AI Coding Moves Beyond Code Generation to Reshape Fintech Software Delivery

The Next Phase of AI Isn't About Models Alone

Large language models keep getting smarter, but the competitive frontier has shifted. The real question now is how to build reliable agents, mature AI engineering systems, and keep AI stable in complex business environments. That's the conversation driving AICon Shenzhen, a global AI development and application conference set for August 21–22.

The full agenda just went live, and it's packed with sessions on agent engineering, AI infrastructure, AI-native development, and embodied intelligence. One talk that stands out for anyone in fintech or enterprise software delivery comes from Li Weining, HSBC Technology's internal open-source lead. His session, titled “From Code Generation to R&D Closed Loop: AI Coding in Fintech SDLC,” digs into how HSBC is turning AI coding from a developer convenience into a governed, full-lifecycle capability.

Why Fintech Needs More Than Smarter Autocomplete

Fintech software development is a pressure cooker. You're dealing with strict regulations, security requirements, and systems where a small bug can cascade into serious financial or reputational damage. Li, who has spent 15 years in the field across HSBC, GAC Auto Finance, Xiaopeng Leasing, and Hang Seng Bank, knows these constraints firsthand.

Early AI coding tools were essentially fancy autocomplete. They generated snippets, saved a few keystrokes, and made individual developers feel faster. But that's not enough when you're building payment systems or trading platforms. The context is huge, the quality bar is high, and the compliance overhead is non-negotiable. Li's argument is that AI coding needs to evolve from a personal productivity hack to something that supports the entire software development lifecycle (SDLC) in a controlled way.

From Snippets to Workflow: The Real Shift

The core shift Li describes is moving from code generation to process improvement. It's one thing for an AI to write a function; it's another for it to help you understand a vague requirement, flag risks in a design, or verify that a test actually covers the right scenario.

At HSBC, different teams were already experimenting with AI in their own ways. Some were using it for requirement analysis, others for code reviews, still others for test generation. The problem was that these efforts were siloed. There was no way to share what worked, no standard for quality, and no governance over what the AI could access.

That's where internal open source comes in. HSBC created a community around AI coding practices, letting teams contribute their prompts, workflows, and lessons learned. Over time, these scattered pieces turned into reusable tools and what they call Agent Skills—packaged capabilities that can be plugged into different stages of development.

Agent Skills Across the SDLC: A Practical Walkthrough

Li's talk breaks down how Agent Skills apply to each phase of software delivery. It's not just about writing code anymore. It's about weaving AI into the whole fabric of how software gets built.

Requirements and Design

In the requirements phase, agents integrated with Jira and Confluence help teams parse user stories, ask clarifying questions, and surface missing details before anyone writes a line of code. That's a huge time sink in most projects, and AI can cut through the ambiguity.

During design, agents assist with architecture options, impact analysis, and technical decisions. They can pull up relevant past designs, flag potential bottlenecks, and even compare trade-offs—something that used to require a senior architect's gut feeling.

Coding, Review, and Testing

For coding itself, HSBC combines VS Code, GitHub Copilot, and custom Agent Skills to keep developers in flow. But the interesting part is how they use AI for code review. Agents check for security flaws, style violations, and logic errors before a human reviewer even looks at the diff. It's a safety net that catches the obvious stuff early.

Testing is another area where AI shines. Agents generate test cases based on the code and requirements, analyze why a test fails, and suggest fixes. They even help close the loop by verifying that the fix actually resolves the issue. This moves testing from a tedious chore to a more intelligent, semi-automated process.

Making Agents Work Inside the Enterprise

One of the biggest hurdles in fintech is getting tools to talk to each other securely. That's where MCP (Model Context Protocol) comes in. Li emphasizes that MCP is the glue that connects agents to the tools and data they need—whether it's Jira tickets, Confluence pages, or code repositories.

The goal is to move from single-purpose assistants to cross-tool agent workflows. Instead of a developer copying a context from Jira into a chat window, the agent can pull the relevant context directly, act on it, and update the ticket when done. It's a more seamless experience, but it also raises serious governance questions.

Security, Compliance, and Governance: The Non-Negotiables

In fintech, you can't just let an AI run wild. Sensitive data leaks, unauthorized tool access, and accidental compliance violations are real risks. Li's talk addresses how HSBC sets boundaries around AI coding. They have strict data controls, audit trails, and permission systems for Agent Skills.

Every skill is versioned, reviewed, and tested before it's shared. There's a clear process for who can use what, and everything is logged for compliance. It's not about shutting down innovation—it's about making it safe enough to scale.

Scaling from Pilot to Organization-Wide

Getting from a small pilot to a platform used by thousands of developers isn't easy. Li shares how HSBC started with internal open-source projects and grew them into a company-wide AI coding platform. The key was finding high-value use cases that could be replicated across teams.

They also focused on developer adoption. Training, feedback loops, and community management were just as important as the technology. You can't just roll out a tool and expect people to use it. You need to show value, address concerns, and build a culture that embraces AI as a partner rather than a threat.

The metrics matter too. How do you measure the impact? It's not just about lines of code generated. It's about time saved, defects caught early, and consistency across teams. HSBC is working on quantifying these benefits to make the case for broader investment.

Key Takeaways for Anyone in Enterprise Software

Li's session offers a clear roadmap for any organization trying to move AI coding beyond the hype. Here's what stands out:

• AI coding is not a silver bullet. It requires careful integration with existing workflows and tools.
• Internal open source is a powerful way to capture and share best practices across teams.
• Governance isn't a blocker—it's an enabler for scale.
• Start with specific pain points, not with technology for its own sake.
• Adoption is a people problem as much as a technical one.

The conference will feature more than 50 experts from companies like Alibaba, Tencent, Huawei, Kuaishou, and Google Cloud, covering topics from agent safety to embodied intelligence. But if you're in fintech or any regulated industry, Li's talk might be the most practical one on the list. It's about making AI work in the real world, where rules matter and every line of code counts.