How We Got Our Engineers Using AI — Without Breaking Everything

At Salesforce, our developers work with AI tools like Cursor, Windsurf, Claude Code, and our own CodeGenie daily - across every part of the product lifecycle.

But reaching enterprise-scale AI adoption wasn't as straightforward as buying the tools and telling developers to use them. It required rethinking our approach and building new infrastructure - new measurement, standardization, governance frameworks, and workflows that could handle thousands of engineers using dozens of AI tools simultaneously.

Here's how we built that foundation, how it's shaping the future of productivity for our engineering organization, and what we learned along the way.

1. Your old metrics won't tell you if AI is actually helping

Traditional engineering metrics - lines of code, commit frequency, cycle time - don't capture AI's real impact. Worse, they can mislead you into thinking AI is creating efficiency when it's actually creating new challenges.

Even with solid governance infrastructure in place, we discovered that AI adoption created unexpected downstream effects.

As it does at many companies with global operations, our engineering data had become siloed: It was scattered across security systems, production logs, code repositories, CI/CD pipelines, and release tracking. We needed a framework with standardized metrics across all engineering teams and a unified view to understand what was actually happening.

So we built Engineering 360, a platform that centralizes engineering data from hundreds of systems to track security, availability, quality, and developer productivity.

This sets the foundation for the metrics we really need to actually tell us whether AI is making engineering better, not just faster. The new ways we are exploring to measure outcomes include effective output, low churn, and highly maintainable code.

While developing the right metrics can help you evaluate the value of new AI tools, you also need a strong strategy to ensure those tools are used securely and reliably.

2. Governance at scale means infrastructure, not policies

Manual oversight works fine for small pilots. It breaks down when you have all of your engineers using a variety of AI tools.

To standardize how AI systems connect with tools and data, we adopted the open-source framework Model Context Protocol (MCP). Then we built an MCP gateway that works like an API gateway for AI agents : registering tools, enforcing permissions, maintaining audit trails, setting guardrails, and monitoring usage.

But developers still need flexibility. Our internal AgentExchange provides a curated marketplace of vetted AI configurations while establishing consistent governance across teams.

The result: Developers can choose the right AI tool for each task while we maintain enterprise-grade security. Choice within guardrails, not blanket restriction.

Even with solid governance infrastructure in place, we discovered that AI adoption created unexpected downstream effects.

3. More AI-generated code meant rethinking our review process

Here's what no one warned us about: AI tools made code review harder, not easier.

AI generates more code more frequently. Reviewers were drowning in volume while trying to spot AI-specific issues - subtle bugs, security vulnerabilities, or performance problems that require different review approaches than human-written code.

We built PRizm, an AI-powered code review tool that provides logical breakdowns of changes and intelligent recommendations based on our quality standards. It handles the first-pass analysis so human reviewers can focus on architecture and complex logic.

The upshot: AI doesn't eliminate human oversight - it changes what that oversight looks like.

Solving the code review challenge taught us an important lesson about AI adoption: Success comes from embedding AI in existing workflows, not creating new stand-alone tools.

4. Meet developers where they already work

We developed the Agentforce Engineering Agent to work alongside engineers right in Slack - where they already work and collaborate. Engineering Agent is a digital teammate focused on automating noncoding tasks like planning, modeling, and investigating bugs or resolving incidents throughout the development cycle.

While embedding AI in familiar tools drove adoption, we also realized that AI agents drive value when operating autonomously - in the background - without disrupting developer workflows.

Now available in 1,000+ Slack channels, Engineering Agent handles routine tasks, provides insights, manages workflows, and searches our documentation and incident history. It's one of our top three most-used agents across Salesforce.

Why it worked: Adoption happens when AI fits into existing workflows, not when you force new ones. While embedding AI in familiar tools drove adoption, we also realized that AI agents drive value when operating autonomously - in the background - without disrupting developer workflows.

5. Deploy autonomous agents for mechanical tasks

Not every AI improvement needs to happen in a developer's workflow. Autonomous agents can handle mechanical tasks like code coverage, freeing humans for higher-order work.

Our autonomous code coverage agent works independently - analyzing gaps, generating tests, validating quality, and submitting pull requests for human review. Developers get the benefits without workflow disruption.

The principle: Use AI to eliminate routine work, not replace strategic thinking.

What we learned

Reaching full AI adoption required both a top-down mandate and grassroots empowerment. We established AI coding tools as the "new engineering baseline" while building community-driven initiatives like AI Productivity Thought Lucks, where engineers share use cases and best practices.

For other engineering leaders tackling similar challenges: Build even as you deploy new tools. Measure holistic outcomes, not just activity. And expect your development practices to evolve as AI becomes part of how teams work every day.

We're still learning, still adapting. But one thing is clear: AI infrastructure isn't optional anymore. It's how modern engineering organizations stay competitive.

Go deeper: