Orion Severhill, VP of Digital Channels & Strategy at Mission Fed Credit Union, has a rule: Junior engineers don't get access to AI coding tools until they can think through architecture on their own. In this episode, he explains why he has that policy and what it means for regulated institutions trying to move fast without getting it wrong. His thesis is simple. The AI can only give you back what you give it.
Q2 Code: Practical AI for Financial Institutions
"Burning Chrome" by William Gibson
We're going to get with someone who's using AI on a daily basis and down where it meets the customer in their business at their financial institution. Great talk with Orion Severhill of Mission Fed Credit Union. And he says that he does not let his junior engineers have access to AI because he wants them to develop that first set of skills without having to rely on the kind of cargo cult mentality of allowing the agent to do the work for you.
It's an interesting thought and I think there's some real validity to it. Join us on Cut to Context where we talk through what some of the ramifications are and some of the broader thoughts we've got about AI and the financial industry for community financial institutions.
Adam Blue
Hey everyone, welcome to Cut to Context. I'm here with Orion Severhill, VP of Digital Banking at Mission Fed down in San Diego. Orion is a Q2 customer and has done a tremendous amount of great work on the platform. And he has a long and storied career in product management, project management, a little bit of everything that comes to banking and technology. So really happy to have you on today, Orion. Thanks for joining.
Orion Severhill
I'm excited to be here. Every conversation with you is a highlight of that day, so I can't imagine what this is going to bring.
Adam Blue
Yeah, that's very kind. Yeah, it'll go where it goes. So I came down and visited last year after CONNECT. San Diego, a fantastic city, and I was struck because—and this is something you told me I think while I was there is San Diego—is the biggest border city in America, which is really interesting. And you guys have such a fantastic member base, and it's got such specific and interesting challenges.
And so why don't we kick off and just get into what's it been like to try and use AI to pursue the mission of Mission Fed in the context of where you are in the community and the specific needs for your community?
Orion Severhill
Yeah, I think when it comes to AI, that's, man, that's a tough topic. It's so broad, right? Back of office for doing things, front of office for doing things. I'm excited by the outcomes, so I guess that's really more front of office. And we're just getting into the swing of things with, you know, like you said, there's a conversion, got to stabilize, normalize, and then see what we can offer.
We have a few things in the works that I can't talk about just yet because then it would seem like I'm promising something. And if it's not going to work for members, I don't want to have them thinking it's going to happen. But you know, when it comes to AI and it comes to impacting members' lives, I think we have a real opportunity to move the needle for people who …
Being the largest border city, there's a lot of underserved, underbanked people from all demographics. Because people are coming in, moving, and our border isn't just with Mexico. It's kind of … we are a military community. We have a lot of people who come in for colleges. Like the borders are on all four sides. And so we have to think through what's going to impact every type of demographic. And I think AI gives us that chance for the first time in never. Like it's never been that way before.
Adam Blue
Yeah, yeah, that's great. I think that's an interesting way to think of it. Some of your borders are kind of physical and some of them are a little more metaphorical. You know, San Diego, big convention city as well. And so you've got a lot of people that are just there for a little while. And you know, in some sense they're an important part of the community as well. And so that's an interesting thing to think about.
So, you know, I'm sure you're aware of this, but banking in the United States is reasonably well regulated. And I think it's been interesting to look at the intersection of AI and a regulated industry. And the default kind of conclusion would be, I don't know, banks have to go slower than everybody else. And you might say, well, smaller institutions, community FIs, maybe you're at the tail end. I'm not always finding that to be the case, and I suspect you're not either.
So I'd love to hear how you guys think about being in a regulated industry. And then what you've done to take your size and where you are in your market and make that a leverageable point and not an obstacle in kind of looking at the intersection between regulation and compliance and the use of AI.
Orion Severhill
Regulation, regulation, regulation. It's always top of mind, but it does not mean slow to me at all. You know, it means the cost of being wrong lands on someone who couldn't choose the risk. And that means that we have to think of any solution that we bring from multiple angles, not just “this would be cool, this would be awesome.” Think of how it would impact this person or that cohort. We have to think it through from the entire lens.
I think once you've got your head around that, then it just becomes fun opportunities for people who want to build cool products. It gives you a chance to stand out. To me, it's all about that member context, right? So what individual members are doing is where I'm thinking about AI, not demographics.
Going back to that border, right? Not having hard borders. Having borders that people move through, not necessarily a cohesive community, but a community of communities means we can't think specific demographics like we used to. And so we don't have a typical member. And so we have to think through what solutions are going to work for every member, whether it's someone at the end of their career just starting out, you know, a teenager who's wanting to start a college fund or save for a car. We have to find solutions that work for both of those member types.
Adam Blue
Yeah. Yeah, I totally agree.
It wasn't me that coined this, but a lot of people talk now about the transition of the idea of community as moving away from geography and maybe a little more toward affinity and identity as driving what's in a community. I think that's a useful concept to think about.
So you and your team have been pretty heavy users of the SDK on platform and built a lot of functionality on your own, which I think is fantastic. And I know that you're one of the early adopters of Q2 Code. So talk a little bit about what the experience has been like in transitioning from handwritten artisanal, you know, Brooklyn coffee shop style software development into a different kind of SDLC where you're driving the “AI software factory” is the current term everybody loves to use. I'll be tired of it probably in another week. But talk about working in the software factory as opposed to you know working in the bakery and like hand folding the dough for the croissant. What's that been like for your team?
Orion Severhill
Well, I'm the wrong person to lean toward that factory model just from how I grew up and how I … You know, in the intro you talked about it. I have a ton of different fields I've played in, you know. Started writing code on my own long before there was AI and struggled through it. It was very humbling to be a software engineer in the earlier parts of my life. And I so I think when we're looking at it, we're trying to keep it artisanal, right? There's a human element that I think makes it artisanal and not that fact … I hate …
Man, I'm going to be with you on that factory thing. There's a human element and the AI element. And we're starting with humans first. In fact that's I think the most critical piece about it and it's the one you hear about the least is how important it is to keep humans involved in software engineering, even if they're leveraging the AI tool.
So our tools, in order to keep it small and keep it controlled, we make them learn to code before they're given any AI, right? That human judgment is really what drives the output on AI. And so when you start to have the AI lead the judgment call, you don't get the same quality and you don't get necessarily the impact that you wanted. You can get a breadth of ideas and whatever you put in, it's going to come back with a correct answer, whether it's right or wrong. But really for us, we've been taking that more hands-on approach and then validate through the beta.
And the results, I wouldn't say they're shocking, but it's amazing to watch human judgment and automated judgment align more often than not. That then the key thing is when we see that happening and we know that the human can push back and they can get to that alignment or they can write the prompt that helps them get to that alignment quicker. That's great.
But we're starting off limited to who could actually access Q2 Code’s beta. And I couldn't be happier that we did. For me, part of it is leadership, right? Like my responsibility is to help level one engineers learn how to become better engineers. Every great engineer I know has been a tinkerer and a problem solver, a puzzle master. And if someone's solving that puzzle for you, you're never going to be the expert in the domain. I want them to grow their career. I want them to move to a point where the AI is a peer or AI acts as a series of assistants helping them get more done. But until they're there, they can't do it.
And the analogy I use with anyone who's asking why—why don't you just let them go at the AI real fast—is autopilots in airplanes and training pilots. Right. Pilots learn how to fly airplanes, and they have autopilot in those airplanes. And yes, they can be taught how to use that particular tool, but they put in tons of manual hours because if autopilot fails, they need to know, this looks like the beginning of a failure. Here's what I do, and here's how I move forward. And that's human judgment. Once they've mastered that human judgment piece, let them use the AI or the autopilot all they want.
Until they get to that point, it's garbage in, garbage out. It's you can't spot the defect on that factory floor, so you’ve got to rework things. And in our environment, going back to it being regulated, you can't really get it wrong too many times before people come asking a lot of questions. So that actually would do the opposite and slow us down.
Adam Blue
Yeah. Yeah, I think you know, for us the way we think about it is there's code that you intend to go into production, right? And that's a pretty special thing. And I think about that one way. And then there's what can I get in 30 minutes of fiddling around that would have taken me a week before and I would have just not done it? And you know, I'll do both, honestly. And I think intention kind of matters, right?
And so, you know, I've in the past, there's been code and I've had an element of that code that I wasn't happy with, but it's like I don't want to go change nine method signatures to move that from a global to something that's passed down a lengthy call chain. So I just make an excuse for myself and then I don't do it. And now with Claude, I just say, Claude, go map this global across the set of available functions that needs it. And it's done in like, you know, 72 seconds, and then I review the code and it looks good, or I say, you know, you keep using this style I don't like. I need you to unwind that.
Claude really likes to put like three different expressions onto the same line of code. And for me at least, this is a personal preference, I don't care for that in terms of readability. It probably ends up being the same Python bytecode. So it's not an optimization issue. It's just like I don't want to have to struggle to read something that was written two months from now when there's a, you know, when I'm tracing it out. But I find it pretty amenable. It'll say, OK, yeah, I'll pull that code out. It's totally fine. And it'll replace it. So that trade-off is interesting.
But I also agree it's interesting. I think of it a little like a musical instrument, a little bit, right? Like you can play the guitar or the piano by ear, and you can learn songs, and you can be really good. But if you really want to understand why a song feels the way it does, you should probably learn the scales and you should learn more than one scale and you should learn to play the song and you should learn a little music theory. Learn how to play the song in another key and see why the key feels different.
And I think there's a place in the world for both of those things, right? I wouldn't want to encourage anyone, don't touch the keys on the piano unless you can tell me what, you know, the Mixolydian mode in C is. That feels … I think our engineering culture sometimes can trend in that direction. I think it's OK for someone to say, you know what, I wanted to try putting this together and see what it was. And I think it's also OK for engineers to say, I love this. It's a fantastic prototype and this will not go anywhere near production because it's just not production level code.
And so I think there's some interesting dichotomy in the way that these tools get used that's thoughtful. But your experience is super, super interesting to think through. But I agree with you. I think people that want to ship code into production should understand how the code works independently of the AI constructing it. And we may be on some arc where that changes, but I don't know.
Orion Severhill
Yeah. I like your musical instrument analogy. It may be better than the pilot one. Although both if you're not having fun doing a barrel roll or whatever when you're flying, or you said just picking it up and noodling around to figure out what you really enjoy, yeah, don't want to kill that part either. I think you hit it on the head. If it's shippable code, if it's going to production and you can't explain what's going on, let's rethink that. It's time for an extra code review.
Adam Blue
Yeah. And sometimes I find the hardest part is accurately describing what success looks like from the piece of code. Like I know how to write it. I know what the algorithm is, but when it comes down to writing unit tests and then integration tests to say, did this code reliably do what you needed it to do? That's where I step back and think, you know, I didn't really think through what the point of this was. And it's the point of it that's got the interesting value.
So how do you guys think about that measurement part of the cycle? Not measurement like, hey, we saved six hours doing it this way instead of that way. Like that's fine. I'm talking about the measurement of how do you know what done is? How do you know what good is? How do you guys think about that? And is there a way that AI can help us with that? Not just generating the lines of code, but helping us getting better at measuring what complete features are and what good features are.
Orion Severhill
Yeah, that's, I mean, that's the trick of product management, right? What's the minimum viable product that you can sell? I haven't quite cracked that. It's interesting to apply AI to product management and watch just incredible ideas that would never work or would work for a very niche market.
But when we're talking about technology, I think one of the cooler benefits is as much fun as unit testing and automated testing, as much fun as code coverage is, giving that to AI and saying, OK, help us think through this problem statement and what are the tests we would need to get there so that we can go in and validate those. Because validating those is pretty cheap. But coming up with all those questions is really expensive. It takes a lot of brain power to sit there, and you could spend days writing out all the cases. If something's already done it for you and then can summarize, here they are, now we're talking about speeding up the process of getting code to members.
And that's the ultimate answer. Like how do we know we're done? When it's something someone can get any value out of, let's put it in front of members and get them that value. That's the outcome that drives success. That’s why in the user story or that, you know, what the outcome statement is in the epic feature statement. That's if the code should all point toward that, and we can get that pretty quickly through AI, we can get here's what minimum viable product looks like.
Adam Blue
Yeah. Yeah One of the people I worked with that was really, really good at product management, a guy named Carl Ryden, who founded PrecisionLender, he used to explain it this way. He would say, nobody really wants a drill bit or a drill. What people want is a quarter-inch hole in the wall. And ultimately they don't really even want a quarter-inch hole in the wall. They want to hang a picture. And so if you think all the way back to the job to be done …
You know, 3M makes these nice little clips now. I don't know if you've seen them, but they like have two fangs on them and then a little hook. And you put it in with a hammer and they'll hold, I don't know, 50 pounds, 75 pounds. And when you take them out of the drywall, if you're careful, they leave two tiny little marks that are so easy. You can almost just ignore them or you can repair them. And more and more now, if I need to hang a picture or, you know, a piece of art or whatever I'm doing, I'll reach for those clips because they're a more direct solution to the problem of I want to hang this picture.
And, you know, I've got a drill and I've got drill bits and I've got drywall anchors and I've got all this other stuff, and I have a laser level and … but I don't want any of that, right? That's the stuff that's in my way in getting the thing I want, which is I want to hang, you know, my Magic Johnson jersey or I want to hang, you know, a cool picture I got from Mr. Brainwash.
And so I think sometimes we can get very focused on the how we're going to implement something and fall in love a little bit with the implementation and skip over that really thoughtful, like, what is the job really to be done? What is the value to be unlocked?
And I think that's why it's interesting when you talk about product management and AI, man, it'll go off. I agree with you, It'll spiral off into an interesting location. And sometimes that can be really valuable, or it can just be, you know, it at least has value in understanding that's not what we want. But that cycle is interesting.
And so I think as, you know, as we continue to work with AI … It's funny. You know the term “cargo cult,” I'm sure, where, you know, a piece of technology or even just an object kind of falls out of the sky and a relatively primitive civilization comes to attach meaning to it that's not really there.
Like that concept, some days I walk around and I think that's a little bit what it's like working with an LLM or generative technology. It's got a little bit of a cargo cult mentality to me in a way that I think is a little troubling. But I think there's tremendous potential there that we're all still trying to unlock at the same time.
And so let's switch gears and just philosophically, let's talk a little bit about, you know, one of the objectives I think for your credit union and almost all financial institutions is to help people build wealth in a way that gives them financial security so that they feel comfortable doing the things that are important to them. And so maybe talk a little bit about how you see AI and the intersection of AI and that as like if that's the job to be done, make people secure, help them build wealth, help them plan for the future and do the things that are important for their families. Where does AI figure into that for you?
Orion Severhill
Yeah, that’s a fun one. Whoever cracks that one absolutely kind of wins the race. I think the way I see AI helping us as a, you know, an industry, the smaller, not the big five banks, but the rest of the industry is we now have access to tools that really can help us help the communities. We don't have to be the laggards behind Erica or Fargo. Right.
It's not eight years ago and they came out with that and everyone was wondering, well, where are we going to get that budget to build that internally? We can now work with partners who can bring that to us. Meaning we can offer personalization to members. We can spot trends in their data and help them understand where they're maybe about to trip over something they didn't see and save them a bunch of NSF fees and going to you know, payday lenders. We can offer them the right product to just get them to where they need to be and make it hyperfocused on them, not here's a loan for anybody. It's here's a loan because we've noticed things that are happening with you. I think that's where the AI technology is leading us. And what I'm hoping for is, you know, we can capture enough ground swell there that we actually see …
You know, I’m proud of being born and raised in San Diego. Some of the internal metrics we're talking about is how can we just make everyone in San Diego have better generational wealth and pass on what, you know, instead of living day-to-day, have enough saved to pass it on or have assets that, you know, future generations can take advantage of.
How do we measure that? I don't even think that metric was possible without AI unless you hire an economist and sit them in a corner and kind of leave them alone to not work for the institution but work for the community. And we can now do that. So we can look at solutions, we can think about our features we want to add, and we can actually start to toy a little with what those metrics look like and what the projected outcome would be. And if we do it at a finite scale, we can actually see, OK, last year or last month did that move the needle?
Getting data, tracking data, analyzing data has been a pain point for my entire career. It's been the fun part for a lot of it, because you’ve got to source it and you’ve got to analyze it. I used to love dabbling in Perl for it. AI will let you just ask the right questions that you're building the right things and then measuring and adapting quickly with that long horizon, still possible.
Adam Blue
Yeah. Yeah, I think there's a real unlock there. And I think, you know, to your point, that the visibility and the ease of getting to an understanding of where you are is really critical for where you're going. I think, you know, people make a lot of financial decisions. Multiple times a day people make a financial decision. And oftentimes they're probably making it with an absence of some of the data or context that would be helpful for them, you know, to understand.
And this goes way beyond education. Some of it's just like understanding what cash flow is or understanding what, you know, prices are likely to be. And so bringing people more of that information and letting them make decisions in a more rational way because they have access to a broader set of information that they can coalesce. And then at the same time, you know, I think that's a noble goal is how do you lift the entire kind of wealth level of an entire city, a large city.
And first you've got to be able to measure it and you've got to be able to understand which things seem to have a positive impact and which have a negative impact and then incent those behaviors. And so that's interesting. And I think there's a lot of data analysis that AI is starting to unlock that could be really interesting as well.
Is there anything specific you guys are working on in that area you want to share, or is this just kind of a vision that you're holding as you go forward?
Orion Severhill
It's a small cohort of us that are just noodling through how we do that. You know, the member-facing platform is the primary key though. You know, it's the biggest branch. Digital banking is our biggest branch in the industry. Everyone should agree with that. It's not a nice to have, it's not just the branch for most people. There are entire banks that don't have a physical branch and they're rather successful.
So figuring out how to push ideas and push knowledge out without being prescriptive in a, you know, a general “you have to do this” sense, but “hey, here's an idea for you.” And how do you attribute that to people's movement? It is tricky. So we're having to play with it a little bit. You know, I know you've got an economist background. I'll probably tap you on the shoulder at some point and say, hey, help me think through some of this math.
You know, it's an interesting way to think through things, but I think it's also directionally an indicator of what we're doing with every feature. That question is asked with every feature we build. Is this going to make a specific member's life better? And can every member take advantage of this? So it's part philosophy, part we're trying to make it practical so we can, like you said, raise a large city and, you know, give it the support that it deserves.
Adam Blue
Yeah. I hear a lot of empathy in that kind of talk track. And I think that's something that some days can be in really short supply. So it always warms my heart, you know, to hear people talk through, you know, what that might look like.
All right. Well here at the end of the podcast, I like to throw in a piece of art or literature for people to think about and enjoy. I've got a good one for today. So William Gibson, who's a very famous author, wrote a collection of short stories called “Burning Chrome.” And there's one in it that people don't really talk about very much. I think it was first published in Omni magazine. And it is it's called “Hinterlands.” And it's about a cargo cult that sits at the Lagrange point in which humankind comes into contact with radically advanced technology from time to time and what it does to the people who are at the edge of it.
And you know, our talk today just really reminded me of that story and how prescient I think a lot of that writing was. So if you get a chance, it's in the “Burning Chrome” collection. You can pick it up just about anywhere. Certainly you can get it on Amazon Kindle. You can probably check it out of the library. And “Hinterlands” is the short story that kind of matches up with today.
So thanks very much for being on today. Orion Severhill has joined us from Mission Fed Credit Union, and this was really great. Really appreciate it.
Orion Severhill
Appreciate the invitation.
Adam Blue
Absolutely. All right. Well, thanks for joining on Cut the Context. You can find us wherever your fine podcasts are bought and sold.