Citigroup's AI Coding Agent Devon: Saving 30% on Tech Outsourcing Annually, Already Used by 14,000 Engineers

Citigroup conducted a bold experiment in technology spending: introducing the AI coding agent Devon into its internal development workflow. Within a year, the bank cut its technology vendor bills by 30% and accelerated product releases by 30%–40%. This is not just a cost-control story—it may reshape how banks work with IT outsourcers.

What Is Devon, and How Does It Replace Traditional Outsourcing?

Devon is an AI coding agent capable of autonomously planning tasks and writing code. According to Citigroup, 14,000 engineers now use Devon in their daily work. Previously, these engineers had to pay external vendors for code or development services; now Devon handles that portion of the work directly.

An industry insider noted that Devon's uniqueness lies in its ability not only to generate code snippets but also to understand development goals and formulate implementation steps—precisely the high-value work that outsourcing teams used to take on. Citigroup CEO Jane Fraser said at an internal meeting that the reduction in vendor bills "has a long way to run," hinting that there is still considerable room for cost optimization in the future.

This shift directly impacts the traditional IT outsourcing industry. In the past, banks relied on vendors for a full range of services from requirements analysis to coding; today, AI agents handle most standardized work internally, forcing vendors to shift toward higher-end consulting or customized development.

AI coding agent is fundamentally different from common code completion tools (such as GitHub Copilot). Code completion tools typically predict the next line or block of code based on context, functioning as "input assistance"; an AI coding agent, by contrast, is given a high-level goal (e.g., "implement a certain API endpoint") and can break down tasks itself, search the codebase, write multiple files, run tests, and iterate based on error reports. This closed loop of "plan—execute—verify" enables it to take on entire blocks of development work previously handled by junior outsourcing teams. Current mainstream agentic coding tools are mostly built on large language models, using tool use to access the file system, terminal, and version control, so their efficiency advantage is most pronounced on standardized, well-bounded tasks, while scenarios with vague requirements or strong business coupling still require human oversight.

Product Releases Accelerated 30%–40%, Covering About 100 Transformation Processes

Beyond cost savings, the efficiency gains from Devon are equally significant. Citigroup deployed Devon across about 100 product transformation processes, shortening product release cycles by 30%–40%. For example, some coding tasks that previously took weeks are now delivered within days.

Fraser emphasized that these efficiency gains do not come at the expense of quality. She believes AI agents free engineers from repetitive labor so they can focus on architecture design and complex problem-solving. However, some observers caution that AI-generated code still requires rigorous testing and review, especially under the financial industry's high compliance requirements.

Compliance and code review requirements in the financial industry are essential context when evaluating AI-generated code. Bank systems are bound by Basel accords, national regulators (such as the U.S. OCC and Federal Reserve), and internal model risk management frameworks; any code entering production must undergo static analysis, security scanning, peer review, and change approval. Although AI agents generate code quickly, provenance traceability, license compliance, and potential security vulnerabilities all require additional auditing. Therefore, Citigroup's claim of "shortening release cycles by 30%–40%" actually compresses the coding and rework phases while leaving existing review processes unchanged—rather than skipping quality control. This also explains why the efficiency gain (30%–40%) is similar to the cost reduction (30%): both are constrained by the fixed overhead of review and integration.

Potential Impact on the IT Outsourcing Industry

Citigroup's case could trigger a chain reaction. Will other banks follow suit? There is no public data yet, but Citigroup's 30% reduction is enough to make peers reassess their outsourcing strategies. One analyst commented: "If a top-tier bank can save 30% on vendor costs through AI agents, other banks have no reason not to follow."

However, AI agents are not a panacea. Devon mainly handles standardized, repetitive coding tasks; for projects requiring deep business understanding or innovative design, human engineers and external experts remain indispensable. Citigroup's remark that the vendor bill "has a long way to run" may also imply that the remaining portion is precisely where AI is hard to replace.

The value chain stratification of IT outsourcing helps clarify where the impact lands. Traditional IT outsourcing is usually divided into three layers: the bottom layer is standardized coding and testing (labor-intensive, low unit price), the middle layer is system integration and project management, and the top layer is business consulting and architecture design. AI coding agents first erode bottom-layer coding work because such tasks have clear goals, are verifiable, and their knowledge can be codified. The impact on the middle layer depends on how deeply agents are integrated with existing DevOps pipelines, while the top-layer consulting business may actually gain new demand as enterprises need to redesign AI governance frameworks. This is what analysts mean by "vendors shifting toward high-end": outsourcers that still sell coding hours on a headcount basis will face pricing pressure; those that shift to outcome-based delivery and domain expertise may maintain or even improve margins.

Future Outlook: Collaboration Models Between AI Agents and Human Engineers

Citigroup's practice provides a template for AI application in fintech. With 14,000 engineers using AI agents, human-machine collaboration has become routine. The future challenge lies in balancing automation and innovation, and in managing the intellectual property and security risks of AI-generated code.

For other enterprises, Citigroup's experience shows that AI agents can be an effective tool for cutting IT spending—provided a robust governance framework and testing process are established. As tools like Devon mature, the relationship between banks and vendors will shift from "outsourcing" to "partnership," and vendors will need to offer value-added services that AI cannot replace in order to survive.

One-line Summary

Citigroup used the AI coding agent Devon to cut vendor bills by 30% and accelerate product releases, disrupting the IT outsourcing model.

Paragraph Summary

Citigroup introduced the AI coding agent Devon into its internal development workflow, with 14,000 engineers using it daily. Within a year, technology vendor bills were cut by 30%, and release cycles for about 100 product transformation processes were shortened by 30%–40%. Devon can autonomously plan tasks and write code, replacing some standardized development work previously handled by outsourcers. CEO Jane Fraser says there is still room to reduce vendor costs, signaling that AI substitution will continue to deepen. The article argues that this case may prompt other banks to follow suit, pushing the relationship between banks and IT vendors from "outsourcing" to "partnership." Vendors will need to provide value-added services that AI cannot easily replace in order to survive, while testing, compliance, and IP risks of AI-generated code still require governance frameworks to constrain.