Tech • AI • Robotics • Game
How Salesforce Uses Claude Code
8/10TL;DR
Salesforce says Claude Code helped it rapidly modernize legacy software, double effective engineering output, and expand autonomous agent use across thousands of developers without a corresponding rise in bug rates.
KEY POINTS
Legacy systems were made headless in months
A major breakthrough was converting decades-old Salesforce applications to work in a more headless, API-driven way. Executives said a project that would normally have taken six months to a year was completed in weeks to a few months, pushing API coverage across the company’s clouds to roughly 80% to 90%. The effort was notable because much of the codebase predated modern API design and required change across long-established products.
Dreamforce launches were shaped by AI-assisted development
The company said many of the new applications and user experiences shown at Dreamforce were built with some level of Claude involvement. Teams reported shipping work in a few weeks that previously would have taken months, turning AI from a coding assistant into a force multiplier for product delivery. The gains were described not as isolated experiments but as visible in production software.
ROI is being measured with “effective output”
Rather than counting lines of code, Salesforce worked with Stanford researchers on a metric called effective output, intended to score whether AI-generated code makes meaningful logic changes or creates new functionality. Low-value additions such as comments score poorly, while substantial functional work scores higher. The company then tracks dollar per effective output to judge whether AI use is materially improving productivity, and said that measure has risen by about 2x.
Engineers were pushed from basic prompting to agentic workflows
After an initial phase of ad hoc use, the company introduced an agent maturity curve to move engineers beyond simple prompts into more structured, autonomous workflows. In one to two months, a large share of engineers moved from basic usage to more advanced levels focused on planning, decomposition, orchestration, and parallel execution. The result, executives said, was a clear 2x increase in output without a matching decline in quality.
Autonomous agents now handle parts of the software lifecycle
Thousands of employees are now using agents in a more autonomous mode across the SDLC, including responding to failed tests in CI/CD, fixing issues, and triggering code review and merge steps. The approach shifts engineers away from manually breaking every task into tiny units. Instead, larger feature requests are handed to models or orchestration layers that decompose work into smaller parallel tasks.
Incident response has become a major proving ground
AI agents are also being used outside feature development, especially in incident management. In one example, an agent reviewed a Slack incident thread, identified that an inner service rather than the paged outer service was causing the problem, named the responsible team, and escalated directly to the right contact. The reported benefit was avoiding the common cascade in which multiple teams are awakened and manually triage ownership.
Knowledge sharing evolved into a marketplace for reusable skills
Bottom-up experimentation was formalized into an internal marketplace for agent “skills,” with evaluation tools to test whether shared workflows actually improve outcomes. The system is tied into the engineering harness and supports reuse across thousands of repositories, helping developers work in unfamiliar codebases and still produce quality pull requests. Leaders described the model as encouraging local innovation first, then scaling what proves effective.
Quality controls are central to the rollout
The company’s core guardrail is validation. Teams define objective success and failure criteria, ask agents to show evidence for both, and measure output against labeled ticket and pull-request data already embedded in standard engineering workflows. Despite nearly doubling code volume in recent months, leaders said bug rates did not scale proportionally and were lower on average relative to the amount of code generated.
Model choice is treated as a cost-performance decision
One of the biggest lessons was to match the model to the task. Simpler jobs do not justify expensive large-model use, while long-horizon debugging sometimes does. In one case, a difficult bug in a critical business product resisted repeated human investigation; a large model was allowed to debug, compile, and iterate for about 24 hours before finding a root cause and proposing a fix, which was then subjected to strict validation because an incorrect change could affect customer financial operations.
CONCLUSION
Salesforce is treating AI coding tools less as a convenience feature than as a new operating model for software development and support. The broader test now is whether agent-driven speed, autonomy, and scale can keep delivering measurable gains without weakening reliability.
Ask a question