking the question because so much of the AI interface is a natural language interface. And historically, we’ve used low code platforms to provide that interface. So, what will we present to the end user? Is it still going to be something graphical or is it just going to be like, type in your natural language query here and something magical will happen? How does this manifest?

Michael Beckley: Yeah, we believe that there is a great role for process and workflow diagrams, quite frankly, as a way to provide transparency and auditability into what AI does. So, there’s still a chat component for the way natural language allows a human to say, “All right, let me interrogate some document, very large document, and I’ll ask AI to find the relevant factors in it.”
If, for example, we’re working with a global investigative agency right now, and they get these large reports in for surveillance, people have been watching the bad guys, and you want to ask, “Well, who did the suspect leave with?” And you don’t want to have to read through a hundred-page transcript. The AI can do that faster for you.
But to understand how AI makes important decisions, we want to visualize them. To be able to audit how a mortgage was approved or rejected or proposed to be approved or rejected, you want to do that with process models because they provide the best, most accurate way, to logically represent a workflow and in a way humans can understand.
And so, that’s the important element here is AI, if it generates lots of code for us, well then very few people can read it. But if we use AI to generate and change and modify workflows, well, at a glance, a human can understand what are the inputs, what are the outputs and validate that the AI is working as expected.

Mike Vizard: What is your sense of what percentage of workflows today are actually digitized? My sense is, for all the hype around digital transformation, maybe 10% if we’re lucky, and of that 10%, half of that is probably not working all that well. So, will AI make it simpler to automate a lot more of these processes in ways that are more interesting and more compelling?

Michael Beckley: Well, you raise a good point in that digital transformation itself is nowhere near a completed project. And yet, we think that that foundational work of modeling a process and digitizing it is a critical component of getting the value from AI. AI needs data. Data is everything to AI. And so, before we can even really effectively use AI, we want to harness the power of digital process and technologies like our data fabric to marshal all that data together to bring it and digitize it.
This is a good point, before we talk about data too much, to emphasize that when we say AI, what do we mean? We’re typically these days talking about ChatGPT large language models, we’re talking about new generative AI technology, but we shouldn’t mistake the fact that there’s this new really cool powerful technology for the reality that there is a whole family of algorithms that we should be looking at using together.
Lightweight AI algorithms are great for digitizing, for solving the problem you’re talking about. We have all this paper out there. We still have about $12 billion of waste a year in just financial services and insurance alone where we have paper processes. And so, we digitize those and AI can help dramatically accelerate that by simply using lightweight AI, low-code algorithms for OCR, for classification, for extraction, for document understanding.
And then, we prepare that data. We’ve learned something from your paperwork. We scan the bills, we scan the invoices, we scan the different unique ways in which all of your different insurance brokers collect applications, and then pass them on to a reinsurance company in a standard structured way with lightweight AI models. And then, generative AI can do really interesting things, like understand all that data, automatically generate responses, do a lot of the manual email interactions that used to require humans to type out each thing.
But my point is, generative AI doesn’t replace lightweight AI algorithms. Lightweight AI algorithms allow us to feed generative AI more efficiently. And that’s so important because generative AI models are huge, and that means that they’re computationally expensive, and the GPUs to run them are in short supply.
So, what you want to do is as much work as possible with really simple AI, which costs a thousandth of a cent per query before you run it through a big heavyweight generative AI algorithm. And the end result is we can digitize… I’m less pessimistic than you are. I believe we’ve probably achieved about 20%, 30% of the achieved target market for process automation, but we can digitize the rest with a combination of low-code AI skills and generative AI.

Mike Vizard: I’ll concede your 10% to 15% difference, but a lot of the processes that I see that are so-called digitized, they’re really just somebody quite literally lifted and shifted the paper-based process onto a bunch of mobile tablets, and they didn’t really re-engineer the thing.
Well, generative AI surface recommendations to make the processes more efficient because today I feel like all we’ve really done is shift at the responsibility for filling out the paperwork from somebody who works with the company to the customer, and they’re not all too excited about that approach.

Michael Beckley: Yeah. Certainly, there are plenty of examples we can point to of poor attempts at digitization or self-service processes where the work is just passed to the customer. But a good process automation platform provides opportunities for actually automating the data entry and automatically performing data validation and automatically capturing information that was previously having to be manually keyed.
And the secret to solving that problem isn’t AI yet. It’s a data fabric. A data fabric eliminates the data integration requirements. A data fabric allows you to access all of these different APIs and systems out there inside and outside of a company so that when a customer’s filling out a form, we filled out most of it for them, and we validated the information is correct automatically.
So, another human doesn’t have to manually re-key that data somewhere else. And we’ve automatically linked that data input to your many different underlying systems, whether it’s here. Imagine onboarding a new employee, you’ve got to connect them with IT systems and provision them with that. You’ve got to connect to financial systems and make sure that they get paid and that their benefits systems are also updated so that they get health insurance.
And so, anything we do in the modern world is complex. There’s many different APIs, many different data systems. A data fabric allows you to operate with all that remote data as if it’s local and do it without expensive and complex data integration and without creating database views and stored procedures to make it perform well. So, that’s the essential first part to making processes great experiences for employees, for customers, for partners.
And then, the other critical element is that data becomes so much more valuable because with AI we can act on it in real time. In the past, that was still too much data for a human to process in real time. Now, with AI, we can take all that data fabric combined information and summarize it, synthesize it, analyze it, understand it instantaneously.
And so, that’s where your data becomes so much more valuable, and those processes can become so much more efficient. And we can turn those words into new workflow improvements by simply asking the AI, “Here’s all the data, can you optimize this onboarding workflow for me? Which steps do we really need here? Give me some suggestions as to how I might eliminate a few steps. Because I’ve already got the information I need, I don’t need to do things the old way where I asked three times for your income.”

Mike Vizard: How will the way we build these workflows change? I feel like we’re still doing the same old methods we used in the past. A developer, they may show up with a low-code tool, takes a crack at something, and then they give it to a subject matter expert to review, and they come back with their suggestions and recommendations. And then, we toss the stuff back and forth until finally we have something we’re happy with, and it takes a long time. Is there a more efficient way to think about doing all this going forward that reduces the time it takes to build these things?

Michael Beckley: The end result we want is an optimal process that can change and grow with us, that can reflect reality, which is business changes all the time. So, a business has new products they want to launch, they have a changing context of regulations they operate under that evolve over time. They have new strategies they want to deliver every year. And the end result is we want to have an AI which helps us stay on top of those things.
To get to that, to change the way we go through the traditional software development life cycle you talked about, which as agile as we want to make it, is still requiring us to manually gather all the data, and then write a bunch of code and maybe low code can accelerate writing to deploying that code. But then, when we want to make change, we have to go rediscover where are the business rules because they’re buried in code.
So, a truly model-driven approach that starts with a data fabric allows us to eliminate all the time wasted with data migration and data integration, and then bring that data to the AI so that we can engage in a collaboration between the AI and the developer where the developer says, “All right, here is my workflow. Let’s make a change to it. Let’s change the business rules around what are the eligibility and enrollment criteria for this new benefits program. Let’s raise the income threshold. Let’s target a new set of demographics, and let’s see what the end result be. Show me what that looks like.”
And rather than the end result being some opaque software that millions of lines of code someone’s got to understand, it’s a workflow diagram. It’s a decision table. It’s a form with clear skip logic. It’s the sort of thing that a visual set of models underneath a process that a developer can instantly engage in this useful dialogue and see how it improves. And if it doesn’t, they have the low-code tools there to just make an adjustment and get to the final result.
So, our strong opinion about AI and the future of AI is that if humans are going to remain in control of our complex world and our complex systems, we need to have a language that we can collaborate on with AI that’s more efficient than natural language, and that is a visual model. That is the process model.

Mike Vizard: There are, of course, a lot of workflow platforms out there these days. What should folks be looking for to determine which one to go with? What are the attributes that you’re hearing from customers that have the most important today?

Michael Beckley: Yeah, it comes down to three simple things. The first is the data, and so customers should be looking for a data fabric that allows them to work with remote data as if it was local data, eliminates the data migration requirement, and dramatically accelerates the speed at which they can deploy new processes because they have access to all the right information.
The second is the process engine itself. How flexible is it? Can it actually reflect all of the vagaries of human work and exceptions and ad hoc types of work. Is it as appropriate for humans as it is for events and documents? Usually, workflow engines have been built for one use case, but your business in today’s world has to span a variety of different types of work to effectively complete anything meaningful. And that’s form-driven workflow, event-driven workflow, microservices orchestration, as well as of course handling all of the complexities of human decisions and human exceptions. And of course, it has to scale to many millions of processes a day.
And then last, data plus workflow, you have to do something useful with it. So, what kind of integration do you have to a total experience that is going to be compelling for customers and employees and partners? And so, does that process engine tie directly into web mobile, and even future virtual augmented reality experiences?

Mike Vizard: Are these workflows getting more complex? It seems to me, even in my own organization, everything we want to do has at least a half a dozen APIs involving some external provider that I don’t have a lot of control over or visibility into. So, how do I extend a workflow that I created internally to encompass all those things if they’re not on the same platform that I am? Is that challenging? Or how do we solve that particular issue?

Michael Beckley: Well, it used to be very challenging, but now technologies like a low-code business process automation platform like Appian, the way we address it is, first, by using a data fabric to connect all those APIs and allow you to have control over how the data works, even though the data isn’t directly under your control.
The data fabric allows you to combine data elements, create new custom fields, allows you to guarantee consistent performance, eliminate the latency of calling those remote target systems. So, the data fabric automatically provides you this virtual data layer that you have total control over, and then synchronizes any changes you make back to those target systems for you, so that you still have the consistency of the data and the atomicity of the transactions.
The same kind of asset guarantees that a developer would expect if they were working with a local relational database. That’s an incredibly powerful technology that simplifies the complex world we’ve inflicted on ourselves, because microservices are great, APIs are great, everything’s got one. But then, now you’ve got to deal with 12 or 20 or 200 different APIs of requirements data systems. So, data fabric simplifies all of that.
But then, how do you still be a good corporate citizen architect and deliver back the value of what you’ve done to other developers, whether they’re in your organization or outside of it in a true platform economy? And so, with technologies like Appian, once you’ve created something new and wonderful, that new definition of what a customer is that combines insights from your different systems, well, you can publish that securely as an API or as a microservice itself, a microservice portal that can be independently deployed and scaled elastically with complete low-code.
So, that’s the value of a low-code approach to business process automation is that it truly allows you this end-to-end automation in terms of gathering all the data, using every bit of insight you have about your customer, your products, your business, and then deploy it back to the world through a new microservice that anyone can consume.
And if no one shows up, it doesn’t cost you anything, but if a million people show up and want to see that data, well then, it’s no problem. You can elastically scale that independently of the rest of your environment and all without writing a line of code, and that’s the true beauty of a model-driven approach.

Mike Vizard: What’s your ultimate best advice to folks? Sometimes I see folks trying to boil oceans, other times I see folks they’re doing something outside of a context of the larger enterprise, and then it just becomes an orphan. So, what are you seeing that people are doing to be successful when they do attempt to transform their business?

Michael Beckley: Well, it begins by actually deciding what kind of transformation you want to undertake. The reality is, today’s modern enterprise has a number of different transformation projects that it’s undergoing and are required. And so, my advice is, first, diagnose what you’re trying to do, and then choose the right platform to support it.
Everyone comes at this with their own background and their own toolkit, so you’ll see people inside an organization advocating for using the CRM platform to solve every problem, or trying to use the IT platform to solve every problem, or trying to use their financial system to solve every problem. And the fact is, if you are just working with IT and IT data, then maybe that’s fine. Or if you’re just working with the customer data in the CRM, then a transformation project can probably use the CRM tools and that’s fine.
But when you want to bridge a process that needs to be transformed about how you onboard customers or about how you regulate an entity or how you comply with a new regulation that affects data and requires data from many different systems, then maybe use an agility platform that’s purposely built for that requirement, one where you can actually work with information equally well if it’s coming from CRM, ITSM, HR, Finance.
And so, that’s where you choose a special purpose business process automation platform if you have a process problem. If you have an IT problem, use your IT platform. If you have an HR platform, use your HR platform. But if it’s truly something that spans those platforms, then choose the platform that’s designed for customization, that’s designed to unify them all.
And then, think about after you’ve modeled your process and you’ve gathered your data with your data fabric, how can I use AI to dramatically accelerate that transformation? But don’t start with the AI looking for a problem. Start with, “What is my process?” And then, it will naturally fall out, where there’s still large amounts of data that AI can read for me, where there’s a lot of manual labor that AI can actually assist with, and then AI can truly be a partner in transforming that process. AI can be a workflow participant that helps people make better decisions, because they’ve got better access to information where AI proposes, and yet it’s the human that decides.

Mike Vizard: All right, folks, as always, it’s so important to take a good long look before you leap, right? And here we are. Michael, thanks for being on the show.

Michael Beckley: Yeah, my pleasure, Mike.

Mike Vizard: Thank you all for watching the latest episode of Techstrong.ai. You can find this and other videos on the techstrong.ai website. We invite you to all to take a good long look at all of those as well. Until then, we’ll see you all next time.