Q&A: How AI Coding Assistants are Reshaping Software Development Across Financial Services Organizations

As Artificial Intelligence (AI) coding assistants continue to gain traction across enterprise and government environments, organizations are moving past early curiosity and into a more critical phase of adoption. The question is no longer whether these tools can generate code, but whether they can deliver reliable, secure, and scalable outcomes in high-stakes environments and highly regulated industries where performance and accountability are non-negotiable.  

As expectations rise, so does the need to understand how these tools actually perform beyond the hype, and what it takes to use them effectively at an enterprise level. 

To learn more about the benefits of using AI coding assistants across critical business processes, we sat down with Alex Duan, Chief Executive Officer at Holokai, an AI company focused on applying advanced language models to real-world software development workflows. 

Financial Technology Today (FTI): How did you initially start working with AI coding assistants?    

Alex Duan: I had used AI coding assistants occasionally before starting Holokai, but we really leaned into them once we began building the company. We wanted to “eat our own dog food” and see firsthand whether they could deliver meaningful efficiencies. 

“Vibe coding is low-stakes and driven by experimentation and speed. Enterprise software is high stakes, where reliability and accountability matter a lot more.” –Alex Duan

Coming from 20 years in development, I’ve developed strong opinions about how things should be done, so I was skeptical of the hype around AI coding. As we explored how to apply our enterprise middleware experience to the AI space, we decided to test coding assistants directly – starting with tools like Claude Code, which were popular at the time. 

Ultimately, it was about validating the hype and understanding whether these tools could genuinely accelerate our product development. 

FTI: What’s the difference between Vibe coding and developing enterprise software?  

Alex Duan: The difference really comes down to things like scalability, durability, supportability, and security – that’s what defines enterprise software. 

Vibe coding is more about those immediate endorphins. You type something in a chat window, and it generates something that works, looks cool, and you can play with or share. It’s that “good feels” moment – the magic of creation without having to do much beyond prompting it. 

Enterprise is the opposite mindset. You need to support the solution for the rest of its life. If something breaks, there are real consequences – you could lose money, lose customers, or worse. If you’re a bank, payments have to go through every single time. If it’s mission-critical, failure isn’t an option. 

Vibe coding is low-stakes and driven by experimentation and speed. Enterprise software is high stakes, where reliability and accountability matter a lot more. 

FTI: How did you develop into more sophisticated AI coding approaches?  

Alex Duan: I evolved into more sophisticated approaches by realizing that a lot of the early methods were over-engineered. Traditional approaches try to recreate enterprise workflows with multiple AI “roles,” like architects or Quality Assurance (QA), but it’s still the same underlying model, just asked to wear different hats. So, you don’t actually get better thinking, just more complexity. 

“[Using different models for different roles is] basically like having two different brains working on the problem instead of one model pretending to be everything.” –Alex Duan

What worked best for me was using different models for different roles. I’d use Claude Code to actually plan and build the code, then use ChatGPT to audit it. I’d have ChatGPT evaluate the output against a rubric, score it, and generate specific feedback, then feed that back into Claude to refine the code. 

It became an iterative loop: Claude builds, ChatGPT critiques, and then Claude improves. It’s basically like having two different brains working on the problem instead of one model pretending to be everything. That’s where it became more sophisticated, and the quality really improved. 

FTI: What are the pros and cons of coding with AI agents?   

Alex Duan: The biggest con is code quality. If you’re not experienced, the AI will make bad design decisions, and you won’t know how to correct them. You end up with “AI slop,” code that works, but isn’t maintainable. Then you keep building on top of that, and now you’ve got a bloated, spaghetti codebase that’s hard to debug and expensive to fix. 

There’s also the context problem. The AI doesn’t remember everything, so it can duplicate logic or make inconsistent decisions, which adds even more bloat over time. For example, in enterprise software you want common core components, libraries, security frameworks, or messaging layers that are leveraged by all other parts of the software.   

This gives you a logical, scalable, and supportable codebase over time.  AI will tend to build duplicate pieces with every new feature.  So, as you scale, you must deal with multiple, interdependent pieces to troubleshoot, fix, and improve your code.  And as that grows, every fix or feature gets more expensive – both in tokens and effort. 

“Each developer can effectively multiply their output without needing a huge team.” –Alex Duan

Cost is another con. It feels cheap at first, like wow, I just built this instantly, but at scale you can be spending hundreds of dollars a day on tokens. And once you’re deep into that codebase, it’s hard to unwind. 

On the pro side, it’s a massive productivity boost if you know what you’re doing. You can guide it well and get really solid, even enterprise-grade code. At that point, you’re not just coding; you’re managing multiple agents across multiple models. I can work across backend, UI, and other parts of the system simultaneously. 

It also lets small, senior teams scale a lot faster. Each developer can effectively multiply their output without needing a huge team. As a developer, you don’t have to rely on other people as much; you can just build. So, it’s extremely powerful, but very dependent on the operator. If you know what you’re doing, it’s a huge advantage. If you don’t, you can create a mess really fast. 

FTI: Your team used AI to help manage a backlog of work for an outsourced development team. How did that work? What were the challenges?    

Alex Duan: AI was effective at generating and managing the backlog itself. If you asked it to build a feature or an application, it could create a detailed and accurate backlog based on standard development practices – things like sprint planning, SDLC steps, all of that. 

The challenge was that we almost didn’t need the backlog anymore. Instead of breaking work into small sections and going through traditional backlog grooming with a project manager, we could just sit down with AI, plan a feature in detail, and execute it directly.  

That created some friction. Project managers still wanted structured backlog sessions, but from our perspective, it felt like overhead. At the same time, skipping that process introduced gaps – things could get missed, and there was less built-in accountability compared to traditional tracking. 

FTI: At the end of the day, did AI coding work for you personally or not?   

Alex Duan: Yes – it worked tremendously, but with a lot of caveats. 

If you know what you’re doing, it’s incredibly powerful. I can move faster, work across multiple parts of the system at once, and basically act as if I have a team of developers working for me. That’s a huge win, especially at the enterprise level where speed matters. 

“For senior engineers to get real value, they need that experience to validate, guide, and correct the AI.” –Alex Duan

But it’s also really easy to shoot yourself in the foot. If you’re not experienced, you’ll end up with a bunch of AI slop, and then you’re stuck dealing with that later. And even when you do it right, there are tradeoffs with cost, complexity, and long-term maintenance. 

So, it worked for me, but that’s because I know how to control it. If you don’t, it can create more problems than it solves. 

FTI: What needs to happen for experienced, senior engineers like yourself to get real value from AI in coding enterprise software?  

Alex Duan: For senior engineers, AI is amazing – you can get a lot of value out of it because you already know what looks good and how to guide it. 

The issue is that AI makes everyone feel like an expert. It accelerates people up that initial confidence curve, but they’re still lacking the depth and experience underneath it. So, you get people building with confidence, but without judgment to know if what they’re producing is good. 

For senior engineers to get real value, they need that experience to validate, guide, and correct the AI. Without that, it’s easy to misuse it and think you’re doing great in the short-term than you actually are in the long-term. 

Originally published on Financial Technology Today.