Mission-Critical AI Moves From Experiments to Business Workflows

Hey guys, thanks for the thrill. We're here with Thomas Robinson, AKA T. Rob, who's the newly appointed CEO for Domino Data Lab, and we're going to have a little chat about what his plans are for this company.

T. Rob, welcome to the show. Thank you so much for having me, Mike.

All right. Walk us through exactly how you view Domino these days, and where does it fit in the ecosystem of things, and ultimately, what attracted you to the company? Absolutely.

So first I'll say the most important thing to know about Domino is we are the best place to run mission-critical AI in the enterprise. We have a decade building a platform that allows people to run very sophisticated models. And what's happened in the past couple of years with the rise of generative AI is it's been able to make fantastic use of our platform to enhance further what people can do.

And really there are two main ways that that happens. The first is coding assistance or the killer use case, and that's really, really changed how people who used to be simply data scientists, simply building models, are now able to actually go build enterprise software as long as they have the right platform to do so. And it's also really, really changed organizations' ability to more rapidly develop and replace legacy systems, if you will, with much more fit-for-purpose software.

And so we're really guiding the company towards supporting that future where basically every organization is building a custom software stack for their mission-critical workflows that has modeling and AI embedded within it. So what are your plans for the company as you look at where you are today? Is there anything on your mind or anything that you see as an opportunity?

Because, well, I think a lot of organizations are still struggling with the whole data side of this AI equation. Yeah, certainly they're struggling on the data side, but I think where we see the most important thing going on is there's still a last-mile problem in actually applying AI into the business. And I think early on in the generative AI movement, there was a lot of hope that, I think hope by executives, I would say, that generative AI could rip and replace a lot of work that their teams were doing and really bring down costs fantastically.

I say that with intention, that it was a little fantastical to think that. But what we've seen is that the real trend that we're zeroing in on is the cost of coding has basically gone to zero. What that's done, though, is made the cost of speccing, the cost of understanding the business problem, the cost of understanding integration and governance has ballooned as part of building these AI systems.

And so what our company is focused on is product that allows for that appropriate governance and deployment of the technology and the people, the talent on our side to help organizations go integrate it. And I think you'll see that theme broadly across the industry as AI is moving from a desktop tool, essentially, end user compute really, to something that's embedded in the core workflows of an organization. There is still a need for judgment and people who know how to actually implement those technologies.

Is part of that whole issue is that I need to bring AI to where the data is, and in a lot of instances, organizations have data that's been strewn all over the place. So do I need to centralize the management of my data in some way or fashion before I really get this whole AI and enterprise scale movement going? I don't think so.

I'm fond of saying, as somebody who's run data projects at other organizations, I'm fond of calling out the sort of perennial three or five-year cycle that every organization is in, bringing or moving their data, or finally going to solve the data problem. And I think the reality is just give up on that. It's never going to happen.

We're always going to have data in disparate systems. And now I think a lot of organizations are actually seeing a new theme that's bringing more concern to the organization is the whole sovereign AI topic. Which is I actually want to isolate my data mainly from potentially future competitors or some of the large model providers, but also within my organization.

I want to do a little bit of siloing might actually be a good thing when AI tools have the potential to be a little more run amok and leaky. So I think if anything, there might be, boy dare I say, a little bit of a move away from some massive consolidation and a little more siloing and isolation of critical data. There's two different things.

There's a big difference between the data required to go train a custom generative AI model. That's something that requires a large corpus of information. That's one dimension, but on an operational basis and a day-to-day basis, once you have those models built into systems, the data isolation is not a big concern.

So what makes Domino the best platform, in your words, to go build an AI application per se when there are so many choices out there? What ultimately differentiates you? So I'd say there are two main things.

One is we've integrated the end-to-end. So from data and development, literally says it in our product. It's data development, deploy, and govern.

That entire life cycle is a core part of our product, and there are a ton of benefits from pushing apps to production with the same environment and setup and work that you've done to develop those applications and models. That's often counter to what's historically happened in large enterprises, where maybe there was a business team or data science team that would build something and then throw it over the wall to IT. It would be recoded.

Ton of waste in doing that, and a ton of problems with reproducibility and actually making sure the software works. So basically building that really streamlined end-to-end process to go from idea to production. The second thing that really differentiates us is governance, and we think this is critically important.

And I'm not talking about data governance. I'm talking about governance of the entire process from what are you building? How are you building it?

Think about governance running alongside the project set up and the development process through into production, including monitoring and tracing, sort of as the second level. That's critically important. And then strong believers always in human in the loop.

We have prototypes in the product for allowing people to keep humans in the loop for reviewing decisions model makes and eventually improving those models over time. So two things, streamlined end-to-end interface that speeds delivery, and then governance that makes delivery safe. Do you think we will maybe revisit how those IT teams are structured in the age of AI because of these integration requirements?

Because for so long now we've had all these silos and different folks who are data specialists and compute specialists and DevOps teams, and is it time to rethink all that? I hope so. I'm praying for a little shadow IT maybe, but I think this is the mindset shift I would love for CIOs to have is everybody's always thought of shadow IT as bad.

Okay. It's bad for some reasons because you have risk, you have controls you need to- Be respectful of. But I think there's a lot of ego around the budget and the building of things that maybe the IT team would want or desire to do.

So I think if you can have the sort of freedom to let go of that idea and build a model that gets the best of both worlds. I think the prototype for it is forward deployed engineering. That's something we're doing at Domino, and I think that's the right model for CIOs to adapt, build the forward deployed engineering team in your group that instead of sitting in IT and being walled off doing mega projects, is embedded in the business.

So running on great rails, having the right sort of platform in order to have that safety still, but embedded with the business, actually building solutions. But I think generative AI makes that possible. The idea that many more folks can build a lot more rapidly and often business users can build alongside technologists.

So I think there's a really opportune moment for CIOs to rethink the structure away from sort of the mega project, the mega team, and more into embedded teams in the organization. Of course, everybody's talking about AI agents, and especially the rogue ones. How do I think about this going forward?

" And so does that mean the controls need to be somewhere in the data platform, or is that where I need to kind of focus? Because otherwise, man, I'm not going to have any control over this and it'll just be chaos. So I have two main reads on what is going on here with agents.

I think the first one is, this is the hammer to nail problem in that everybody is wanting to use generative AI as one type of AI to solve most every problem. And so agents are kind of an extension of that. Let me go allow the generative AI to write and run its own code and act independently.

And frankly, our belief at Domino is that's not the right way to build AI systems that can deal with mission-critical workflows where there's real risk. Our belief is you need an all of AI approach. You need the entire toolbox, not just the hammer.

And so we think that simply putting guardrails around agents is not going to be enough. We think that there's an integration of traditional machine learning models, especially models, computer vision or simulation, statistical models, rule-based systems. My God, let's not forget rule-based systems.

Putting those things together where the particular decisions are maybe made by a model that is proven to have the right sort of risk tolerance rather than something that can veer in a random direction. So we see the future being a blend of those traditional models with generative AI together in systems built for more critical decision-making. Where do the security teams come in, in this conversation?

Is that an extension of the governance motion now, or how does that kind of fit? I don't want to put too much in the governance umbrella, but I certainly think security is part of it. With our governance product in Domino, we've tried to serve all folks, and those can be the cyber team, that can be an IT team, that can be somebody like a product manager who's responsible for making sure things are built in a good way, maybe a business analyst, and can also be for really deep governance and risk experts.

Think model risk management or somebody responsible for model validation in a financial services company. And so, yes, absolutely, there's a place for cybersecurity in this. And what's going on in the industry, and I think you spoke about this recently is, gosh, there's a massive increase in vulnerabilities and the attack surface area given the new models, right?

That are able to basically just troll the internet looking for more and more vulnerabilities. And so our approach on cyber is to build in all the tools to the development process, doing the analysis of the packages used, build materials, look at the static analysis of the code, et cetera. So yes, we think there's an important place for cybersecurity in the future.

We also seem to be trying to manage data at some level of unprecedented scale, and we weren't very good at managing data before AI. So what's going to change here in terms of how we think about data management, and is there some way to make this kind of, put it at the front end of our brain versus something that we kind of have historically saluted while we drove by it? Yeah.

I think it goes back to what I said before, really identifying the difference. " And basically, I think all the storage teams tried to get all the historic data online to go do some mythical training process, and I think that's an interesting moment. I don't necessarily think that's right.

I think we've seen the model's capability accelerate ahead of the need to have that much bespoke data in the enterprise. And so where I think we're probably going is that split between some modality for training, which is probably something that happens periodically, and then much tighter operational data stores for the particular workflows or business process that is had. But yes, there's no question that I think data volumes are going to scale and already everybody is grappling with the move from, of course, tabular data to many more modalities of data, documents, images, et cetera, videos.

Yeah, so more storage, always. A lot of folks are also talking about the cost of AI, and part of that seems to be, or at least the solution to it, is to rely more on the context in the data platform and maybe even throw some skills in there so that I can reduce the overall overhead of the AI agent in terms of what it needs to run in memory. Is that where we're headed with this, and there'll be knowledge graphs and all this other stuff that will come together in a way that will enable us to kind of run AI at scale without breaking the bank?

Yeah, I think you're absolutely right. " Or I had a customer tell me they had a new hire start in a finance organization and use Claude to go build a web scraper to pull market data. " Yes, this is the thing where there are very specific tasks that need to be offloaded to skills, and one of the things that we've built in Domino is all of our important enterprise primitives are built in the platform and accessible through skills, to the coding assistant.

So you sort of get this ability to offload hardened services while still being able to code against that for custom applications. But I think the other thing that's going on with ROI in AI is, gosh, we've been really focused on individual productivity, and I think there's this great pivot to organizations going into the mission-critical processes. At the end of the day, if you're focused on adoption by individuals who are doing things in the scope of their individual job, the productivity increase isn't that high.

And when foundation frontier model companies are charging based on token, the cost goes through the roof, and the ROI goes through the floor. What we need to see is the rotation, call it the great AI rotation, into critical business processes, and that requires a lot more care. Those decisions are much more high consequence.

There are fewer experts, and so the stakes are a lot higher. But that's where the real value comes from, and the cost is massively offset by the multiplicative impact of those decisions being made by AI in your business. I can't help but feel like the ROI conversation is going to be difficult because, let's say company A comes up with some brand new AI workflow, rival B will copy said workflow and create it themselves and deliver that same capability within days and maybe a week or two, depending on how fast it is.

So what's going to be defensible in the age of AI? Or is it just going to be basically the new table stakes and if you're going to be relevant, you've got to have AI? Yeah.

A rising tide lifts all ships. And so in some way, I think you're correct that more distribution and diffusion of those techniques between companies is going to be good for everybody. At the same time, I still strongly believe there are two moats or two things that keep companies competitive.

First is their data and their data sovereignty, right? In some ways, especially when you get to companies that are using AI for innovation and new product development, think about pharmaceutical organizations, that data is critically important and is a big differentiator, so sovereignty continues to be important. The second piece, though, is I just think there's a lot of challenge in the last mile, and organizations have thought very carefully about how they build their business process.

And so if you take, say, two mortgage companies, the idea that you could diffuse an innovation in one to an innovation in the other, maybe. But they have likely unique markets, unique customers, unique business process that makes them competitive, and that just sort of peanut butter spreading the AI from one company to the other isn't necessarily going to drive a lot of improvement. So I think the two things are certainly the sovereignty and certainly the bespoke nature of the business processes are still differentiators.

" Boy. I go back to the-- I think there's a couple things, and some of them I've highlighted in this conversation. One is what drives me nuts, and I say this time and time again, is when CIOs measure AI adoption, and that's the check mark of success.

That is not a metric for success. That is a metric for cost. " And of course, they're thinking in a cost mindset, so it's very tangible.

It's a terrible metric for the outcomes for your business, right? It's like saying how many people in an organization have access to the internet or have a computer or whatever. That may be anachronistic, right?

Of course, it's 100%. But if you think about that through those evolutions of business, was that really the thing that made the change? No, that has nothing to do with the underlying business.

And so thinking about it in the context of this is a technology we are giving people, rather than starting from business first, what are the three or four core metrics that drive this business unit or overall company, and then working backward in terms of how AI can impact that. So I think the AI adoption metric is the number one thing that drives me nuts, and I think it's terrible. I think the shadow IT piece that we talked about with forward deployed engineering, the idea that everything still needs to be a massive mega scale project, it's just not the future.

There's going to be so much more customization. I think in a world of generative AI, individual business users deserve something that's more fit for purpose. I think that's achievable, and so I'd like to see our technology stacks and our technology build approach orient around that, too.

I think that one irks me a little less, but I think that's kind of the future of where we're going. All right, folks. You heard it here.

We may still be in the middle of this AI frenzy, but it's pretty clear it's time to have a more mature business conversation about it. Hey, T. Rob, thanks for being on the show.

Thanks, Mike. Appreciate it. And back to you guys in the studio.