Why Legacy Systems Are the Hidden Constraint on AI Agent Adoption

ai Leadership Insights series. I'm your host, Mike Bazar. Today we're with Jeremiah Stone as the CTO for SnapLogic, and we're having a little chat about AI and legacy systems and how we kind of stitch all this stuff together.

Jeremiah, welcome to the show. Thanks for having me on, Mike. Uh, it's a really fascinating topic, and we're working with many companies right now that are really grappling with, you know, this fundamental question of how do they introduce AI into their existing landscapes and accelerate their businesses and do so in a sort of cost, you know, responsible way, uh, and get the speed they want without having to throw everything away they've been working on for a long time.

So it's a really interesting topic. Well, that is the nut of the thing, right? So if I look back at everybody's environments, we have applications in some cases that have been running for decades.

We're not gonna throw those away overnight, and that's where all the data is. And sometimes that data's on premise and sometimes it may be in the cloud. And then we have all these great new AI tools and models, and most of that stuff probably is running in the cloud, and we're trying to mash all this together.

So, um, how do I think about this? Or is there maybe a set of best practices that you've seen people adopt to help navigate this stuff? Well, I think we are starting to, uh, see repeatable patterns emerge, and that's, that's, you know, often what we look for in these times of rapid technological progress where we definitely have a high intuition and then start to see proof points of, of technology starting to, um, really have an impact and then look for patterns to be able to put that to work that doesn't take, you know, another decade to, to get to work.

And I think one of the really interesting, um, things that I'm starting to see, you know, relative to, to the companies we're working with is a reevaluation and a different approach to thinking about technology strategy with existing landscapes and trying to isolate the systems that have enormous amounts of value that can loop forward. And I, and then systems that need to be refactored need to be replaced. And the term that, um, we're using quite often these days, and I think it really kind of came out of financial services, is starting to distinguish between legacy systems that are gonna need to be put on a timeline for retirement, refactoring renewal, and heritage systems that need to have a, an architectural approach that preserves potentially decades of, you know, continuous investment, business logic, you know, deep organizational, um, information that, that, you know, is going to carry domain expertise and a high degree of value, but do not need to be refactored and thrown away and, and starting to take a more nuanced approach to understanding the, the existing landscape.

And then an architecture that allows us to deliver rapid, uh, value and new capabilities, you know, for, for the business without getting, you know, caught up in business value neutral, you know, sort of shiny object activities, hacker news architecture, I think somebody's called that, uh, you know, that that, uh, you know, is just the implementation of something because it's contemporary, but not necessarily because it's required. Is this, when you peel it all back, you know, just a huge exercise in business process re-engineering that, you know, we've been trying to get at for decades now, but in the age, AI may be easier to accomplish, but we're now doing it at a level of scale and sometimes in production environments we're just moving ahead and seeing what happens. But, you know, actually it feels like, like deja vu for me.

I I came into the industry in the late nineties, early two thousands when we were having a very large scale move to, um, you know, local gooey interfaces, three tier architectures moving away from mainframes and saw the rise of, um, a completely different penetration into the workforce with, you know, these systems and were able to take, you know, basically knowledge work really rose at that period of time, and you were able to, you know, give people personal, you know, computers, but techno technology on the desktop and expand what used to be, you know, just sort of a terminal operation. Now you, you are giving technology to every employee, and I think we're in another one of these sort of expansions of high value, high interaction, um, technology being provided in different ways now via conversational interfaces primarily, but also this idea of digital labor where we have this entirely new domain of work that can be automated and it can be automated because predominantly language models are able to handle unstructured semi-structured data in addition to what we do well already, which is highly structured data. And so, you know, I think that that is, is, is what is a redefinition of how do we get to those same outcomes?

How do we support new improvements to business processes, taking out cycle time, taking out cost, delivering more value, but apply these technologies in a way that allows to support that. So yeah, I think that I keep thinking back to, you know, business process for engineering, business process automation, that we were traditionally looking only at transactional workflows where we could use a, you know, interface that was delivered to the employee to be able to go through checking a purchase order, onboarding an employee, you know, doing other elements of the business. But now we're doing that in a different way to the end user, but also we're able to automate things that we couldn't automate previously without.

And we tried with expert systems without proved to be far too heavy a lift, uh, to be able to do. And now we can do with these native systems now the element of how quickly we can do it and how heavily it's happening at scale, I do think that there is a, um, difference between, you know, what is being rushed out into production very, very quickly is every, you know, what some people refer to as everyday ai, the ability to include these, these tools in productivity tools in your document editing system, in your email, in your spreadsheets, et cetera, versus a core business process that is not, you know, direct knowledge work, but it's actually something that would've only been in a transactional system, seeing a ton of the everyday AI incorporating these systems. You know, all of us, you know, over Christmas, everybody's got, you know, I don't know, you know, uh, you know, Claude, um, you know, running now, um, their desktops doing knowledge work, that sort of thing.

Um, I've got a couple of windows open on my desktop for, for stuff that I need to do. Um, but what I'm really starting to see with businesses is now, okay, how do we take these things and now go after those core workflows, maybe create new products, different user experience, different customer experiences, but that has been happening more slowly than the introduction to individual employees. And now we're starting to see these larger, more scalable architectures, predominantly, I think two things.

One is orchestrating across transactional systems using agents, um, as, as did the labor, seeing a lot of that in financial services with processes like loan underwriting, um, other traditionally heavy document based personnel based processes, um, but then also better ad hoc analysis systems where we're using often a, a conversational interface like a Slack or teams or, or even a, you know, self-developed ui. But for doing things that would've typically been a dashboard, now we're able to use MCP across existing, um, systems to be able to answer a business leader's question or a customer support person's question, et cetera, to be able to increase, you know, customer support readiness and customer intimacy, people would say by decreasing, again, cycle times to useful information to take action more effectively. To your point, I think this is maybe occurring in phases.

Um, we are seeing employees being more productive because they're using AI tools, but they're also having to spend some additional time maybe cleaning up after those AI tools, what some people occasionally refer to as AI work slot. But you know, on the whole, they're better off. But when I look at the workflows that you're describing with AI agents, I think part of the challenge is, is that a lot of the workflows are, shall we say, deterministic, and they're supposed to be done the same way 100% of the time.

And we have a lot of AI tools that are probabilistic being added to the mix. And well, many of them don't do the same thing the same way twice ever. Um, and I think people are struggling with how to kind of meld those two or figure out, you know, exactly where do I kinda overlay a probabilistic tool to a deterministic process and well, how deterministic Can I make that ai?

Well, yeah, I think it's a very important distinction, and actually I'm not seeing a lot of value in yield going back at well understood deterministic processes and introducing AI into them other than cross deterministic process orchestration. Uh, what I am seeing a ton of is existing human processes, which were probabilistic already now getting the value of, um, decreased cycle time better, basically raising the level of an individual employee. I'll tell you, for example, in our business, um, one of the things that we really prize is great relationships with our, our customers because we're operating at the infrastructure layer, and it's just high, high, high contextual complexity.

In principle middleware is not that complicated. You're connecting sy, you know, a system to system application, or you're taking data from applications and putting them in in databases. The technology in the middle isn't complicated.

What's incredibly complicated is the context around it. Every customer has different systems, they have different workflows, they have different data extensions they've made, um, et cetera. So for our staff, maintaining a level of deep individual customer understanding and getting that understanding before you have a conversation with the customer is pretty onerous.

It means going back and reading all the information about the customer, reading all the customer records, um, those sorts of things. And when, when we have turnover or a customer has turnover, you have to reestablish those relationships and reestablish an understanding of the customer. We've developed a system we call Sonar, uh, which, uh, reads across 1520 different, um, you know, areas of data and then creates automated, um, updates to a customer service representative or to myself as a sponsor of a customer on a monthly, weekly, or even an event driven basis.

It just increases the ambient knowledge of what's happening in the business. That was already a non-deterministic Q and centric process. We were already trying to do it, we've been working forever.

There's a core value of the company, but now we've been able to add, add ai and basically take the toil of going and reading Salesforce, reading Zendesk, reading our other information, and having that abstractly summarized and provided up to us, as has been completely transformative in terms of taking these lossy, um, processes out of, you know, individual burden of, of work and then putting it into an automate system. That wasn't a transactional process, it wasn't something that was well handled in aggregate across the different systems, but now we're able to get, get value there. On the other hand, have we adequately changed the way we onboard a new vendor?

Not yet. But even within that transactional process, there are things like credit checks, you know, risk, risk, uh, scoring or, you know, other types of things that typically would've involved a person in our finance department to look at these things. We can now, you know, know, establish, um, you know, individual workflows that again, this is a probabilistic judgment on whether the vendor, um, meets these particular criteria that a human would've had to do.

We can now automate those and have evaluations within those individual workflows that give us a level of confidence that they were performing well. Do I think you're gonna do your gross to net or your general ledger operations or your payroll using arrange model? No, and I just don't think there's a lot of value there.

What's challenging the industry, however, is that for the last 30 years, we've been stuck in these common business processes that became best practice in the last BPR wave. And it's almost like we as professionals have to go back and reread the books from the late nineties and early two thousands when we created these processes and apply that same mindset to an entirely new domain of opportunity. Um, I also think that a lot of business leaders thought that if they invested in ai, that they were gonna see some massive ROI and there'd be some competitive advantage, but as we get into it, I can't help but start to look around a little bit and figure out that, well, a lot of this stuff is just the new table stakes, right?

Everybody's trying to automate the same processes. And so I have to distinguish between an AI investment that I'm gonna make just to stay competitive, and the one that I might think is actually gonna, you know, give me some sort of sustainable advantage. But I'm not sure everybody has a firm handle on what those are.

I think you're right. I, I, I think the table stakes are the efficiency outcomes. So where we can gain more efficiency in processes, where we can run leaner, where we can, um, essentially instead of looking for a labor cost arbitrage, I think that was table stakes previously.

Are you offshoring, are you nearshoring for certain non-core portions of the business? Heard somebody refer to it as digital shoring? I, I don't know if that works, but, you know, automating, using, um, you know, essentially a language model in a loop, an agent to, to be able to do these things, I think that's table stakes.

I think that is where, um, it, it's just expected in business to be competitive. What I think we really haven't seen a lot of great examples of are new types of services, new types of things that actually increase the top line, um, that, that improve the business through new products, new services, different ways of, of, you know, growing the business or creating new value. And I think that's still an area where, you know, we're, we're working harder.

I do think that there are examples of that underway and there are, you know, folks working on this, but, you know, think back to, you know, how I would guess, you know, predictive analytics really had massive changes. And you think about, I dunno, originally when, when Uber was just a good idea, it was, could we locate the car, the black car originally before there were ever, you know, individuals involved? Could you predict where the ride demand would be, create an entirely new industry and new new way of doing things?

I think that that moment is still to come, you know, for language model-based systems where we see things and the where it's really happening right now is in the software development space. I think that's the first really interesting product market fit, where you have fundamentally different ways of doing things where you look at spec driven development tools like AWS hero or, or other, um, things that come out, you know, what, what cloud code is doing right now. We're definitely seeing, you know, an emergence there, but you know, more broadly in the economy, I think it's still early days.

Mm-hmm. Um, so as we kind of noodle this all for a minute, um, do you think we will start to see the flattening of organizations because as workflows change and maybe my sales and marketing motion is tighter, and maybe my salespeople are creating their own collateral instead of waiting for the marketing department, but you know, that's just one example, but, um, today the way organizations are structured is around, you know, some time honored things, right? There's sales, marketing, manufacturing, whatever.

Um, will we flatten that in the age of ai, Social norms changing is, is hard, right? Those, those things are pretty durable. Um, I I do think that we're continuing on an arc that we've been on for quite some time, which is more empowerment of the individual.

And, and I do think that, you know, certainly these tools put more capabilities in the talented, motivated, creative individual to not have dependencies on the product marketing team to be able to help build, build a collateral deck or something like that. On the other hand, um, it is, you mentioned AI slop earlier, it very much is, um, an age where, you know, the more seasoned higher expertise individuals are getting a lot more out of these tools then individuals that don't have the domain expertise or, or capability to, to utilize them. So, you know, I I think in some ways I, i one wondering if very talented individuals, the great deal of expertise that may have, you know, been put in managerial positions, that there would be almost a return to individual contribution.

And yes, flattening in that sense that we'll have more player coach type, um, organizations where because you've managed to take a lot of the, um, administrative overhead in some cases out, you're able to then take these, you know, more seasoned individuals and put them into the, the frontline. You know, go back to my, my projects owner example here, here at SnapLogic. I definitely think our customer, um, success leaders are more client facing now because they're doing less, um, administrative stuff in the backend.

Um, does, does that ultimately lead to, you know, less hierarchy and, you know, leaner, tighter organizations? I think it certainly could, and I think that's what we've seen traditionally is you gain more and more productivity, you have less and less need for, um, just large employee populations, and that requires less hierarchy and less, um, administrative control. On the other hand, um, I don't think there's ever been a greater need for good leadership and, and leadership and, and guidance when you have, um, such a fast dynamic market.

You know, the, the burden and pressure on consistent direction, clarity of goals and, and execution is, is I think more profoundly felt than ever. I I think that's what we see continuously is now we're having to improve our processes around corrective action plans, around, you know, program management, those sorts of things. Because following through and ensuring that the outcome is achieved, um, these tools don't fix that, you know, element of, of the process.

And that's to a large extent, the responsibility of management of leadership is to ensure longer arc, um, delivery and, and outcomes. Mm-hmm. So what is that one thing you see people doing these days that just makes you shake your head a little bit and go, folks, I think we need to be a little bit smarter than that.

Um, it's, it's hard to pick. Uh, I think there's, there's a lot of that going on. I, I do think that the, um, widespread, oh, next year knowledge work will be obsolete, will have, you know, a large, um, technical unemployment.

Um, I, I shake my head at that. I'm just, I'm just not seeing it. Um, in, in fact, what I'm seeing is almost a, you know, jevons paradox that the more productive these tools get, the more demand there will be for people to actually, you know, develop these systems.

And I think that, um, what people are not doing is really seeking to invest in upskilling and, you know, deeper education for, for how to use these tools and doing better workforce planning and management. So I think, and now is the time to be really recruiting and hiring new grads. And I think that there's been a lot of stories where, um, you know, companies are deferring bringing new graduates in into the workforce.

Um, and I think that's gonna hurt, um, companies because as we see this explosion of creativity, development and growth, um, that you're going to want to continue to build out the ability for people to, to develop and, and deliver these systems. And so even though we have individual productivity increasing, we also have this huge expansion of the domain of things we can now go after and, and improve with these technologies. And I think the thing that shakes my head is the, um, I guess the, the pause, um, on investing in talent and, and coaching and growing, um, the, the, you know, early and mid-career, um, individuals.

And I think that's more important than ever, um, given these tools. All right, folks, you heard it here. Hey, maybe this one play after all is to invest in the people first and the AI will follow.

Jeremiah, thanks being on the show. Thank you very much, Mike. Been a fun conversation.

All right. And thank you all for watching the latest episode of the Text Drawing AI Leadership Insight series. You can find this and others on our website.

We invite you to check all those out. Until then, we'll see you next.