The interview transcript below has been edited for length and clarity.
Sears Merritt: A set of AI engineers and developers pulled all of the COBOL copy books, and then at the same time they took screenshots of our green screens, and then loaded all of that up. We have a series of prompts that we're using to actually have the AI decompose the code, rebuild UIs, right? A, a whole bunch of things ki- running in the background.
We would end up with just about a working prototype of what we would need to rewire and reintegrate and move that workload into a modernized little web app. The team did it in seven days. It's just amazing. Wow. We're in a situation now where we can encourage everyone to use as much AI, the best in class models, right?
The most token consumptive workflows and capabilities possible, so that we can actually understand what the benefits are relative to some of the quote unquote simpler, lower cost LLMs. We're also now gaining access to analytics that let us in a very granular way look at usage patterns, developer workflows, and begin to make sense of who's using what when and for what types of tasks.
Matt Marshall: Today we're talking with Sears Merritt, who runs AI as CIO of MassMutual, a hundred and seventy-five-year-old Fortune 500 insurance company. He's deploying inside one of the most regulated legacy environments in America, and he's doing it with discipline. He's got a trust score rubric, a skepticism of single vendor lock-in, and a willingness to attack problems everyone else calls too hard.
Today he tells us how AI rewrote his build versus buy math. Welcome to Beyond the Pilot, Sears.
Sears Merritt: Thank you so much, Matt. Really looking forward to our conversation today.
Sam Witteveen: Sears, welcome to the pod. A lot of enterprises out there are signing these huge single vendor deals with lock-in. It sounds like you've deliberately, very deliberately by the sounds of it, gone the other way.
You let software developers try out a number of different tools. How do you let them actually decide what they want? And then are you actually seeing these bumps in productivity that people are talking about?
Sears Merritt: For all the tools that, that we choose, particularly for software development tools, we establish, a set of evaluation criteria effectively a rubric.
And you can effectively map the evaluation criteria down into two categories, cost and experience. And we will work with dev teams to balance th- those two dimensions and ultimately choose the tool that, that gives us the best outcome, both, both in terms of cost and developer experience.
And again, w-wi- within experience, think quality, code quality time to release, th- things of that nature. So that, that, that's how we do it. And then we'll, typically sign twelve-month agreements with our providers so that we have enough time to get the value out of the tools that, that we bring in, and then at the same time also have enough of an opportunity to switch and replace to the extent we need to do that.
Matt Marshall: Sears, our survey data at VentureBeat backs you up, right? Most enterprise companies are f- following this hybrid strategy. They're using deterministic approaches in areas that are really mission critical. They're using maybe more reasoning in, in, in LLM, less deterministic approaches for areas that are less critical, right?
Then you've got this hybrid approach. You've compared this to the early internet, right? Every- everyone wanting their own LAN before a TC/PA, I-IP one. Where are we in that cycle? What's the forcing function that eventually pushes towards standards?
Sears Merritt: Number one, to the extent you, you believe the analogy it still feels like it's pretty early, but there's some signals that I think suggest there's gonna be, again, a move to standardization and consolidation at some point.
I think a l- a lot of that is around standards development, so MCP certainly A2A and things like that where you get some of the big providers actually investing in those capabilities. And I think the market will just kinda continue to force companies to, to operate like that. I think open source actually and some of the bigger open source model providers will probably also encourage standards development.
You might even expect, if you kinda go back and look at the history of open source and how some of the bigger firms kinda got behind it, one of the reasons they did that was to neutralize kinda that monopolistic competitive factor that, that some of the bigger players like, like Microsoft had.
So you can imagine to the extent that those kinds of dynamics continue to play out in the field of AI sh- we should expect to see standards and open source-like mechanisms emerge that kinda neutralize those threats, for smaller businesses trying to compete. And then of course, companies like, like us can stand to benefit because that gives us a lot of, a lot more optionality and at the same time some minimum level of functionality and interoperability, right?
To keep the core infrastructure and kind of the core AI services running.
Sam Witteveen: Currently, what sort of, bump in productivity are you seeing from the AI tools that you're using? You mentioned some interesting stuff about actually seeing some of the software development pick up. Are you measuring that?
How, how-- what sort of is the percentage increase?
Sears Merritt: For, first of all we are seeing material improvements in productivity. For the things that we measure, SDLC we've seen a lot of numbers floating around. We've, through a lot of different ways that we've measured that recovered about a thirty percent boost in developer productivity.
And with that productivity boost, the other cool thing we see is developer experience also goes way up, right? So they're just happier to come to work and be able to use these tools and again, make a bigger impact on the dev teams that, that they're part of, modernizing infrastructure, enhancing mobile apps and and some of our digital tools that things like that.
Certainly seeing big boosts there. In things like contact center, we're seeing time to resolve a call come way down, like ten minutes down to one minute for specific call types. And similarly from a cost perspective, dollars down to cents. These are really really big im- big improvements.
And again, at the same time, we're getting those efficiencies realized, and we're also seeing experience improve. So effectively, NPS con- continues to go up. And I find that to be some of the most fascinating outcomes here. Oftentimes, when you think about efficiency. There's always a balance where you've got to set a floor or a minimum level of experience you wanna keep in place, right?
Subject to driving as much efficiency as you can. And with AI, we just haven't seen that.
Matt Marshall: Sears, congratulations. That, that, that's a great story. Over the last month or two, there's been a lot of publicity about the exploding costs of using tokens as organizations have realized the power of agents.
You heard from Uber recently that they blew through the entire budget in that they had for AI in the first three or four months because those engineers went ahead and did that. So how are you managing the cost side? You've said you brought the cost down, but how is that working over the last month or so?
Sears Merritt: Yeah. So I, on, on the one hand, you could say we were strategic. On the other hand, maybe you said may- maybe we can say we, we got lucky. In either case when we brought in all of this AI infrastructure and we evaluated, signing a consumption-based versus kind of a seat-based or effectively an unlimited license with our providers we just chose to go unlimited because the cost differential be-between the two seemed to be immaterial at the time, and we would obviously hedge out a lot of the potential cost overrun risk if we just got seat-based licenses.
So for everything that we've done today we're seat-based. That's helped us avoid some of the challenges that, that other companies might be going through. But certainly we can't avoid those forever. And over the next six, six to nine months we're gonna have to adapt pretty quickly to make sure we've got, the FinOps, so to speak, in place to manage token budgets and model calls and, all that good stuff.
Sam Witteveen: You've gotta be looking like a genius at the moment for negotiating this sort of all you can eat from a, from the different model providers, et cetera. How does that affect your thinking? D- are you, one constantly trying to push your team to be token maxing now to work out what can they actually do with a lot of tokens while you've got the sort of all you can eat buffet?
And two, are you also, on the other hand, scared or, constantly concerned about, "Hey, how much is this gonna cost us when this is actually over?"
Sears Merritt: It's funny it's yes to both. Okay. So we're in a situation now where, we can encourage everyone to use as much AI, the best in class models the most token consumptive workflows and capabilities possible so that we can actually understand what the benefits are relative to some of the quote unquote simpler, lower cost LLMs.
So we're certainly encouraging everybody to do that. And then at the same time and this is w- talk about fast moving, this is like within the, in the last week we're also now gaining access to analytics that let us in a very granular way look at usage patterns developer workflows, and begin to make sense of who's using what when and for what types of tasks.
And once we have that data set established then you can imagine what we're gonna be doing and have started to do right now, is effectively optimize those workflows and make sure that we've got the architecture and some of the core infrastructure in place to facilitate that optimization.
Particularly think, model calls and routing and, choosing prompts based on cost and experience and response times and all those good things. That's infrastructure that, that we are putting in place right now that will be built as a function of the analytics that we've been able to start to, to collect from fr- from our providers.
Got a lot of work to do on a relatively short runway, right? So we're gonna have to, re- renegotiate, think of it by end of year.
Sam Witteveen: People like Nvidia talking about, token maxing and that every, e- every engineer should be using X dollars worth of tokens and stuff like that.
My guess is you're in a position to actually see that for certain tasks or even certain people in the organization, the return on those tokens is very different than other areas of the organization or other people. Is that correct?
Sears Merritt: Absolutely. And, I guess I, I might see it maybe a little bit differently.
We're not starting with a, token max everything. That's not the mindset we have. For that, that focused AI development work where we're really trying to optimize and improve a particular process or an experience, we're always focused that's the outcome. So when we have a cross-functional team working on those types of projects, so AI engineers and developers and the like, they're all working together not to see how many tokens they can use But to see how close we can get or how far we can exceed a particular goal tied to improving the efficiency or improving the experience of a particular workflow in, in the company.
And then we work backwards from that. So you can imagine there's a trade-off. We can spend $40, every time we run that process using Opus 4.7, or we can spend $1, using some legacy GPT. We will make those choices based on co- based on the outcome and the impact we're trying to have on ultimately the people that, that are gonna benefit from what we're doing.
So that could lead in certain really important high-value, high-impact processes, we could say, "Yeah, we wanna token max the heck out of it." We're willing to spend $500 a run on that because the return is gonna be so high. Whereas we might say, "Hey, for these lower cost things we're gonna run those, bare bones and ma-ma-make it as cheap as possible because it doesn't really make a difference to the experience."
So l-long story short, staying focused on the outcome and the impact as opposed to how to maximize tools getting there is m-maybe a nuanced view that we take here.
Sam Witteveen: What are the things that are actually worth paying for the really expensive models versus w- I think people jump to conclusions about some of this stuff, but clearly you're actually seeing it and getting the results.
Sears Merritt: Yeah. So a great example actually is a lot of work that we did in our IT contact center. As we were building out an agentic solution there to effectively triage inbound calls and workflows and again, just try and speed time to resolution up, we had a choice to make. And again, we focused on outcome, so cost and experience.
During the development process, we exposed our user base to two LLMs. One of them was expensive, right? And the inference time ran multiple seconds. And the other one was a smaller model but it provided a near real-time type of experience to the user, but the quality of the results was more noisy it was lower.
And as part of all of our development we have these we use a concept called a trust score. As we're developing these things, we ask users, "What was the quality of the response? Did you like it?" And again, we have A, B, C to, d-different variants where we measure a lot of that. And in both cases, cost wasn't an issue here by the way, because as we were just seeing the results, again, we were going from minutes that people were spending talking to a person down to seconds when they were using the AI.
And the cost of either one of these things when you factor that in were negligible. But nonetheless, we exposed both of those models and both of the corresponding experiences to the users and asked them, wh- which would you prefer? And the users actually said we want the more expensive one.
We're willing to wait, but the quality difference is so high that the two extra seconds actually is worth it to us." So we factored that experience piece into the decision-making, and that led us to say on a relative basis, the costs were immaterial, so we're gonna use the more complex model.
Matt Marshall: Sears, there was this explosion of excitement around Claude in the first half of this year as a result of a lot of the developments they made.
And you've seen Anthropic struggling with that growth, and you're an example of that. You've really been using Claude. Clearly, we haven't talked much about OpenAI, but they, the OpenAI did come up with Codex, and you are still talking about potentially using them and whatever is best. You have the gateways for the models, and that was one step, right?
Is, okay, you could use any model. But now that it's gone agentic, how do you deal with that choice? And are you working is it still possible to use either one?
Sears Merritt: With respect to agents and kinda map- mapping that down to architecture we do envision, a multi-harness type of environment where different people that are either building agents or using them are gonna have different harness and environments.
What we've started to do, though, going back to kinda architecture and making sure we've got some common ways in which we can manage cost, enforce security and things of that nature, is just making choices on how those harnesses are gonna integrate back into our environment and then of course tie into particular LLMs.
So w- we've got API gateways where w- we've made choices to implement identity and access security controls, and we're gonna consolidate all of our agentic tool calls and things right on, onto tho- those types of things. And then similarly where we're gonna be using multiple LLMs and multiple harnesses, those are all gonna transit common gateways.
So think things like Amazon Bedrock and other things that, that you can leverage for proxying and also of course leverage, control plane capabilities, FinOps and the like. So that's kinda how we're, I would say, prudently building out the AI infrastructure that we're gonna need to make sure we've got, again, enough optionality and flexibility so that we can use the tools that are gonna balance cost and experience while at the same time not introducing a whole bunch of technical, complexity in the environment.
Matt Marshall: Sears, you said earlier that you were a little nervous about what you would find out when you got access to the analytics. Could you, can you go into, like what were you scared to find?
Sears Merritt: You never know what you're gonna find until you see it. And so when you read all these headlines saying, companies are blowing budgets and a quarter of the time they thought they, they had to use all these tools.
I I just didn't know what to expect. Things could have been mu-much more I would say expensive than we've found so far. But luckily for us, as we've actually got the detailed analytics and got to see user patterns and how they're using some of the models and token consumption and things of that nature.
We do anticipate spending more all things being equal, but it's not anything like double or triple what we are doing, right? These are kinda like somewhere between 20 and 30, 30% more and that's with all things being equal.
Matt Marshall: So which, when you move off the all you can eat plan, you're anticipating every, everything equal about 20 to 30% increase.
Yeah. That's great. Yeah. We saw headlines, I think it was Amazon, I think the final attribution was, of spending half a billion dollars on token usage as a result of some of these policies that your engineers are using this toward a goal. You don't necessarily have a leaderboard where you're seeing engineers token max and...
But I think there's this interesting trend in the industry where the folks on the top of the leaderboard, if there are, if there is a leaderboard, are the folks who are a little less sophisticated. They're using Opus 4.7 or now 4.8 on everything when they don't need to or when they can go with a smaller, more efficient model.
I-- so how are you dealing with that right now? Are you pr- are you praising the usage or are you actually locking in on costs?
Sears Merritt: So right now I would say we're actually just observing and continuing to encourage folks to use the models that they want to, with the caveat that we will be implementing changes, and we will be putting in infrastructure that allows us to optimize, usage and constrain or manage cost.
So we haven't made any kind of big dir-- changes in, in direction or asking developers to kinda change behavior at this point. Ask me in a week and depending on kinda how we're building out our FinOps capability and s- again, some of our kinda core, core infrastructure, that, that might start to change.
But again I do think it's fascinating when you get into the data. There, there's certainly a very heavy kinda power law-like tail to usage and who's consuming what and what types of mo-models are being used. And like you said, Matt it very much correlates to a small number of people using the most sophisticated token-consumptive LLMs is what's kinda driving, dri-driving a lot of the usage at least when you measure it through tokens.
Matt Marshall: Yeah. You said the top 10 or percent are using about 80%. Are is that- Yeah ... top 10 the boneheads that are using it inefficiently or- It,
Sears Merritt: It's a good question. Y- we actually just got that data last week, so we haven't really dug into the workflows that are sitting underneath what's actually generating the usage patterns yet.
Sam Witteveen: What can you do now that you couldn't do before? So you told us this interesting story about how you had stuff on mainframes that had been around perhaps from the 1980s, if I recall, and they just... it just wasn't worth the sort of bang for the buck to update them in the past, but suddenly that has flipped now.
Can you tell us a little bit about that?
Sears Merritt: We've had mainframe systems like many other, financial institutions. Certainly ma- many life insurers have had mainframe systems for years and that's largely due to the fact that our products la- last for decades. This is a really fascinating piece.
You, you can't actually leverage the speed of your business to have an organic transition mechanism because our policies stay in force for so long. So there has to be some intentional migration of policies off of one sy- system to, to another. Now what we've been able to demonstrate is even for these smaller applications that sit on our mainframe Using things like AI to very quickly analyze copybooks, define new architectures, implement them in a new kind of web-based standard is e-effectively free as, as far as I'm concerned.
You don't need, small armies of contractors debugging and analyzing code. What we've been able to do is basically use AI to do all of that and rapidly unlock new opportunities for us. So that's actually a big focus area over the next, call it 18 months. We- we've got an an initial core mainframe migration initiative where we're taking our big mainframe applications, so these are big policy administration systems, and consolidating those down.
There we can make the CBA work. But again, there's a lot of surrounding smaller apps and things that sit on those platforms that, that could never do that. So now we can actually move all of those things over at a relatively low cost with a huge amount of speed.
Sam Witteveen: How do you do that?
How do you actually... 'cause you're talking about really old legacy systems that-
Sears Merritt: Yeah.
Here's the workflow.
So a set of AI engineers and developers pulled all of the COBOL copybooks, and then at the same time they took screenshots of our green screens and then loaded all of that up, screenshots and code, into AI.
And we ha- we have a series of prompts that we're using to actually have the AI decompose the code, rebuild UIs, right? A, a whole bunch of things ki- running in the background. We would end up with just about a working prototype of what we would need to rewire and reintegrate and move that workload into a modernized little web app.
I- if you rewind the clock, that type of work you'd, again, you'd have probably a team from a system integrator, 15 people sitting there working 90 days to do all that analysis. The team did it in seven days. It's just amazing. Wow. It's just absolutely amazing. So once they demonstrated they could do that, then of course it, we said hey, that's a major unlock, so let's get everything and see how fast we can parallelize this process, enhance the development things where, the prototypes didn't quite work as we needed to.
But again, 80% of the work was done by AI, and that remaining 20%, we can have our experienced developers and and other folks smooth out the kinks and get the bugs sorted out.
Matt Marshall: Is this why IBM is down 27% this year? If a regulated insurer can do in days what used to take consultants months, these mainframe as a service integrators maybe didn't see it coming.
Th- there's a number of these companies that are just really struggling because you can now put together these software platforms together more easily yourself or more cost effectively.
Sears Merritt: Yeah, I think it depends on how you look at things but certainly I'm sure the market looks at some of this stuff and says, "Hey, we're gonna, we're gonna discount what we used to think were high value businesses and core capabilities given the threat that AI presents to them."
Companies that are mainframe as a service providers I think should very much start to think about how to use AI to maybe become AI as a service or maybe level up their core p- core product offerings, and at the same time help their customers actually have them speed up so that they can actually get them onto a, a revenue positive new product or platform in a way that, again, help- helps their customers modernize and get off some legacy tech.
But yeah, I definitely think, for businesses that are looking at the world statically and assuming the TAM for MIPS is what it is and they're just gonna continue to try and take their 10 or 20% I think that's a recipe for a decline.
Matt Marshall: As you're seeing these frontier labs, OpenAI and Anthropic, go for, they- they... Their business models are really driving up the the costs of what they're doing. And so they're charging a premium for right? W- whereas you, open source models from China, right? F- really focusing on efficiency, on more commoditized hardware. So you s- saw DeepSeek release a model that's almost frontier, right?
V- the V4 Pro. But that is I don't know, a th- a third of the cost. It depends on what you're doing. Of what OpenAI or Anthropic is gonna charge you. So when your contract is up i- in November, are you considering open source? You talked about, that choice earlier between O- Opus maybe 4.7 or maybe a cheaper GTP model, the difference between $4 and $1.
But what about the 10 cent model or the one cent model from China? 100%. So we're
Sears Merritt: looking at, at all all models. And again, can map down how we're evaluating these things both strategically we wanna maintain optionality. But when we do make these choices, there, there's this cost and experience trade-off that, that we balance.
And I think open source has a very big role to play in how companies like ours and ma- many others are gonna use AI. We're certainly gonna need frontier models and leading edge capabilities to, to do what today is impossible and tomorrow will be possible. But I think as the industry matures, certainly as the technology matures, the hardware matures, right?
You-- I think we just saw Nvidia say they're gonna launch, desktop chip, and so inference that's being done in the data center, again, kinda go back to internet streaming and kinda how this has all played out. You can imagine a lot of the inference is gonna be done, on device for one, one reason or another, probably most importantly for energy reasons.
But again, e- open source LLMs I definitely think are gonna play a big role. Now, whether they come from a particular country or not then I think there's ba- bigger questions and more complexity to, to kinda sift through. But I w- I would, strongly believe that there's gonna be a thriving set of open source providers, certainly in the US.
We've already seen it obviously in, in China, but I definitely think the US has room for more open source here as well.
Matt Marshall: But as a, but as a high- highly regulated company is there any is is there any reason why you wouldn't use a, Chinese model? Are there any regulatory reasons or concerns?
It doesn't appear that there, because these are controllable, because you c- you can't, they're open weights, and you can ha- have it so they're, no data is exfiltrated back to Chinese servers. Presumably these are fine for you to use, right?
Sears Merritt: Possibly. I think, you've got to ask yourself, so what data's been used to, to train these things?
What were the methodologies? There, there's a lot of, I think, evaluation criteria that need to be applied to any by the way, a- any open source type of model and any closed source model. But I think, making sure that you actually do that evaluation and you really understand where these types of things can and cannot work.
I actually think, going back to the data, y- you will, ultimately that evaluation criteria will, I think, very quickly show you where there might be data advantages and disadvantages for some of these models. And going back to, to open source, and the business models that, that need to be created there, I definitely think for financial institutions that want to pursue open source, they're gonna wanna be working with businesses that have transparent controls and transparent development methodologies and good kind of data management practices, either maybe bring your own data, you could imagine or so- something in between.
But I think that, we're still kinda, kinda watching that mature I would say.
Sam Witteveen: Just as you've had like the wins that have come from taking on the mainframe stuff and things like that, my guess is that the flip side of that is that y- there are also a lot of concerns around security that have come up now that you're starting to see the power of these models, whether they're open or proprietary models.
That's gotta be something that's on your radar as well, right?
Sears Merritt: Absolutely. We are certainly evolving our security infrastructure and control environment such that it can be applied to this new agentic world that we're living in. So paying particular attention to things like identity and access management and how we're gonna grant and delegate privileges and all of tho- those things internally.
That- that's one big bucket of work. And then certainly as things like Mythos have got- gotten people's attention and we're starting to see what these LLMs can actually do the, the threat landscape has completely changed. Maybe less so from the actual types of threats, but more so in, in the rate at which those threats can appear, compromise a company, and actually cause re- real damage to a business or to a consumer.
I think that's really what Mythos showed everybody. So we've been hard at work evolving a lot of our core security processes Less, less so really creating anything like dr- drastically new. I think when you really boil it down detect, respond, contain, right? All, all the standard cyber playbook things still hold, but the rate at which you have to be able to operate those processes in that playbook.
Some companies, might be operating those things on days, and now they've got to do it in hours. Some might have chosen to do things, monthly, and now they've got to do it weekly or daily. That's where we're at right now. And similarly for us, right? We've gone from days down to hours for some of our core cyber capabilities.
And of course you can't do that with your old processes. We've actually had to build AI solutions and build in agentic solutions into our kinda tier one and tier two cyber capabilities to be able to fight AI with AI, I guess you could say.
Sam Witteveen: Is Mythos and more specifically Project Glasswing something that's helping you in that way? Like even if you're not testing out the model are they advising you on how to think about these things?
Sears Merritt: Yeah. I think, the Glasswing and the... I would say the broader community of CISOs and other CIOs in financial services, right? Folks have really come together as they always do by the way to share best practices and talk about implications and how folks are thinking about either approaching or solving some of these new threats with durable solutions.
Certainly we've spent some time talking to cyber leadership at Anthropic about what Glasswing is and how they're using it to systematically make sure that we're gonna reduce cyber risk in a hand- handful of different ways. So we, we've certainly taken that and kinda fed it into our existing cyber strategy to figure out what things we might need to speed up, reprioritize, do ourself, versus anticipate working with partners on.
I think in the long run, we should almost expect to see software probably get more secure as the LLMs and the cyber capabilities that, that come with these models actually get embedded into the development life cycle. So you're actually finding these really critical or even handfuls of low critical bugs, but when you chain them together, turn into something pr-pretty nasty.
Those are actually getting discovered before they ever even reach a release. So I think in the long run, we're gonna see more software that, that's more secure for some of the big providers that have modernized their development practices. And so if you're a consumer Of technology, right? If you've got a, big SaaS shop, in the long run you're gonna be a a benefactor of that.
I think if you're writing your own software, what you very quickly need to do is make sure you're embedding AI into your SDLC, not only to speed up, developing code, but to also make sure it's more secure. So that, that's actually been in, c- call it about kinda two and a half weeks we've actually spent a lot of time developing that capability ourself.
So we're gonna be launching that into production think of it over probably the next 14 days, where we'll be able to use a Mythos or, choose your next Mythos-like LLM, whether it's OpenAI's or somebody else's. We'll be able to pull those LLMs in and bake it into, our cyber scanning and other kinda AI-driven security capabilities.
Matt Marshall: So wrapping up, big picture, you've lived through the internet standardization once already. H- how long until the agentic world kinda shakes out from this sprawl of options into real standards? And what should builders, your, your peers do in the mean- meanwhile?
Sears Merritt: It's hard to know when. Certainly it's started. We've got things like MCP. You've got the beginnings of A2A and these other things. So I just think staying up to date on those standards and watching how they're evolving and making sure that on your own infrastructure you've got the important components and pieces in place to, to leverage those standards.
And again they're gonna change as we go. New ones will certainly appear. I think just monitoring that trend and building with that trend in mind is probably the most important thing a company can do right now. And where there's uncertainty, layers of indirection always help right?
They give you ton- tons of optionality and a lot of lot of flexibility with, I would say, kinda minimal additional complexity. So where you don't have a clear answer, I think layers of indirection can help you kinda get through that.
Matt Marshall: Great, Sears. Thank you very much for joining us.