EY achieved 4x to 5x productivity gains in software development by connecting AI coding agents to its internal engineering standards, compliance frameworks, and code repositories, according to Stephen Newman, EY’s Global CTO Engineering Leader.
The gains apply across teams building EY’s audit, tax, and financial platforms. But Newman is clear that flipping a switch on a new tool did not produce them. The company spent 18 to 24 months on cultural groundwork and technical integration before semi-autonomous coding worked reliably at scale.
The Core Problem With Coding Agents
Coding agents can produce thousands of lines of code in minutes. The issue is deployability. “You can generate a ton of code, but it doesn’t mean really anything, right? It’s got to be code that is integratable, that is compliant, and you don’t want to create more work on the back end just because you sped up the code generation process on the front end,” Newman said.
Without access to EY’s internal repositories, engineering standards, and source catalogs, agents generated generic output requiring extensive rework. Newman calls this connected knowledge base the “context universe.” Agents operating outside it simply add friction rather than remove it.
Adoption Built From the Ground Up
EY started with GitHub Copilot-style tools, giving engineers time to get comfortable with prompt engineering and assistive AI. Newman’s guiding principle was organic adoption. “It’s important to bring AI capabilities as a ground-up organic adoption rather than force them onto the users,” he said.
Productivity gains plateaued until EY built deeper integrations. Once agents had access to the context universe, the ceiling lifted.
Choosing an Agent Platform
EY evaluated three platforms: Lovable, Replit, and Factory‘s IDE-based Droids. Rather than mandate a single tool, Newman’s team tracked adoption, usage, and productivity across all three. Developers gravitated toward Factory organically, which the team treated as the clearest signal of real value.
Factory adoption “took off like wildfire” once it moved from evaluation to pilot. EY had to throttle traffic and restrict which repositories could connect before security and compliance teams signed off.
What Agents Handle and What They Don’t
Developer enthusiasm forced EY to build a workload classification framework. Not every task belongs to an agent.
High-autonomy tasks agents now handle independently:
- Code review
- Documentation
- Defect fixing
- Greenfield features
Tasks that still require human oversight:
- Large-scale refactors
- Architecture decisions
- Cross-system integrations
Developer roles shifted alongside this framework. Engineers moved from writing code themselves to acting as orchestrators, directing agents toward the correct databases and repositories.
Measured Outcomes
During the early adoption phase, with security guardrails in place and repository integrations complete, EY measured efficiency gains ranging from 15% to 60% depending on the developer persona. The 4x to 5x headline figure reflects the more mature stage Newman describes as “horizon model development,” characterized by semi-autonomous agent execution at scale with teams of orchestrators rather than individual coders.
Newman acknowledged the gains cannot be attributed solely to the agents themselves. Trial and error, combined with behavioral shifts across developer teams, produced the result together.
Photo by Nguyen Dang Hoang Nhu on Unsplash
This article is a curated summary based on third-party sources. Source: Read the original article
