Shifting Developer Effort from Writing Code to Reviewing It: An AI-First Engineering Model in Practice

A global fintech company serving corporate treasury and finance teams across industries decided to become an AI-first engineering organization. Their platform manages cash, payments, and risk analytics for enterprise customers in dozens of countries. The work requires speed and accuracy to appropriately service their customers, but their complex codebase, a mix of modern and legacy systems and the need to enhance while maintaining existing systems, hindered rapid delivery.

To move from experimentation to an AI-native way of working, the client needed a repeatable development process, clear human-in-the-loop checkpoints, training for engineers on how to direct and evaluate AI agents, and practical tooling that would help teams adopt the model safely. To design and operationalize that new way of working, they hired SPR for AI consulting services.

The client’s engineering leadership saw how AI tools could shorten development timelines if used strategically, not to replace developers but to change their focus. Engineers would guide AI agents to handle implementation, allowing them to concentrate on requirements, architecture, and reviewing outputs.

This model had real promise, but it also introduced a new kind of risk. AI-generated code can look plausible yet still be wrong, making incorrect assumptions, choosing inefficient patterns, or missing dependencies that a developer with domain knowledge would catch immediately. The client needed experienced engineers who could evaluate AI output with the same critical eye they would bring to any code review.

The client’s environment created challenges for AI development due to a complex, poorly documented codebase of about 15 gigabytes across distributed systems that have evolved over many years. When issues crossed boundaries or depended on implicit links, AI tools struggled to understand the full picture. Imprecise prompts worsened this; for example, an AI assistant assumed two data sources were both in US dollars and without explicit clarification, leading to technically correct but inaccurate outputs. Detecting such errors required domain expertise. The lesson: cleaner, self-contained systems allow AI to be more reliable, whereas tangled systems need more human oversight.

SPR consultants embedded within the client’s engineering teams to help turn AI-first development from a promising idea into a working operating model. The work included defining where AI agents could safely accelerate delivery, where senior engineering judgment was still required, how teams should prompt and review AI outputs, and what tooling or automation was needed to make the process repeatable across teams.

• Autonomous defect resolution with Devin AI. For bug fixes, the team deployed Devin AI with direct access to the client’s code repositories. When a defect is reported, engineers hand off the investigation to Devin by describing the issue, pointing the agent at the relevant system, and letting it navigate the source code, reproduce the problem in a virtual environment, make the necessary changes in a new branch, validate the fix, and open a pull request, all without human intervention. SPR engineers then review the changes, tested locally, and either approve the pull request or provide structured feedback to the agent for another iteration. For well-scoped issues contained within a single application, this workflow significantly reduces the time from bug report to resolution.
• Multi-pass AI code review with Claude involved reviewing code in stages instead of treating the first AI response as final. Claude was used for initial code generation or review, engineers then assessed the output for gaps, edge cases, and inefficiencies, and subsequent Claude sessions were used to re-check the revised code or explore additional concerns. This layered approach helped identify legitimate issues earlier, often catching problems in a second pass before human review, while also improving pull request clarity.
• AI-powered automation via playbooks helped the client automate routine processes like generating change request tickets with Devin AI. Engineers instruct the agent to fetch work items from Azure DevOps or Jira and draft change requests, which humans review and confirm. This reduces admin work and ensures a repeatable, auditable process.
• Environment setup scripting addressed a longstanding onboarding bottleneck. The client’s local environment, previously taking two weeks to configure due to size, complexity, and scattered documentation, was streamlined. Our team consolidated the documentation and used Claude to generate error-handling PowerShell scripts that can run safely multiple times. This reduced local environment setup time from two weeks to under one, greater than 50% reduction in onboarding setup for new engineers.

A pattern emerged across SPR’s work: the AI-first model succeeds because of the experience of those running it, not despite it. Less experienced developers may accept AI-generated plans or code without question, but senior engineers see the flaws, understand when assumptions don't hold, and know how to craft prompts for reliable output. This reflects how senior engineers operate: understanding requirements, making architectural decisions, delegating implementation, and remaining responsible. AI agents are a new implementation resource requiring domain knowledge and senior judgment, which SPR’s consultants provided.

With SPR’s consultants embedded, the client’s AI-first model advanced from ambition to reality:

• Faster defect resolution: Poorly scoped bugs were investigated, fixed, and reviewed by AI, shortening cycle times.
• Rapid service development: AI generated code, with self-review, before human review.
• Less admin work: Routine tasks like change requests were automated, freeing engineers for complex tasks.
• Improved onboarding: Scripting cut environment setup from two weeks to under one.
• Responsible AI delivery: SPR established workflows, review practices, and guardrails for quality and speed.

The client’s experience makes it clear that AI-first development is not simply a matter of adding AI tools to an existing software process. It requires a new way of working, with clear workflows, trained teams, human oversight, and the right automation to support delivery at scale. The speed gains are real, but they depend on experienced engineers who can direct AI agents, evaluate their output, and build the guardrails that make AI-native development reliable. For organizations pursuing this model, SPR helps establish the process, practices, and engineering discipline needed to make AI a productive part of everyday development.