Citigroup's 14,000 Engineers Use Devin, Slashing Tech Outsourcing Bill by 30%

Citigroup cut its technology vendor bill by 30% within a year, driven by 14,000 engineers adopting the AI coding agent Devin. The figure comes from public remarks by Citigroup CEO Jane Fraser, but one caveat needs to be clear: the 30% reflects a single cost line, not a reduction of the entire technology budget. In the same week, JPMorgan's CFO said AI-driven gains have yet to show up in its financial results. The two financial giants' stances on AI coding tools reflect the real divergence underway in enterprise AI adoption.

What Devin Does at Citigroup: From "Writing Code" to "Planning Work"

According to analysis by a YouTube channel, the Devin deployed at Citigroup is not a simple code-completion tool. It can autonomously plan tasks and write code, directly replacing some development services previously provided by external vendors. This means the role of AI agents is shifting from "assisted programming" toward "end-to-end delivery."

In the traditional model, banks pay different vendors separately for requirements analysis, architecture design, coding implementation, and other stages. Devin's involvement lets Citigroup internalize some standardized development work, thereby compressing outsourcing spending. When mentioning the vendor bill, Jane Fraser explicitly said this cost line "has a long way to go"—implying the room for cuts is far from exhausted.

AI Coding Agent differs fundamentally from traditional code-completion tools (such as GitHub Copilot). Code-completion tools predict the next line or block of code based on context, with human developers always leading task decomposition, architecture design, and debugging. An AI coding agent, given a high-level goal, can autonomously plan steps, invoke tools (such as terminals, browsers, and code editors), execute multiple rounds of modification, and verify results. Devin is a representative product from Cognition, positioned as an "autonomous software engineer" that can complete tasks end-to-end from requirement understanding to code submission. This leap in capability allows enterprises to convert some standardized development work from outsourcing contracts to internal automated execution, but it also imposes new requirements on code review, security auditing, and engineering management.

Where the 30% Efficiency Gain Comes From: Accelerating 100 Refactoring Processes

The flip side of lower costs is improved delivery efficiency. According to the same source, Citigroup's release speed for roughly 100 product refactoring processes improved by 30% to 40%. This figure deserves unpacking: faster product launches mean a shorter cycle from requirement confirmation to deployment, and refactoring processes often involve large amounts of repetitive code migration and test adaptation—exactly the scenarios AI coding agents excel at.

Note that the "30-40%" here is limited to about 100 refactoring processes, not all of Citigroup's product lines. It shows that AI coding tools can already deliver quantifiable efficiency gains on structured tasks with clear boundaries, but there is still a gap before they cover the entire R&D system.

Product refactoring processes typically refer to structural adjustments to existing systems without changing their external behavior—for example, splitting a monolith into microservices, upgrading framework versions, migrating databases, or clearing technical debt. Such tasks are characterized by clear goals, well-defined boundaries, and high repetitiveness, but they involve large amounts of mechanical code changes and test adaptation, traditionally requiring substantial manpower. AI coding agents happen to excel at these "structured tasks with clear acceptance criteria," and can therefore significantly compress cycles in refactoring scenarios. However, faster refactoring does not directly equal faster new-feature development—the latter requires more business judgment and cross-team collaboration, and the autonomy of AI agents remains constrained by requirement ambiguity and the complexity of architectural decisions.

Key Caveat: 30% Is Just "One Cost Line"

The easiest way to misread the Citigroup case is the scope of the 30%. The original information explicitly states that this 30% cut applies only to the technology vendor bill, not to larger technology budget categories such as servers, cloud services, or internal labor.

This caveat does not diminish the case's value—vendor bills are often the most elastic and most AI-replaceable part of a bank's technology spending—but it reminds us that equating "30% cost reduction" directly with "a 30% cut in the technology budget" would seriously overstate AI's short-term financial impact. Jane Fraser's own wording was restrained: she emphasized that this cost line still has room for optimization, not that AI has already reshaped the overall cost structure.

The Other Side in the Same Week: JPMorgan Says AI Gains Are "Not Yet Visible"

Citigroup's high-profile disclosure contrasts sharply with JPMorgan's cautious stance. According to the same source, JPMorgan's chief financial officer said in the same week that AI-driven gains have not yet shown up in earnings. Both banks are investing heavily in AI, but they are sending very different signals about the visibility of financial returns.

This divergence may stem from differences in deployment stage and measurement approach: Citigroup's 30% is a reduction in a single cost line and is easy to quantify, while JPMorgan may be more focused on AI's contribution to overall revenue and margins, metrics that require a longer observation period. Another possibility is that the AI use cases different banks choose vary in maturity—coding agents are one of the most mature categories of enterprise AI adoption today, while other scenarios have longer return chains.

Earnings in corporate management typically refers to final financial metrics such as net profit and earnings per share, not single-item cost savings. For AI investment gains to show up in financial statements, they must go through a complete chain from tool procurement and deployment to process transformation and scaled replication, and often first appear as increased expenses (such as compute procurement and employee training) before gradually manifesting as lower costs or higher revenue. JPMorgan's CFO's remarks mean that, under its measurement approach, AI's positive financial impact is not yet large enough to offset investment and be identified at the consolidated reporting level. This is not contradictory to Citigroup's disclosed single vendor-bill reduction—the latter is an optimization of a local cost item, while the former concerns the visibility of overall financial returns.

Implications for Enterprises: Which Vendor Bill to Cut First

The actionable clue from the Citigroup case is this: AI coding agents hit standardized, outsourceable development service procurement first. If an enterprise wants to replicate this path, it should prioritize development contracts with clear requirements, high repetitiveness, and vendor billing by person-day or project.

But before replicating, two questions need answers. First, does the organization have the engineering capability to manage AI agent output—Devin can write code, but code review, architectural decisions, and quality control still need human engineers. Second, is the cost reduction coming at the expense of system stability or technical debt—if a 30% bill reduction is accompanied by accumulating hidden risks in refactoring processes, long-term costs could rise instead.

Citigroup's scale of 14,000 engineers shows this is not a small pilot but an enterprise-wide tool replacement. JPMorgan's caution reminds us that between point cost savings and overall financial returns, there is still a road that needs validation. Taken together, the two banks' statements form the most realistic cross-section of current enterprise AI adoption: the tools can already settle one account, but not the whole account.

One-line Summary

Citigroup used Devin to cut its vendor bill by 30%, while JPMorgan says AI gains have yet to appear in earnings—enterprise AI adoption is diverging.

Paragraph Summary

Citigroup cut its technology vendor bill by 30% within a year, driven by 14,000 engineers using the AI coding agent Devin, with release speed for about 100 product refactoring processes improving by 30% to 40%. But this reduction applies only to a single cost line, not an overall technology budget cut. In the same week, JPMorgan's CFO said AI gains have not yet shown up in financial results, a sharp contrast with Citigroup. The two financial giants' stances on AI coding tools reflect the real divergence in enterprise AI adoption: AI coding agents can already settle one account on standardized, outsourceable development tasks, but there is still a validation gap between point cost savings and overall financial returns, and human review and quality control remain necessary.