Most conversations about AI in software development start with code. How much faster can developers write it? How many tasks can an AI agent take over? These are fair questions, but in financial services, faster coding alone doesn’t necessarily lead to faster releases. For leaders investing in AI, the question is whether that investment helps deliver business priorities sooner and more predictably. Research points the same way: in Faros AI’s 2025 study of more than 10,000 developers, teams with high AI adoption completed 21% more tasks and merged 98% more pull requests, yet company-level delivery metrics showed no significant improvement.
Code is only one part of delivery. Every change also needs clear requirements, review, testing and approval. When AI speeds up coding, bottlenecks elsewhere in the process can become more visible. In our experience, the most effective fix is often found earlier in the process.
What we saw in practice
We work with banks and insurance providers. For one of them, a specialist insurer, our team is building a customer-facing mobile application that the client also offers in white-label versions to its partners. The client came to us with a product concept and high-level requirements. Gathering and refining the detailed requirements was our responsibility.
That’s where the bottleneck appeared. Our Product Owner, who also acted as the Business Analyst, was managing the product vision, the backlog and the analysis at the same time. As the product grew more complex, tasks often reached the team without enough detail. Developers got blocked, and work went back to refinement. Coding wasn’t the slowest step – analysis was. Although this example comes from insurance, the underlying problem is relevant to banks: unclear requirements create rework and make changes harder to verify.
The change started inside the team. Our solution architect began using spec-driven development to document how features should be implemented. He soon noticed that the Product Owner could use the same approach for business requirements. Together, they set up a dedicated repository for specifications and a simple workflow:
- Business specification. The Product Owner works with an AI agent that has a dedicated skill for gathering requirements. The agent turns meeting notes and transcripts into a structured specification – definitions, overview, use cases and user stories with Given/When/Then acceptance criteria – and the Product Owner reviews and refines it.
- Technical specification. During refinement, the team uses an AI agent with a dedicated skill to describe how the feature will be built – the architecture, integrations and impact on the backend and frontend. Developers make the architectural decisions and validate the constraints.
- Implementation. Developers give the approved specification to a coding agent instead of a blank prompt, so the generated code starts from agreed decisions.
- Testing. Testers write tests based on the same specification, so they can start before development is finished.
At every step, AI prepares the draft and people make the decisions.
What changed
In the team’s experience, the biggest shift happened before development started. Developers now raise far fewer questions about specifications after refinement. Fewer tasks go back to refinement, sprint scope changes less often and tasks spend less time blocked. More tasks are fully ready for development by the time they reach the team, and preparing specifications and acceptance criteria takes much less time. The team also sees fewer bugs, though that effect is smaller.
The difference is easy to see in the backlog. Before, a typical user story had a brief story statement, a short list of requirements and a handful of acceptance criteria. Now each story comes from an approved specification, with a clear goal, detailed Given/When/Then criteria for each scenario and technical notes on constraints and fallbacks.
Review became easier, too. Requirements and the planned solution are now verified much earlier, before development starts, so reviewers check code against decisions that were already agreed.
Feature lead time has stayed about the same so far. The early gains showed up as fewer blockers, less rework and more predictable sprints. The next step is to measure whether those improvements translate into shorter lead time and identify what still limits it.
The math behind the bottleneck
Simple math shows why speeding up one step rarely shortens delivery on its own. Say coding a change takes five days, and review, testing and release preparation take another 15. If AI cuts coding time to two days, coding time drops by 60%. Total delivery time, however, drops by only 15% – from 20 days to 17. And if both versions of the change miss the same release window, the production date doesn’t move at all.
Why it matters in banking
Banks don’t release software the way startups do, and for good reason. Since January 17, 2025, the Digital Operational Resilience Act (DORA) has required EU financial entities to manage ICT changes in a controlled way: changes need to be recorded, tested, assessed and approved.
At the same time, trust in AI-generated code is limited. In Google Cloud’s 2025 State of AI-assisted Software Development report, 30% of respondents said they have little or no trust in the code AI produces. In a regulated environment, that code still has to be checked by someone who can take responsibility for it.
The real challenge is making sure that every change reaching review, testing and approval is clear, complete and easy to verify.
What we learned
Control works better upstream. Reviewing more and more AI-generated code at the end of the process doesn’t scale. Agreeing on intent, scope and constraints before implementation means there’s less to fix later.
A specification isn’t more documentation. Its value lies in when decisions happen. A good spec moves decisions from expensive stages, such as review, testing and approval, to a cheaper one – before the first line of code is written. For banks, there’s an extra benefit: a specification records the intent, scope and constraints of a change before any code exists. In this project, topics such as permissions, access control, change history and rollback are covered when a feature requires them. In a banking project, we’d go further and make data and privacy, security and access controls, audit requirements and a rollback plan a standard section of every specification, scaled to the risk of the change. Risk and compliance teams could then review it early, instead of piecing together approval evidence just before release.
Process comes first, AI second. AI makes specifications fast to create, but the real value comes from a consistent structure the whole team agrees on.
Measure the whole cycle from the start. Looking back, we should have established an end-to-end delivery baseline before evaluating AI productivity. Time saved on coding tasks alone can’t show whether changes reach production sooner.
Where to start
For banks planning their next step with AI in software development, this is what we’d suggest:
- Find where work waits the longest – unclear requirements may be a bigger constraint than coding.
- Pick one team and one type of feature to try a spec-driven approach.
- Agree with risk and compliance teams on what a specification must include, in proportion to the risk of the change, such as security controls, audit needs and a rollback plan.
- Measure blocked time, rework and returns to refinement alongside lead time, then scale what works.
AI is already changing how financial institutions write software. Whether it also changes how fast they deliver depends on everything around the code, and much of that work happens before development starts. In banking, three days saved on coding pay off only when the rest of the cycle keeps pace.
About the author
Tomasz Krakowczyk is Financial Services General Manager at Software Mind. With over 15 years of experience in IT management, Agile coaching and transformation leadership, he helps financial organizations align software delivery with business strategy.
About Software Mind
Software Mind, a global technology partner, provides consultancy and software development powered by AI accelerators. Delivering deep AI, cloud and data expertise, Software Mind turns technology into measurable business value so that companies across industries can thrive in competitive markets.
www.softwaremind.com
