AI in banking product development: Where tools actually matter

Banks have invested millions in AI tools, and the impact is already clear. Code generation time has collapsed. At Goldman Sachs, GitHub Copilot is now running on 12,000 developer desktops. Citizens Bank is reporting 20% productivity gains. However, features often still take 30 days from idea to release. The bottleneck hasn’t moved; it’s just shifted. 

AI tooling doesn’t fix delivery unless it does one of three things: reduces friction, improves quality, or gives time back. Most banks have added speed to tasks but not to outcomes, which is where the gap remains.  

The coordination tax 

In software and product development lifecycles, the biggest drain on time isn’t the work itself – it’s the coordination around it. Developers spend hours clarifying vague requirements between product and engineering. Often, teams are stalled waiting for approvals from risk, compliance or architecture review boards. Manual handoffs then cascade between teams: epic to story, story to code and code to test, slowing momentum. Context switching only increases the hold-up and entire afternoons dwindle when suddenly someone must justify an undocumented decision made three sprints ago.  

That’s where AI steps in, by clearing the bottlenecks. 

From conversations with banking teams we learned that developers can spend up to four hours per sprint writing handover documentation. By embedding AI summarisation into Jira, that can be reduced to 20 minutes. This gives 180 hours back to engineering each month. In this case AI isn’t making documentation better, it’s making it unnecessary as a separate task. 

That’s just one example. Across the lifecycle, AI is quietly removing friction:  

• Epic clarifications generated before sprint planning reduced ambiguity by 40% in pilot teams.  

• Automated compliance checks embedded in the pipeline catch issues at commit time, not three weeks later during review.  

• Semantic search across past decisions means teams stop re-litigating problems they’ve already solved. 

JP Morgan was the first major bank to roll out an LLM-based assistant, reporting a gain of 2-4 hours per week. Not from writing code faster, but from finding answers without chasing people down. 

The pattern is clear: AI removes waiting, not work. 

The rework problem 

Quality problems create rework loops. A defect caught in production costs 10x more than one caught in development. AI helps to tighten those feedback cycles. 

Code reviews now flag logic errors, security gaps and anti-patterns before human review. Automated testing now generates test cases directly from acceptance criteria, removing the need for quality assurance to write them manually and edge case simulation catches scenarios humans miss. Real-time detection spots requirements drift between story and implementation. 

The value isn’t ‘better code.’ It’s fewer defect cycles and less rework, which can save weeks. 

One bank cut enhanced due diligence report time by 88% with 94% model accuracy. The AI didn’t replace compliance teams, it eliminated the manual validation loops that consumed their time. Compliance officers stopped checking data and started making judgment calls that require expertise. 

Goldman Sachs reports that their fine-tuned AI assistant, trained on their entire codebase, catches context-specific issues that generic tools miss. It knows which patterns violate internal standards before code review even begins. 

Banks are using AI testing platforms to help teams accelerate regression testing and streamline backlog management, using intelligent triage to refine user stories and priorities. The difference between working AI and theater is measurable: error rates down, rework cycles eliminated, defects caught earlier. 

Attention economics 

AI doesn’t just make work faster. It gives attention back to the most important parts of work. 

Product managers stop writing boilerplate user stories and spend time on customer research. Engineers stop documenting decisions and spend time solving hard problems. Risk teams stop chasing clarifications and spend time on judgment calls that require expertise. 

This is attention reallocation, not headcount reduction. The goal isn’t to do the same work with fewer people. It’s to let people focus on higher-value work. 

In conversations with product and engineering teams, we’ve learned that developers save 12 hours per sprint on documentation, testing and review cycles. That time normally goes to refactoring legacy code and reducing technical debt, work that is often deferred for months. With newfound capacity, engineers can significantly reduce technical debt in the months ahead. 

Citigroup equipped 30,000 developers with generative AI coding assistants. The goal wasn’t faster feature delivery. It was freeing capacity for legacy system modernization. Converting COBOL (common business-oriented language) to Java, speeding up cloud migration, work that creates long-term flexibility but never makes the sprint because urgent tasks always crowd it out. 

For example, when IBM’s Watsonx translates 1960s-era COBOL mainframe code to modern languages, it doesn’t replace engineers. It removes the grunt work so engineers can focus on architecture decisions and business logic that still require human judgment. 

Capital One reports that their ChatGPT-style assistant for 20,000 customer service reps measurably improved call handling times. Not because AI answered calls. Because it gave reps instant access to policy information, freeing them to focus on the customer’s actual problem instead of hunting through documentation. 

The pattern repeats: AI removes the low-value work that blocks high-value work. 

Why most pilots die in the demo phase 

Most AI pilots fail not because the tools are weak but because the organization isn’t ready. 

Three failure modes have shown up consistently: 

• Tools deployed without process redesign. AI sits alongside legacy workflows, not inside them. Developers have copilot but still spend two days waiting for architecture review board approval. The code gets written faster but ships at the same speed. 

• No change support. Teams get told to ‘use the copilot’ without training, champions or feedback loops. An MIT 2025 report showed 95% of GenAI pilots failed to yield business impact, only 5% reached production at scale. The core issue wasn’t the AI tech. It was that pilots operated as isolated science projects disconnected from real workflows. 

• Metrics focus on task speed, not outcome velocity. Banks celebrate faster code generation while delivery timelines stay flat. We’ve seen banks automate loan approvals with AI but compliance violations spike because regulatory requirements haven’t been built in from the start. The tool worked. The system didn’t. AI can’t outrun organizational drag. If your approval process takes 10 days, faster coding saves you nothing. 

Goldman Sachs succeeded because they fine-tuned AI on their entire codebase and embedded it in the IDEs and CI pipelines employees already use. They didn’t treat it as a side project. They invested in training engineers and measuring time saved. Enterprise-wide integration, not innovation theater. 

JP Morgan has 450+ AI use cases in production because they baked AI into the C-suite agenda and built a governance framework. Models go from ideation to production with proper controls. Top-down support plus operational discipline. 

The banks stuck in pilot purgatory have fragmented tooling, no ROI tracking and culture gaps. One-off bots built in silos can’t plug into enterprise workflows. Without KPIs tied to business value, reduction in dev cycle time, defect rate, projects lose executive support and fizzle out. Most failures are operational, not technological. 

What actually matters 

Most banks are asking: which AI tools should we buy? 

The sharper question is: where do we lose time, create rework and block progress today? 

AI only delivers when it removes those specific losses. If you can’t name the exact friction points in your software and product development lifecycles, the handoffs that take days, the reviews that loop endlessly, the clarifications that eat afternoons – you’re not ready to deploy AI effectively. You’re ready to waste money on tools that sit unused. 

But if you can name those friction points precisely, the path forward becomes clear: embed AI where work actually slows down, where quality suffers and where people waste attention on tasks that don’t need human judgment. 

Do that, and delivery timelines will compress. Ignore it, and you’ll have expensive tools inside the same slow system, just with better demos. The banks that win won’t be the ones with the most sophisticated AI. They’ll be the ones who know exactly where their time goes, and use AI ruthlessly to get it back.