Artificial intelligence. Is there any term that’s more used in tech these days or that has a wider range of meanings? Any one that conjures up more excitement, hyperbole and fear? In this episode Jon Prial talks with Adam Devine, the CMO of WorkFusion, one of Georgian Partners’ newest portfolio companies, about a very practical application of the technology: Using AI to improve and even automate what have traditionally been human-driven processes in the workplace. You’ll hear about robotic process automation, an emerging field that is bringing AI-powered software robots into the workplace to help make companies more efficient and effective.
Jon Prial: Artificial intelligence. Is there any term that’s more used in tech these days or that has a wider range of meanings? Is there any one that conjures up more excitement, hyperbole and fear? Today, we’re going to focus on a very practical and a real application of this technology, using AI to improve and automate what have traditionally been human-driven processes.
We’ll take a journey, looking at how technology has evolved to help automate the work of traditional back-office business processes. The latest step in the evolution has been the development of robotic process automation, an emerging field that’s bringing AI-powered software robots into the workplace to help make companies more efficient and effective.
We’ll find out how on today’s episode, when I talk to my guest, Adam Devine, head of marketing at WorkFusion. WorkFusion is one of the newest members to our portfolio, and it’s using AI to help large companies use intelligent automation to work more efficiently.
I’m Jon Prial, and welcome to the Impact Podcast.
Jon: At one point, I was looking at a survey. I’m not sure if it was on your website or something I found, but McKinsey had said that 49 percent of the activities that people are doing today in the global economy can be automated with a currently demonstrated technology. Can you take me through your view of what you think of when you think of automation?
Adam Devine: Sure. First, I would invite everyone listening to close their eyes and imagine the huge expanse of a back office of a large financial institution or insurance company. Hundreds if not thousands of super-smart, capable people spending 30, 40, 50, 60 percent of their day doing things like operating the UIs of SAP or Oracle, super-repetitive swivel chair work, or looking at a PDF on one monitor and an Excel sheet on another monitor and simply, routinely transferring the information from that PDF, which you can’t manipulate, to an Excel sheet.
I think McKinsey is very much right. There’s a high percentage of work that the average so-called knowledge worker, people who work with information all day long, can be automated.
Jon: The thought of taking the data from the PDF to the Excel spreadsheet has to get codified somewhere. How do you approach that in terms of that’s something that could be done more efficiently, that needs to be automated?
How do you figure that out? How do you get the algorithms behind all these changes, perhaps?
Adam: There is this notion of writing rules or having rules learned. In the old days, like two years ago, there was scripting — if/then/else automation. You’d have teams of engineers and maybe some data scientists writing rules for scripts to follow, and that meant, as you say, codifying each and every action that a machine would take so that there is absolutely no ambiguity about how the work is done.
This, today, is an old-fashioned way of automating a process. What we can do today with machine learning — and it’s not just our business, this is a growing trend — is having machines that learn. Learn is the key word.
Rather than writing the rules, people do as they do. They open up an Excel sheet. They open up a document. They click here. They click there, and over the course of time, machines can detect patterns that people can’t. This is what I mean by learning. Where someone clicks on a document once it’s been digitized, what the context is of that information.
With enough repetition — typically 400, 500 repetitions — the software is able to identify a pattern and train an algorithm to do what a person had been doing.
Jon: I started, one of my early careers we did a lot in the world of workflow and image processing, taking electronic versions of paper and moving it through a process, maybe reading the paper, managing workflow. That evolved from paper-based processes to human-based processes.
Can you talk to me more, then, about robotic process automation, what that market is and what it was a few years ago and what it’s evolving to?
Adam: Sure. If workflow yesterday was the movement of paper, the movement of information, RPA is one level above that, or one step up the ladder, in that it doesn’t just move the information, it can transform it and transfer it. A good example would be moving structured information from SAP to Oracle or from Oracle to Workday.
These are systems that don’t inherently talk to one another. They’re different formats, and they require what you’d call human handoffs between these applications. RPA can operate these systems at either a UI level, meaning at a virtual desktop a bot will enter credentials automagically and run an operation to do a transaction or to move the information, or it can operate at an API level where — I guess you could call it— diplomatic code serves as an intermediary between these two applications.
I would say that RPA is the next level up above old‑fashioned BPM or workflow.
Jon: Does that involve AI, or does AI then come to the next level?
Adam: It can involve AI. One of the problems with scripting and with RPA is exceptions. What happens when something changes about the process or the content and the bot, which has been programmed to do a very defined task, says I don’t know how to do this? That means the process breaks. That means the bot breaks.
What happens, with just RPA, is that a person discovers that a bot has broken, because the business process has failed, and has to go in and manually retrain that bot and fix the business process. When you add AI to RPA, you have automated exceptions handling. You have an intelligent agent identify that the bot doesn’t know what to do and route that work to a person.
The person handles the change if the bot can’t figure it out, and that creates a contribution to the knowledge base. It teaches the bot what has gone wrong so that the same mistake or a similar mistake doesn’t happen in the future. What AI does for RPA is business continuity.
Jon: When you talk about RPA getting improved by managing the exceptions, and you’re managing the exceptions because you’re learning things — it’s a learning opportunity. Obviously, you’re learning from data. What new type of data is being brought in to a system to allow that learning to take place?
Adam: There is a lot of new learning that takes place when AI assists RPA. One of the more interesting things is workforce analytics. Rather than having opacity around who your best human performers are, around what their capacity is, what their capability is, what their aptitude is, when AI gets involved and can monitor the actions of a person that’s intervening in a process, you very quickly figure out who your star performers are and what they’re good at. You very quickly have transparency on what the capacity is of a workforce and how work should be routed.
A good example would be the back offices of a large bank. Most offices are highly distributed across Latin America and India and the US and Europe, so when workforce A in Costa Rica blows out of capacity or doesn’t have the capability, AI can look at that workforce and say, “OK, I’m going to move this task, this business process, to a supplementary workforce in the Philippines or in Omaha.”
The number one set of data you get when AI is involved in a business process is not just the automation of the work, but an understanding of how people are performing it and how best to perform the work in the future.
Jon: As you get started, as you do an implementation, I assume the first focus area is how to make a process better and focus on that data. I know you even do some crowdsourcing of data around that. Let’s talk about making a process better, and then we’ll take a step back and do a little more about the people.
Adam: We get this question a lot, about how our software enables transformation. I was talking to an executive from a shared services organization just yesterday at a big conference down in Orlando, and I used the word transformation, and he flinched. Apparently, transformation is a four-letter word in a lot of these big organizations.
They’re not necessarily trying to transform. They are truly trying to automate. What we see is that by using software such as ours, there needn’t be a focus on transformation for the sake of transformation. When you allow an intelligent automation to do its thing by automating — for example, import payments in trade finance, or claims processing in insurance — the byproduct of automating that work, by letting algorithms see how data is handled, see what the sources are, see how people extract and categorize and remediate information and thus automate it, the process, the byproduct of this automation is transformation. Does that make sense?
Jon: The transformation, it still involves automation. I’m talking about the conflict you had with the customer you were talking to. Doesn’t that transformation get them to automation, or not? I’m trying to think what the end goal might be here. They’re not mutually exclusive, are they?
Adam: They’re definitely not mutually exclusive. Most businesses simply have a remit to either cut costs or improve service and capacity. It’s one of those two things, and in these days, it’s both. Most shared services, product lines, operations, wherever the genesis of automation is, wherever the genesis of these initiatives are, they’re starting with their KPIs.
Their KPIs are not impacted by simply transforming work. Their KPIs are impacted by eliminating the amount of manual work done in the operation. That elimination of manual work and the freeing up of human intelligence to focus on higher-value work is, in effect, transformation.
Jon: The results, you’re looking at the KPIs and you’re getting better business results, then everybody should be happy, because the topline numbers matter the most.
Jon: In terms of industry, you’re mostly in fintech, but what do you see is the opportunities for the automation of these types of processes against different industries? What’s your take on that?
Adam: That’s a great question. Martin Ford, who’s the author of “Rise of the Robots” — we’re actually featured prominently in that book — super-smart guy, true futurist, he spoke at this conference I was at yesterday in Orlando, the Shared Services and Outsourcing Week, and he said that to ask what the impact of AI will be on different industries would be like asking the impact of electricity on different industries.
His perspective, and I share it, is that AI will have a ubiquitous impact across every industry. It’s going to touch everything that we do. We’re not going to feel it until it’s ubiquitous and we stand back and say, wow, it really has transformed everything.
To drill into it specifically, we at WorkFusion have made a strategic choice to focus on banking, financial services, and insurance. We’re now getting into health care very quickly. We have a lot of interest from utilities, from telecom. I don’t think there’s any one industry that we won’t touch. It’s just a question of sequence, and it’s also a question of internal drive.
Banking and financial services, because of regulatory compliance, have had an unusually high amount of pressure to digitize, to automate. I think health care is very closely behind, and then the general Fortune 1000, I don’t think there’s been quite as much pressure, but some of the things that’s happening with the optics of offshoring and outsourcing will probably catalyze automation efforts even faster.
Jon: Let’s talk a little bit about the analytics. A lot of it’s rooted in basic machine learning. There are semantical challenges sometimes for people understanding the difference. I see a lot of you saying this is AI, and it’s really just machine learning. Give us your thoughts on how a technology like machine learning can evolve into something a little richer in terms of a solution set with AI.
Adam: As I understand it, machine learning is really the only practical application of what we refer to as artificial intelligence right now — algorithms that take in massive amounts of data and are configured, the feature sets are programmed to do something like extract data from an invoice.
Another subset of that question is what’s the difference between analytics and AI. A lot of businesses, as they get into their AI journey, confuse the two. Machine learning automates the manual work in business processes. Analytics, that may or may not be powered by AI, tells you something about the way a business is performing. These are two very different things.
Automation replaces manual effort. Analytics tells you something about the way things are happening. That’s the simplest definition I could give.
Jon: You earlier mentioned about training. What’s your approach to training and making sure you’re learning the best algorithms, you’re actually codifying the right actions — avoiding biases, avoiding codifying bad behaviors. What’s your approach to training?
Adam: There are two things to remember about how intelligent automation and machine learning learns to do the work of people and automates the work of people. The first problem is how do you get lots and lots of good quality data. This was a problem that we solved back in 2010 at MIT’s Computer Science and Artificial Intelligence Lab, which is where the company was born.
The researchers back then used the same approach to identify human quality that banks use to identify fraud in financial transactions, and that’s called outlier detection. If there are 100 workers doing a specific task and 90 of them are performing the same keystrokes, the same speed, on the same content, but 5 of them are going really fast and using only numeric characters, whereas the vast majority are using alphanumeric, and 5 of them are going really slow and using just alpha, then that means those 10 on other end are outliers, and that means that they get an increased level of scrutiny.
Maybe there is adjudication between two workers where two workers do the same thing and the results are compared. That first problem of how do you get quality data was solved by using machines to perform outlier detection and statistical quality control on workers, the same way that assembly lines ensure quality.
The second challenge is how do you take that quality data and train machine learning models. In the old days, you’d need data scientists and engineers to perform countless experiments on Markov models, traditional random fields, deep learning neural nets, and run all of this sequentially to figure out which model and which combination of models and features was going to best perform relative to human quality.
We solved that second problem about three or four years ago, which was to do with software what data scientists had done. We call it the virtual data scientist, where once the good-quality data is generated, the software automatically performs experiments with different models and feature sets and then compares them all in real time.
Once confidence is high enough, that means there’s been a winning algorithm. You can think of it like an algorithm beauty pageant where they’re all competing to do the best work and the software chooses the best model to deploy as automation.
Jon: Does this go across both broad processes, long-running processes across a day or multiple days, as well as some of the small micro tasks that individuals might do? What’s the difference?
Adam: The question is around what’s the level of granularity and application of this process of learning and training. It happens at an individual worker level in a matter of seconds, where maybe a worker handles a task and that makes a small adjustment to the way an algorithm performs, and it happens across entire lines of business where the data generated by hundreds of workers impacts the way automation performs.
That’s the beauty of machine learning is that it isn’t a blunt object. It’s an incredibly specific and — I guess you could say — perceptive capability in that the slightest adjustment by a person can change the way a machine performs.
Adam: You really are sourcing the logic from the data that you’re collecting across multiple companies, individual companies? What’s your view of data rights and learning and normalizing and learning from different companies together?
Adam: Jon, this could be a whole other podcast. We get this question a lot from our customers. Dealing with some of the most data-rigid security concerns, compliance-ridden businesses in the world in these giant banks, one of the very first questions we get from C-level stakeholders is, what are you doing with my data, and do I own it?
There are a couple of different answers within that question. The first is that our customers own their data. You could think of us as a car wash, the car being the data. The data comes into our business, into our software. It is manipulated. It is improved. It is stored into a place in a customer environment, and we no longer touch it. Our software does not store our customers’ data.
The second part of that question is do our algorithms retain the intelligence created by one customer, and can we take the intelligence from one customer and apply it to the next? There is some level of retention, but if, for example, we’re talking about something like KYC — know your customer in banking — we do not and will not take the insights generated by one customer’s very specific, proprietary business process and apply it to another customer’s.
We consider that proprietary to the customer. Would it be nice if we were like Google and my search history could be applied to your search history to improve the search results for all of mankind? Sure, but that’s not what our customers want.
Jon: You mentioned that you might change into other industries over time and grow. Absolutely that makes sense in fintech and potentially in health care, although perhaps learning about how health care procedures work across different hospitals and different solutions they may be willing to share some of the data rights and allow you to aggregate.
Today, the answer is here you are today within fintech. Could that change in other industries?
Adam: It really could, even within fintech. There’s another process called anti-money laundering, and this is essentially massive-scale fraud detection for banks. Banks don’t really consider the way they execute compliance a competitive differentiator.
There is a stream of thought among our customers to pool their intellectual resources on our software to create a ubiquitous software utility to solve non-differentiating processes like anti-money laundering, like KYC to an extent, like CCAR and BCBS 239. There are 80,000 regulations out there or something like that.
Industries and our customers, we may be a forum for them to decide where they want to compete and where they want to collaborate. In healthcare, this is particularly true, given that it is outcome-based. There’s actually a really cool company in San Francisco called Kalix Health, and they have the remit to democratize outcome based on democratizing good-quality data.
I think we’re going to see a big trend in healthcare to do more sharing than siloing.
Jon: Let’s put some CEO hats on as we get toward the end of this thing and think about CEOs. I’m going to make the assumption, the mental leap, that they have already figured out what we’ve been calling applied analytics. They’ve got analytics. They’re beginning to inject insights into processes, but they haven’t really taken the next step yet, in terms of degrees of automation and efficiencies that they could get.
They’ve got better outcomes, but they really haven’t thought about where these efficiencies are. If you’re talking to a CEO and talk about the managing and improving of processes and leveraging of the AI, where would you have the CEO start? Where should he or she start?
Adam: There’s a strategic answer and a tactical answer. Sometimes the tactical one is more insightful. In terms of the geography of a business, we see a lot of companies beginning in what’s called shared services, where there is a large aggregated workforce that serves as an internal service provider to the business, like handling processes like employee onboarding or accounts payable, the high-volume, common processes that different lines of business all employ.
Shared service is a great place to start, because it’s where outsourcing typically happens, and where there is outsourcing or offshoring, there is a large amount of work that can be automated. I would say, tactically, I would say CEOs should look toward their global business services or shared services to start their automation journey.
Strategically, I think every business needs to decide, do I want to do the same with less, or do I want to do more with the same, or do I want to be extreme and do even more with fewer? That’s the existential question of a business. If a business is healthy, they’re going to want to continue to grow their headcount but then exponentially grow their productivity.
That is the power of our software, of other software like it, that can do with machines what had been done before with people. The other thing, too, is how do you elevate the application of human intelligence? You do that by automation. You hire the same great people, but you expect them to perform at a higher level because machines are doing for them what they had done before with their hands and their minds.
Jon: As I’m building my team and I’m thinking about this and I’m going to embrace this model of being more efficient, you mentioned that you deliver a virtual data scientist. How much do you look for your customers to help, in terms of them owning and building a data science team? What are you looking for these companies to provide as they get started?
Companies are going to embrace this. These aren’t necessarily your customers, but customers are going to embrace AI. We’ve talked about data quite a bit. It does get rooted in that. What kind of skills should they be looking for?
Adam: The honest answer is that the only skills that a business truly needs to be effective with WorkFusion are subject matter experts on the process — people who understand the progression of work. Any big, successful company is going to have people within the organization that understand the methodical flow of work — first do this, then do that.
The problem with AI in the past, and actually some other vendors that are out there, is that they’re black boxes, and these black boxes require data scientists and engineers to build — you could say to fill the gap between the black box and the practical business process.
The reason why WorkFusion has been so successful and why we’re going to be such a big, important company to C-level executives across all industry is that there is no need for them to fill the gap between our capabilities and the practical business process, because we are rooted in the practical business process, and we do not require teams and teams of data scientists and engineers.
Sure, if you’re a power user and you want to radically automate across an entire business, you’re going to need some level of technical capability. You’re going to need some java engineers. IT is certainly going to need to provision environments just like they do with any software, but the beauty of this next generation of practical machine learning powered software is that it can do the dirty, highly complex work of teams of data scientists automatically.
Jon: I like the thought of making sure the CEOs stay focused on what they do well — know their business process. Stay in your swim lane. Keep going and the application of AI might get you into deeper water, but you’ll keep going. That’s the key.
Adam: That’s the key. There is the D word — disruption. I don’t see this as a disruption to the way a business works. I think a disruption would be expecting a business to go out and hire 500 data scientists and use some magical black box that has no transparency. That’s disruptive.
What is non-disruptive, what is purely evolutionary but exponential in its benefit, is a software that can seamlessly integrate into the way a business is doing their processes now and make those processes slicker and faster and more automated. That’s what every CEO wants.
Jon: I can’t think of a better way to end it than that. That’s a perfect message. Adam, thank you so much for being with us today.
Adam: Jon, my pleasure. Thank you.