How to Nail Business Automation in the Agentic Age with Micha Kiener, CTO at Flowable

(Intro Narrator at 00:00:00) Today, we're talking to Micha Kiener, CTO at Flowable, about how business automation is evolving around agentic technologies and what he learned going from CEO to CTO. You're listening to Joel Beasley, Modern CTO.

(Joel Beasley at 00:00:20) One of the things that stood out to me a whole lot about you is that you were a CEO that then went to CTO. Can you explain that transition?

(Micha Kiener at 00:00:28) Well, as a founder, when you start up a new company or new product, you have to be kind of everything, right? So from product innovation, I even did product development down to the code, up to marketing, sales. You basically run the company with your couple of people that you have around you. But as you grow as a company, you need more structure, you need more management capabilities. And at some point, I felt like I want to be where my true passion is, and that's product. So that's where I went to become a CTO and had implemented a CEO along the line.

(Joel Beasley at 00:01:15) Yeah. And so today, you're at Flowable, but this isn't your first company, right?

(Micha Kiener at 00:01:19) No, it's not. I actually founded my very first company at age 14.

(Joel Beasley at 00:01:25) Okay.

(Micha Kiener at 00:01:25) Yeah, it was software. It was also software. When I was a young kid, basically, I taught everything myself. So I read books, but back in the days, there was no, you know, like we know today. So it's pretty hard to get knowledge and stuff. But yeah, so kind of a self-taught, the tall person, so to speak.

(Joel Beasley at 00:01:49) I like that. I think that we have that in common. So at 14, what type of problems were you solving?

(Micha Kiener at 00:01:54) My first gig—I don't remember the name of the software, but I got my hands on some software. I cracked it. That was still, you know, back in the days with dongles and all that stuff to protect software. And I approached them saying your software is not secure, but I can make it secure. And that gave me my first gig. And in instead of getting money, I wanted a faster computer because they were super expensive in those days. Yeah, that's kind of how I started.

(Joel Beasley at 00:02:29) So you exchanged patching up their software for compute.

(Micha Kiener at 00:02:32) Exactly. And bigger hard drive.

(Joel Beasley at 00:02:36) Yeah, of course.

(Micha Kiener at 00:02:37) Oh yeah.

(Joel Beasley at 00:02:38) And so, and then now you're at Flowable. And what does Flowable do?

(Micha Kiener at 00:02:41) So we basically have an intelligent business orchestration platform. And recently, in the last couple of years, we added the agentic capabilities to it because we feel like that adds the next level of valuation in such an orchestration platform. We basically orchestrate and digitalize end-to-end processes, if you will. That really goes across business units, across systems to build full end-to-end user journeys.

(Joel Beasley at 00:03:14) Can you go a little bit deep? Because that kind of means a little bit different to everyone. As you're saying that, things are popping in my head from, like, business process management all the way through customer onboarding. There's all of these things that are popping through my head when you use those words. Can you boil it down for me?

(Micha Kiener at 00:03:29) Yeah, absolutely. You nailed it. So when you said something like customer onboarding, that's a really good example that actually crosses a lot of different systems and business units within a bank, for instance, right? From client advisory back to back office, compliance, there's just a lot of people involved, typically systems. That's what Flowable does. It basically is—you can think of Flowable like the glue sitting on top of different systems. You don't want to put that data into every system, like from CRM to ERP, core banking system, maybe your compliance system that you have yourself. You want that to be fully automated, and that's what we call orchestration. Orchestration across the different business units, systems, people, various integrations. So you really have to be able to integrate with the different systems. You have to be able to get people on board that review stuff or fill out information. And if you really want to truly go end-to-end, you also need something that we call multichannel capabilities. It can reach out through various channels, including email, online, like through a portal, e-banking, if you're in the financial world. Even up to real-time messaging like WhatsApp, text messages, WeChat. All that stuff needs to be fully aligned in one platform.

(Joel Beasley at 00:05:01) That's pretty cool. That's a whole lot going on. I like the word orchestration too. You must have, like, a project inside of your software called, like, Conductor, you know, that groups everything together. No, that's great. That is really cool. And how did you first identify this as a problem in the marketplace? Because it takes a lot to start a company, man. And for you to commit to it, you must have really seen an opportunity.

(Micha Kiener at 00:05:28) Actually, when we first started out in commercial software, always in the era of open source, open source users at the beginning, contributors, and eventually our product has an open source core as well. But the idea came along when I was implementing an end-to-end solution for money transmitting. You probably know things like Western Union. That was in the year 2001, 2002-ish, something like that. And we started to implement a lot of different rules and process snippets, which are process fragments, because every country had its own regulation, its own rules, and its own set of how the money actually gets transmitted or wired. And it came to our mind—first, we started to code everything, right? That was how you did software back in those, in that era. And afterwards, I realized, like, wait a second, I need to be able to do that more descriptive through a model instead of code. That's where the foundational ideas around process management, case management, started to light up in my head. And yeah, that's when, basically, we started creating the first process and case engine.

(Joel Beasley at 00:06:52) And when you're working with a lot of big companies, big banks that have huge processes, when you get to work with these partners, do you notice, like, when they first come to you that they're doing a lot of wrong things? Are they getting things wrong? Are they getting most of it right and you're just the tool that's going to help them put it together? Tell me about that.

(Micha Kiener at 00:07:10) I wouldn't say they're completely doing it wrong. No. But—

(Joel Beasley at 00:07:15) I just asked you to talk badly about your clients. Let's say talk badly about your clients. Let's do that.

(Micha Kiener at 00:07:22) Yeah. Let's say they're not getting into their full potential. Let's call it that way. So for instance, to give you some examples, a couple of years ago, they started with something they called automation. So single tasks getting automated. You could do that through various ways. You probably know about RPA, robotic process automation. That's where you have these kind of robots that did something that you would do as a human, and you automate it. But you automate single tasks, but there was no—and to pick that word, orchestration again, there was no orchestration. So nobody kind of orchestrated or automated the automation, if you know what I mean, right? And working in what we call silos, that's still—and especially in agentic AI, that's currently the same pattern repeating itself. So you implement those single siloed agentic solutions, a single agent that answers some questions around your policies or rules for, you know, financial institutions, cross-border rules, whatever. And that has value. But if it's not orchestrated in an end-to-end way, you lose a lot of potential on the ground. You still have to trigger it left and then go right and then go to the next one, or you try to add another automation that takes result A and puts it for the other automation in B. But that's really hard to control. So that's where we believe, like, you need this kind of horizontal layer that orchestrates systems, people, data, events, agents, AI agents, as well the same way.

(Joel Beasley at 00:09:11) And you have four types of agents. Is that correct?

(Micha Kiener at 00:09:14) Well, in the meantime, we are close. In a couple of days, we have our next release. So we even have more. The foundation is still those kind of, like, four agents that we have—or agent types, I would say, not four agents, but four agent types. So utility agent that helps you creating very small tasks, like extracting structured information from an unstructured type of content, like an email, for instance. Extracting information like date, payment, account numbers, addresses, or what have you out of unstructured information. That's what we call utility agent. Then we have a document agent. Like the name implies, it works around content or documents in general and can do classification. So let's say if you have a document coming in through email as an attachment, it can classify: are we looking at a bill? Are we looking at an account statement? Whatever it is. And the second part typically involves extracting structured information from that document. And that's the document agent. Then we have what we call a knowledge agent. So the other two, they are based on what we call language, a large language model, right? Public knowledge, so to speak. But the knowledge agents can actually be on top of your very specific company-based knowledge, like your specific rules, your policies. If you're an insurance company, for instance, the knowledge might be the policy that you're currently looking at for a customer, and the agent can work on top of the public knowledge with your very specific knowledge. And then the most sophisticated one we call the orchestrator agent. That one works on top of what we call a case and tries to solve that case end-to-end by orchestrating itself. So there's no kind of a predefined path, like do A and then do B and then do C. The orchestrator agent is more driven by natural language, like you instruct it what is the use case, what is the goal of the case, what are the tools at your disposal. And then it figures out the best way to solve and to handle a case. And then in the meantime, for the new release, we have new agent types more for external agents. There's a new emerging called A2A, agent-to-agent. So we can also work and orchestrate external agents, you know, like from AI Studio, from Azure, or AWS Bedrock, Claude. So different external agents that can be working together in the Flowable platform.

(Joel Beasley at 00:12:05) I'm excited right now, and we're getting into it. Okay, the orchestrator agent, just quick question before I dive deeper. You say you keep saying "tries to solve a case." Like, give me an example of what's a case.

(Micha Kiener at 00:12:16) Cases typically are on the probably more complex side of business use cases. That's also probably where the name evolves from. So give you an example also from the financial world, like source of wealth. If you're a customer getting onboarded into a private bank, it's super essential that you can declare where your money comes from. Might be from, you know, employment, maybe you had your own business, you sold it, you had dividends, you had inheritance, whatever it is. And that's a rather complex process. So the bank really needs to hunt all the different statements you declare as a customer where your money comes from to find proof. Does that make sense? Also do benchmarking to figure out: is the salary given the function you had, does that make sense, and things like that. So solving a case means at the end of the day, the goal is that all the money that you declare as a customer is kind of, you know, proven that it makes sense with statements, benchmarks, and things like that. Maybe you also need to find public available evidence like, you know, selling a company or whatever. And that's complex. It's not a straight line. You cannot say, oh, I need A, B, C, and then I analyze, and then I'm done. Maybe we can even loop back to the customer saying, hey, for this inheritance, do you have proof? Can you give me some documentation around it? Or you sold a business five years ago. Can you give us the contract, the selling contract, so we can extract information that matches your declaration where the money came from. And so solving a case means the goal is to figure out or to declare all the money the customer to be valid, what they tell you, where the money comes from. And that involves, you know, reaching out for public resources, maybe manually asking the customer to provide more evidence, more documentation, extracting that information from the document to make up the whole narrative, the whole story behind the money. And the orchestrator agent then orchestrates all the given tools that might be another agent, like a document agent. So the orchestrator agent would then tell the document agent, hey, I have new evidence from the customer. Classify it. Extract that information that I'm missing or that I need to. Or it could invoke a loop back to the customer saying, hey, I need X, Y, Z. Or it could even involve a human in the loop and saying, can you review this piece of information, things like that. And that's what we mean by orchestration as a general means of solving a case end-to-end.

(Joel Beasley at 00:15:13) Okay. So yeah. And let me kind of regurgitate it from my experience. You tell me how close I am. You're essentially just building these different specific agents, like, almost like services, where it's like, this is our authentication service. You know, if you're on a SaaS engineering team, you have a team that runs that service or you have a team that runs a billing service or a payment service. And so they're kind of like separating into their own little unique functions as core areas that you need to be competent in: utility, document, knowledge. And then when you have some larger project within the company, you can use this. The orchestration agent is essentially your placeholder term for this is the one that's going to coordinate across all of our available data. So it could be at a bank, it could be the source of wealth thing. At a SaaS company, it could be solving a customer's problem, looking across billing systems and past, you know, analytics data of where they visited and what pages they visited and errors they've experienced and all of this stuff. So it can mean a lot of different things at a lot of different places. These are just the names you're using.

(Micha Kiener at 00:16:17) Absolutely. So it could be any sort of use case we're talking about. Could also be something super simple like unstructured payment. That's still a thing that people just send emails like, hey, I need you to wire X, Y, Z to this account, this person. And then the orchestrator agent has the task to, you know, extract all the information, make it structured, validate it, maybe depending on some rules, adding a compliance officer into the case or asking more information from the customer if something's missing. And could be up to what we just discussed, like the source of wealth complex use case. And it typically tries to break it down into several, you know, like, smaller pieces along the line to handle it. That's what typically—and the cool thing in Flowable is you can mix and match between services, humans, agents, and even what we call micro processes or sub-processes.

(Micha Kiener at 00:17:19) And that's where the case comes in. That's the beautiful concept of a case where it's kind of like the umbrella. It's kind of like the context that you add all your information, data, whether it's structured or unstructured, content, documents, humans, events, services. That's all put together or kept together in this context, in this case.

(Joel Beasley at 00:17:43) So my next questions are going to be around monitoring and correction. Right? So you're doing these things. You're solving these cases. What are the tools like?

(Joel Beasley at 00:17:53) Because I haven't been hands-on software engineering on like a nine-to-five basis in probably four years. That preceded about twenty years of doing it every day. And so I'm not like in there, but explain to me how mature these tools are for watching what the agent did or what happened with the case and then correcting it for future cases.

(Micha Kiener at 00:18:22) That's a really good question because that's one part where we see a lot of companies struggle at the moment. If you look at the market, agentic AI grew up within the last couple of months and years so fast. And you always had this kind of like fear of missing out as a company, like, "Hey, if I'm not doing anything in AI or agentic AI, I lose against my competition or whatever." So a lot of companies just did something. Right?

(Micha Kiener at 00:18:54) I need to do something. So let's just use some sample or easy tooling or whatever out there and then put something in production. But a lot of them missed things like, what about maintainability? What about traceability, transparency, the guardrailing, permissioning? That typically was, in a lot of cases, just left out.

(Micha Kiener at 00:19:18) You give access, you give agents access to even a lot of your data you would never do that to an employee. Right? So that's super dangerous if you don't have that full control. And we are coming from a totally different angle as Flowable. The last couple of years, we spent our time creating a platform that is fully enterprise-ready from, you know, maintainability, versioning, putting adoption or changes of your processes into life, adding feedback like, "Oh, something we need to improve or whatever," and making it transparent, auditable. Everything that happens in the platform gets audited, gets recorded.

(Micha Kiener at 00:20:03) And we just added agents on top of it, on that platform. Whereas there's a lot of tools out there—they started from scratch. They do some agentic stuff, but they say, "Well, let's do that later. Let's add, you know, traceability, all that stuff later."

(Micha Kiener at 00:20:20) And I think that's the danger that we see in a lot of customers today. But next year will be the year where a lot of this kind of enterprise readiness needs to come in. Otherwise, you will fail big time.

(Joel Beasley at 00:20:38) Yeah. I haven't seen really cool—and I haven't been looking for it either—but I haven't seen really cool technology around like agent permission management, you know, because some people just seem to be like, "Well, we're just going to give it all the data at the company," and others are very specific. I think you're going to end up wanting to give certain agents certain permissions to certain datasets and then have other agents that work together because I think it's going to end up modeling like our employee hierarchy, how it works. Right? Not everybody has DBA-level access to everything. You know?

(Micha Kiener at 00:21:10) Yeah. And there's something new that's also in our opinion—you probably heard terms like prompt engineering, I guess. So how can we improve the instructions that we give an agent towards the desired outcome?

(Micha Kiener at 00:21:26) But we feel like context engineering is even more important. What I mean by that—I mean like giving an agent as less data as needed to get the desired outcome. Several reasons for that: the more restricted sort of data you provide to an agent, the better the outcome. It's the same with humans. If we have like a hundred pages we have to read and get a couple of important information out of there, or if that's kind of like prepared and we already worked through a summary with two pages, it's easier for us to grab the essential part of that smaller amount of information. Right? It's no different with agents. And beside the side effects, like it's faster and cheaper if you don't have big context. And that's where context engineering is super essential. And the case, again, helps you to build different or various different, what we call, scopes of data that you give to or give access for specific agents.

(Micha Kiener at 00:22:30) So for instance, the inheritance agent that only focuses on the inheritance piece of your wealth, if we go back to that same use case. Why would you give information like from all your employment, all your payslips? It's not relevant. Right? It could just mess up the whole thing.

(Micha Kiener at 00:22:49) That's where context engineering tries to solve in a way like, "Hey, this agent only needs this particular piece of information, this one this piece," to make it very efficient and very more predictable towards the desired outcome. But of course, you need to think about how you arrange data. It's easy to just throw everything into an agent like, "Hey, figure it out." And it's not that easy. The bigger complex use cases, you definitely want to go into context engineering for sure.

(Joel Beasley at 00:23:23) I can definitely agree with that just as in my personal life. I use it a lot, like quite a bit actually. We use it here at the podcast too. We've got custom projects trained on very specific things, and we have found that less is definitely more. Less is more, and then examples are awesome. So—

(Micha Kiener at 00:23:42) Absolutely.

(Joel Beasley at 00:23:42) If I have like, "Hey, I have here's five examples. I need you to output it like this." That's all you really need to say.

(Micha Kiener at 00:23:48) Yep.

(Joel Beasley at 00:23:48) You know, and then give it constraints. Consistency, I'd say it's fairly consistent. And then what we've been doing, like, for example, with the transcriptions—it would sometimes format the way that they transcribe each speaker slightly different. Sometimes it'd have braces. Sometimes it'd have brackets. Sometimes it'd be uppercase, lowercase, Speaker One Speaker. All these little different things. And so we started running it, and then we noticed this inconsistency in how it was classifying and displaying speakers. And then that was just like one line: "Speakers displayed this way." And then we have, you know, essentially a hundred percent success rate with it.

(Joel Beasley at 00:24:27) But I guess, watching it—like having a good idea, then watching it run, then making sure the guardrails are in place. Because I don't think you—tell me, Micha. Can you ever predict like a hundred percent of it? Or do you always just have to have the idea, let it run, and then kind of shape it along the way?

(Micha Kiener at 00:24:46) I would say, at least in the current state of where we are at with the AI, there's no hundred percent guarantee ever. Even for simple single tasks, where—that's where our utility agent was really designed in a way to, you know, reduce the information like prompts, input, output into a very specific defined way to help reduce the risk of returning something you don't want—but it's never a hundred percent guarantee. That's where even today, having an ability to have humans in the loop is super essential. And the thing is, I also saw attempts where companies would go for like, "Okay, I always put a human in the loop. Everything that an agent does gets reviewed."

(Micha Kiener at 00:25:36) The problem with that is you get what we call button fatigue, which means like if you present a hundred review tasks and 99.9 percent, it's okay. People just click yes, yes, okay, okay, okay, okay, okay, okay. They don't even see that 0.1 percent where it should be their attention. You know what I mean? And so the tricky piece or the challenge is to figure out when do I need a human in the loop. You don't want only that 0.1 percent. Maybe it's 3 percent because you want a bit of grace there that says, "Hey. Rather add a couple of, you know, kind of false positives instead of missing that one that was actually wrong, and I really need a human for review." That's also something that Flowable deals with, like this confidence level that we have for agents that we try to figure out like, "Okay, what's the confidence level? What's the threshold that we add a human in the loop?" Sometimes the outcome of the agent could even see like, "I don't know. I need a human in the loop."

(Micha Kiener at 00:26:50) And then in Flowable, that's pretty cool. We transform an agent task into a human task. And it's fully transparent, and it's pretty cool. And then you can design a form for it, how that user task then should look like, and things like that. Or we do it by thresholding, which is, say, we have another control agent that controls the output and defines the confidence score. The goal is to really have this very small percentage where you only want to have humans. So let's say for them, in 90 percent of the cases, it's really worthwhile looking at the result for review.

(Joel Beasley at 00:27:32) Assuming you use GPT, Grok, whatever model that you're using, Gemini, whatever it may be. And we've all seen in the past couple months the ability to like watch them think or to watch them like grab other agents and work together, like multi-agent problem solving. And then you can watch them communicate back and forth. Is anyone doing that in production today in your world with your software, or is that not happening today?

(Micha Kiener at 00:28:00) That's what we do, everybody, even in production. Absolutely. But typically, that gets orchestrated directly in our platform. So, this kind of like, what I mentioned before—a complex problem gets divided into smaller chunks, into smaller pieces. And the orchestrator agent, he might even orchestrate other agents and, you know, delegating those smaller pieces, and then results thinking like, "Okay. Now I have new information. What do I do next?" Right? And we also have some sort of really nice visualization that gives you this kind of timeline. Okay.

(Micha Kiener at 00:28:44) What agent was invoked by which one? What was the result that turned a next tool to be invoked or a service or a human in the loop? You really can visually see how a case gets solved over time. Who did what, and what was the outcome? Why did I invoke a tool because I had a missing piece, and I had a tool at hand to give me that missing piece. Right?

(Micha Kiener at 00:29:08) And I think it's at the end, what we try to do is to kind of demystify the magic in a sense, like, "Oh, we throw something at an agent and there's a result coming back." We want to know what's going on behind the scenes, and we want to control it. We want to have guardrails where needed. We want to have those reviews to humans in the loop where needed. You want to know what happens at any given time.

(Micha Kiener at 00:29:36) Also for traceability and auditability, of course. Right?

(Joel Beasley at 00:29:40) Especially if you're working in health care or banking. It's all—

(Micha Kiener at 00:29:45) Absolutely. You have to do all of that. Yep. Yep.

(Joel Beasley at 00:29:45) This—we'll wrap up on this context engineering part real quick. Are you building agents that help you create good context prompts?

(Micha Kiener at 00:29:59) That's actually something that we're truly working on. Having AI that actually help you define the prompts, context, or pieces. We used GenAI already for quite some time now in the platform that helps you, you know, define models up from process models, case models, forms, integrations. That's really cool, and we're already at a super good level. You could also say something like, "Give me or generate me a typical two-step review or approval process, and I need information A, B, C, D," and it would already generate kind of like a basic bottom line model for you.

(Micha Kiener at 00:30:46) But the next step we're currently working on is really that we use even GenAI by defining or helping you define those agents from prompts to or reviewing prompts, giving you feedback like, "Hey. This is not specific enough," or rephrasing it or whatever, up to context engineering as well. Yeah.

(Joel Beasley at 00:31:11) Yeah. I'd say maybe 10 percent of my interactions with AI are actually having it help me come up with the right prompt to use as a project or just how do I get the best answer. I go to Grok to ask it to help me create prompts for the Imagine thing because it can create better prompts than Imagine does. So—

(Micha Kiener at 00:31:32) Really? Yeah. Yeah. Yeah. That's pretty cool.

(Joel Beasley at 00:31:35) You guys have done some really cool things for your customers. It's not all just building fancy flashy technology. One of the things that caught my attention is you took a retail onboarding from five days to seven minutes. Is that right?

(Micha Kiener at 00:31:51) Yep. Yep. Tell me about that.

(Micha Kiener at 00:31:53) So that was, I think, it's not everything to AI. Right? It's also pure—we talked about at the very beginning—pure orchestration. It's around breaking up those silos. They had a lot of breaks in their data pipeline, so to speak. So when you sign up for a new account, you could even do that online. Then it went through different parts in the organization, and that took time. That takes time. It's error-prone. People would even, you know, be typing in things multiple times.

(Micha Kiener at 00:32:31) And so a big portion was just pure orchestration and automation, end-to-end. We're back to that picture, beginning. But somewhere in some places, we also added AI agents or that previously done by manual reviews, for instance, for specific, you know, types of customers or whatever—that's not clear. That could reduce the time again. So I would say there's seven minutes.

(Micha Kiener at 00:33:00) The biggest portion of that was online identity verification that's taking up the most time, but you'd be entering in seven minutes. And you could even log in to your e-banking account afterwards. That's pretty cool. So you apply for an account and you get through all that online verification and everything. And after seven minutes, you were handed the credentials to get started on even logging into e-banking, which was very cool.

(Micha Kiener at 00:33:34) Because the problem was not—it's not the pure efficiency gain or saving some time you had to spend for onboarding. It was more the challenge of losing customers along the way. If you have to wait for seven days just for opening an account, you probably just jump off and say, "Okay, I'll try the next bank."

(Joel Beasley at 00:33:59) I just did that last week, actually. I applied to two banks. One came back to me like, "I need this, this, this, this, this, and this." The other one says, "You're approved. Here's your card." I said, "Oh, oh."

(Micha Kiener at 00:34:09) Oh, thank you. Okay. So you canceled the other one, I guess.

(Joel Beasley at 00:34:12) Yeah. I just was like, "I'm not doing the other one" because I needed to solve my problem. I took two shots at it. One came back immediate solve. I can move on with my life. I was like, "I'll just go with that one."

(Micha Kiener at 00:34:21) Yep. So the churn rate is higher than you would expect nowadays, and it's always getting—people are more and more—I wouldn't even say impatient. It's kind of like state of the art for a lot of platforms. You just sign up and you get immediately after you can just go. Right? Why should it be that complex in those regulated industries like banking?

(Micha Kiener at 00:34:48) Of course, there is stuff behind the scenes that still needs to work for sure. It's a bit more complex than signing up for an Instagram account or not. But it's still—we have capabilities today to make that way quicker than it was years ago.

(Joel Beasley at 00:35:06) Yeah. I mean, it's partly—I'm an impatient person. I know that about myself. I'm very aware of it. But also, you know, there's that aspect of it, but we created a new standard on Earth.

(Joel Beasley at 00:35:17) We created the standard on Earth that you can have it right now. You can get it right now. And so anyone who's not rising to meet that new standard is going to get left behind by some startup company who's hungry, who says this shouldn't take seven days, this should take seven minutes, and they're going to go make it do that.

(Joel Beasley at 00:35:37) You know?

(Misha Keener at 00:35:37) Absolutely. Absolutely. Yep. Or another cool example was when we helped automate and streamline the invoice process at a big retailer in Europe. I think they handled a couple of 20 or 30,000 invoices per month throughout the whole chain. And it took them in average, I think, 35 days for processing, from receiving an invoice until it was paid. And so not only took that a lot of effort, they had to review a lot of information and double check with their ERP. Are we even talking about a valid supplier here? Do we have a processing order, a PO number, and things like that?

(Misha Keener at 00:36:29) And so we also used our document agents to classify the documents, extract information, and even prioritize. In Europe, I don't know if it's a thing in the US as well. We have something we call cash discounts. If you pay within, like, typical five or ten days, you get an extra 3% discount. They could not even—they didn't even implement that part of the process because they knew they were always too late. So I don't even bother. And with our solution, we brought back that thirty-five days to, I think, below one day. And we could even prioritize if we still needed some approvals or whatever. Everything was prepared. And we could prioritize in a sense like, hey, those are cash discounts. We want to look at them. And so we could additionally, beside all these savings they had in less work, we could save them like between 8 and 10 million per year just on savings in those cash discounts, just because it was fast enough for payments. That's pretty cool, I would say.

(Joel Beasley at 00:37:34) All right. Yeah. I'd say, coincidentally, my services will cost about 6 million a year, and we'll see. We'll net two. Let's do it together.

(Joel Beasley at 00:37:42) Now that's actually really cool. And so you're a Swiss-based company. Is that correct?

(Misha Keener at 00:37:48) It's Switzerland. Yeah.

(Joel Beasley at 00:37:49) Oh, very cool. And so, hey, does that help ever when dealing with international companies?

(Misha Keener at 00:37:54) It actually does. Yeah. We not too long ago, we even had this kind of Swiss army knife type of marketing thing. Like, this is our platform, like the Swiss army knife that basically solves all your orchestration issues. Yeah. Yeah. The Swissness, the quality, you know, in time and the engineering capabilities, that definitely helped on the international world.

(Joel Beasley at 00:38:22) Play to your strengths. Right?

(Misha Keener at 00:38:23) Yeah. That's true.

(Joel Beasley at 00:38:23) 100%. Yeah.

(Misha Keener at 00:38:26) I should do that.

(Joel Beasley at 00:38:27) Yeah. Your competition will convince you not to. They're like, ah, you shouldn't lean on that, you know. You should definitely play to your strengths. So look. Let's say that we get to this point where we're automating all the work, got lots of agents, very minimal need for human in the loop. What do the humans do with all of their free time?

(Misha Keener at 00:38:47) You know, working at the things that are truly essential, like taking care of their customers. I feel like if you go to probably almost any sort of business, there's not enough time to really take care about the essential, like your customers. And that goes through every domain, I would say, even in healthcare, in hospitals, nobody has time. Right? And if we can free up time with that we have to spend in those silly tasks and admin work and, you know, whatnot, you have time to really take care of customers. So I believe it's not like we are going to lose tons of jobs. I think it's going to slightly shift towards, I would say, more meaningful jobs that you can do.

(Joel Beasley at 00:39:46) Well, the jobs are definitely going to change. I mean, this is nothing new as far as history goes.

(Misha Keener at 00:39:52) It's not the first time. Right? No. It happens all the time.

(Joel Beasley at 00:39:56) You and I aren't lifting giant thousand pound bricks. Okay? We've got hydraulic machines that do that now.

(Misha Keener at 00:40:03) The same for fabrication. Right? If you look back, like, there was a super big whole industry, and nowadays they're gone, but we do it in a different way. And it's not like, oh, they left tons of people behind.

(Joel Beasley at 00:40:17) And how is Flowable doing today as far as growth? Were you—are you early startup, mid stage—where are you at?

(Misha Keener at 00:40:24) I would say we're mid. We're—I don't know how you would call that. I would not say we're a typical startup anymore because we still have something around more than 600 customers worldwide. We're doing pretty good. We have big growth, around, I don't know, 60% year over year overall, or just substantial growth. And also from a geography point of view, I think 45%, probably, of our revenue now is from North America. Although we started off in Europe. And so yeah. But we recently just were recognized by analysts, Gartner, into their new magic quadrant they created because there's kind of a merging of a lot of different technology into one platform. They call it BOTE. I don't know whether you heard of it. It's called business orchestration and automation tools. Previously, they were separate pieces, you know, like, they looked at automation, they looked at RPA, they looked at integration solution, like process orchestration and stuff like that. And they figured out, like, nowadays more converging into single platforms that can handle all those capabilities. And we were selected as one worldwide to be in that new magic quadrant. But we're by far the smallest one. There are others like, you know, Microsoft, obviously, Oracle, SAP. Those big names are in there. And we are the smallest vendor in there, but still substantial that you even made it into that piece of documentation from analysts like Gartner.

(Joel Beasley at 00:42:12) That's interesting. You guys getting acquisition opportunities all the time?

(Misha Keener at 00:42:17) We are getting contacted by a lot of people. Yeah. But currently...

(Joel Beasley at 00:42:23) If you're a fast-moving technology in a nascent market that's kind of coalescing, eventually those M&A people are going to start knocking. And they'll probably happen all at the same time too if they're doing their job right. New magic quadrant comes up, they can see the market, they're in it. Because before you mentioned the BOTE thing, that was my thought. Because I see all different types of technology all day because the podcast getting to meet all these different types of people. And that's kind of where my brain was going to garbage collect. It's sort of like compressed down later. I'd be like, okay. But you said it perfectly. Right? It's the business automation tool. It's all these different tools that are kind of coming together. And that's actually hard for some people to wrap their mind around because you silo all these tools in your head like, oh, that's the RPA provider. That's this provider. You've got the Zapiers of the world. You know, you've got all these different providers for these niche things. One provider coming together and pulling it all. That's pretty newsworthy.

(Misha Keener at 00:43:25) And you can see, like, from all those technologies where their roots are. So for instance, probably you heard of the software called UiPath, probably one of the biggest RPA vendors. They also now went into that angle and said, oh, we need to add process automation or process orchestration to our platform, but their true core is still RPA. And we also have RPA capabilities, but we added them in terms of, hey, we also need to be able to orchestrate those smaller automation pieces, those smaller robots. And you can tell if you look at a platform what is their biggest core or strength and what is—but they can basically solve all those different aspects of an end-to-end platform, but you can typically tell where they're coming from. Or others like Mendix, they're coming from low-code capabilities building applications. Probably heard from them as well. Now they also added those kind of process capabilities, but it's pretty minimal still. And so every—and that's pretty cool. So you have this kind of emerging or converging platforms, but they still have their, you know, their sweet spot, so to speak. If for you, like, RPA is still the biggest thing, you probably go for something like UiPath. If you say, hey, I want a super robust process orchestration or automation, then tools like Flowable is probably your choice. You can still do a bit of RPA. You can still do integration, of course. But it's really depending on what is the problem you want to solve.

(Joel Beasley at 00:45:11) Yeah. It's very important, you know, companies trying to pivot and add on and bolt on versus the ones who were built for the future. You can tell the difference, and that'll just increase the valuation of the company. So kudos to you, Misha.

(Misha Keener at 00:45:24) Thank you.

(Joel Beasley at 00:45:26) You skated towards the puck at the right time. Right? Like, that's the—so the things are moving fast. The LLM evolution's happening. It's almost scary in your words. That's what you've mentioned before. Tell me a little bit, like, what's changing so fast that you'd be like, this is kind of scary.

(Misha Keener at 00:45:45) I think scary in a sense, like, how fast those evolve. Things an agent probably could not do great today doesn't mean it cannot be done tomorrow or in a week's time or month time, whatever. And so it's really something that evolves so fast, and you have to keep up. And that's also another big issue for customers. Think about you build a solution, you have your agents, and they do what they should do. And you're currently running on—let's say you started off in GPT-4 Turbo type of a language model. And that all works. You have your prompts optimized, your context, everything's great. And now you want to switch to the newest one, like GPT-5 or whatever, and you feel like that should be seamless. Right? But it's not. Some tasks might be even possible nowadays, which weren't before, but some others, maybe even simple tasks, don't work anymore. You have to refine the prompts or the data input or whatever. And this really implies another, I wouldn't say an issue, but something that the platform should take over. Like, how can you ensure that it can switch to different models and still your end-to-end use case or workflow agent, whatever, is still working? So that's also something that we have in Flowable. We call that an agent test suite that allows you to define test datasets and expected outcomes, and then you can actually switch the models and see how they differ. And in very different ways, not just the result, but maybe it's the same. Let's say you—so just the other day, we had an example with a customer, and they were just using the best, the latest and the best, biggest model. And we said, hey, this is a small utility agent. You probably want to go for a mini or even a nano. That saves you time and money. Right? And then they tested it and they felt like, hey, it's working perfectly, just faster and cheaper than—I don't need the full blown model for that specific task. So you also want to have a platform that allows you to use different vendors, different models depending on the very task you're looking at. Right? So for instance, if you want to do mathematical calculations, ChatGPT is not the best one. You want to go more for something like Anthropic Claude or Grok. They are way better in those kind of mathematical calculations. And so you want to be able to—your agent can or you can define what model you want for your agent to work.

(Joel Beasley at 00:48:32) Not surprised. OpenAI is still working on their business model a little bit. We swear GPT said the math added up. We just need another couple billion. Microsoft, it'll be cool.

(Joel Beasley at 00:48:45) Let's talk a little bit about leadership lessons. So one of the things I like to ask the leaders that come on the show is for a piece of leadership advice that you received a while ago. You put it into practice, and it stayed with you until today.

(Misha Keener at 00:49:01) I've been especially going through those kind of startup phase, right, where everything matters. Every email you write, every instruction you give, everything matters and has a direct impact. I would say one thing that I always dare and is important for me is leading by example. So instead of just delegating, just do it yourself and then people follow and leading with passion, with being a front runner, so to speak. I still do that today. So that was definitely something that sticked with me for sure. I think it's also probably a bit of how I grew up with everything I did starting at age 14 and going from kind of like company or product to the next one. It kind of came a bit natural because at the very beginning, you're more or less alone. You have to do it yourself. And eventually, you get people in and you still do it. I probably did it quite naturally, I would say.

(Joel Beasley at 00:50:17) And then when you're bringing people on to your—not at the whole company as a whole, but your specific team, they're going to be a direct report for Misha. What's the most important thing that you look for?

(Misha Keener at 00:50:29) I think the mindset attitude for me is more important than the current skill set. You can work on skills. You can develop skills. But the passion you bring with, the mindset you have, how you—is it just a regular, you know, job for you? You just do your eight to five thing, or do you do things because you search for purpose, you have passion, you want to work on a great product, have an impact? That's what I'm looking for. So in skills or tools, whatever, you can work on that over time. We have a great team that is there to help, but it's hard to change attitude.

(Joel Beasley at 00:51:16) I agree. And a lot of people when they hear that, like, you know, eight to five or nine to five, whatever the working hours are, they get really fixated on the hours part. But I have found that it's more important about the energy that you bring to something, the work that you're doing. Because I have days where, like, if you look at the hours worked, it's pretty low, but they've had some of the highest impact of the quarter, you know. And it's like, if you come and you're fully charged and you're there and you want to climb that mountain and take that boat and, like, make it happen, that's the energy that I have found is the most profitable energy.

(Misha Keener at 00:51:55) Absolutely. And it can also be at—or getting into dynamic that gets you excited as a team. So, that also specifically, of course, in the early days when you had to work, I don't know, whether we worked on a specific POC or presentation or the next new feature we want to have and deliver to our customers, next release, whatever. Sometimes we really had to work long hours, night shifts, whatever. And that's also something when I talk about leading by example, I will not just tell my team, like, hey, tomorrow morning that needs to be finished and off I go.

(Micha Kiener at 00:52:31) I would help myself. Some parts of the engines are still written by me because I want to be super hands-on and still understand all the bits and bytes behind the scenes. And that can also give the whole team a lot of energy. I actually lost one of my really cool employees one day when we were more, you know, settled, more like less of those night shifts and all we need to do to get this done by tomorrow or next week, whatever. And he would leave because of that, because it was too boring for him.

(Micha Kiener at 00:53:11) It was like, hey, I can, if I want to, I can work eight to five. I need this energy. I need this kind of, you know, vibration around a weekend, a night, whatever. Sometimes we would even come together, the whole team, because we are fully distributed.

(Micha Kiener at 00:53:29) Sometimes we just gather for a week in a nice sunny place and do twenty-four seven hackathon, whatever. Right? And people get excited. And if you don't have that mindset, then you also don't feel right at the team, I would say.

(Joel Beasley at 00:53:46) I love it. Well, Micha, we did it. We made a podcast. How do you feel?

(Micha Kiener at 00:53:50) Pretty good.

(Joel Beasley at 00:53:51) Thank you so much for listening. And if you found this episode useful, please share it with a friend or colleague who you think would get value from it. And if you have topics that you would like to hear discussed on the podcast, either add me on LinkedIn or send me an email, [email protected]. Every time I get an email or LinkedIn message, it absolutely makes my day and inspires me to keep going.