Salesforce's AI Strategy: Agentic Engineering Enablement

Enterprise AI Enablement: Salesforce's Agentic Strategy

Adopting agentic tooling at an enterprise scale for thousands of software engineers presents a unique challenge. The primary hurdles are not inherent model capabilities, but rather an organization's capacity to cultivate new skills, establish common expectations, and define a unified language for human-agent collaboration.

Left to their own devices, engineers will experiment and develop individual workflows. While this can surface immediate productivity gains, the absence of a shared vision and explicit leadership guidance can lead to fragmented practices and an inability to translate individual advancements into organizational-wide benefits. Salesforce encountered this challenge directly, recognizing that the core issue was not engineer adoption, but the development of a shared lexicon to amplify individual productivity gains across the entire engineering organization.

The Technology, People, Innovation, and Learning (TPIL) team at Salesforce addressed this by focusing on enabling the entire engineering organization, not just early adopters, to learn and grow collaboratively with agentic tools.

Agentic Transformation as an Organizational Learning Challenge

When powerful new capabilities like agentic tooling are introduced, software engineers naturally integrate them into their existing workflows. This might involve developing advanced prompting strategies, focusing on AI-assisted testing, debugging, or documentation. These individual approaches are valid explorations of new technology. However, TPIL observed that the organization's struggle wasn't a lack of curiosity, but the absence of a shared understanding of effective agentic coding practices. Many engineers grappled with questions such as:

• What does it truly mean to become AI-native?
• Which behaviors are most critical for success?
• How can progress be meaningfully measured?

TPIL established a fundamental principle: engineers learn in diverse ways. To foster sustainable organizational change, enablement efforts must meet engineers where they are, building confidence through various learning modalities. This led to the understanding that the challenge was primarily one of organizational learning, not merely a technological one.

Convergence of Insights Guides the Journey

Framing agentic transformation as an enablement challenge prompted a deeper question: what does the journey to becoming AI-native actually entail? As it turned out, different teams within Salesforce were independently developing models to answer this. One team proposed a five-level model, distinguishing orchestration from looping. Another utilized a 2x2 matrix centered on telemetry. Despite their differing approaches, these models consistently revealed a shared underlying journey characterized by similar breakthroughs, frustrations, and periods of uncertainty. Engineers across various products and technologies navigated comparable developmental stages.

This convergence of insights, combined with TPIL's learning strategy, led to the development of the Proficiency Level (PL) Framework. This framework was deliberately grounded in mindset and behavior, rather than specific tools or dashboards. The intent was to create a durable model that wouldn't become obsolete with tool updates. The PL Framework provides a shared language for thinking and behaving differently with agentic tools—understanding how to delegate, verify, and collaborate with agents as a core capability, transferable across different platforms.

Crucially, the PL Framework established a shared definition of "good," offering a common understanding of effective agentic work at each stage, replacing numerous tool-specific interpretations.

Progression Over Proficiency

While the PL Framework provided a definition of success, the greater challenge was facilitating engineer progression through its levels. A significant mindset shift involved validation. Engineers quickly learn that generating code with AI is simpler than assessing its quality. Outputs can be erroneous, requirements misinterpreted, and subtle errors can lead to unreliable software. Many agentic adoption initiatives falter at this point, highlighting the need to learn how to effectively communicate intent, evaluate AI outputs, and construct reliable engineering workflows. The true transformation lies not just in learning to write code with agents, but in fundamentally rethinking the software development process itself.

As engineers advance, they move from using agents for discrete tasks to orchestrating complex agent systems. Each stage necessitates a new mental model, reinforcing that becoming AI-native requires continuous capability development, not just the acquisition of isolated technical skills. For engineering leaders, this implies that success is measured by engineers' consistent development of judgment, confidence, and capabilities to reach the next stage, rather than the speed at which they achieve a final level.

Enabling Progression at Scale

A shared definition of "good" provided direction, but practical enablement was necessary to drive progression. TPIL developed hands-on programs designed to make progression an experiential process. "AI camps" offered dedicated time for engineers to practice and experiment. Weekly sessions met engineers at their current skill levels, adapting to evolving questions. Coaching guides equipped managers with a common language for one-on-one discussions, shifting the focus from "Are you good at AI?" to "What is your next stage of growth?" This empowered managers to actively support engineer progression.

The combination of a framework defining the destination and programs facilitating the journey transformed the PL Framework from a theoretical model into a tangible enablement practice. The four proficiency levels illustrate this practice in action:

Four Proficiency Levels: A Developmental Map

While the "Agentic Coding Maturity Curve" identified nine stages, TPIL condensed these into four manageable levels for practical application. Change management research suggests that focusing on four to five steps at a time is optimal for individual action. The PL Framework offers a clear, progressive path with distinct stages.

Each level focuses on shifting skills, mindsets, or behaviors, rather than dictating specific technical steps. The four proficiency levels were designed for visibility of growth, not for performance evaluation. They function as a shared language for describing evolving thinking and workflows, fostering collaborative discussions between engineers and managers about growth.

• AI-Assisted: Engineers shift from writing all code manually to treating AI as a primary collaborator in daily development.
• AI-Validating: Engineers move from blindly trusting AI output to actively validating it, understanding that effective AI utilization hinges on verification.
• AI-Orchestrating: Engineers transition from directing individual AI tasks to designing systems where multiple AI agents collaborate towards a larger objective.
• AI-Native: Engineers shift from optimizing their personal workflows to codifying them as organizational standards that accelerate the progress of the entire team.

At each level, the emphasis is on observable behavior change, not solely on acquired knowledge.

Measuring Enterprise AI Adoption Through Behavior Change

Instead of tracking attendance or course completion, TPIL focused on evidence of actual behavior shifts across the engineering organization. While numerous engineers participated in AI camps globally, adopting AI tools rapidly, volume and speed alone did not confirm behavior change. These initiatives, however, provided a sufficient population for monitoring such shifts.

The most telling signal emerged from the evolution of engineers' questions. As participants progressed beyond foundational topics, weekly session attendance naturally decreased. The nature of their inquiries shifted from optimizing AI outputs to identifying areas where human judgment remained paramount. This evolution indicated that engineers were not just increasing their AI usage, but were fundamentally rethinking their approach to work.

Four Principles for Building AI-Native Engineering Organizations

Regardless of the specific AI models or tools employed, every engineering organization pursuing enterprise AI adoption will encounter similar challenges. Salesforce's experience offers four guiding principles:

1. Scale Shared Expectations: Agentic practices can only scale when common expectations are established. Software engineers naturally develop diverse workflows when exploring new technologies.
2. Learning Journeys Outperform Assessment Frameworks: Proficiency is a progression, not a certification. Salesforce prioritized guiding engineers from their current state toward future growth.
3. Behavior Change is the Metric That Matters: Adoption is defined by evolving workflows, coaching interactions, and the complexity of engineer inquiries, not mere attendance.
4. The Learning Culture Matters More Than Any Single Framework: Tools and frameworks are ephemeral. Long-term success depends on an engineering organization's commitment to continuous learning through disruptive changes.

Building Engineering Organizations That Grow with AI

As access to increasingly powerful AI models becomes ubiquitous, not all organizations will achieve AI-native status. The differentiator lies not in the technology itself, but in an organization's effectiveness at enabling its engineers to learn, adapt, and grow collaboratively.

Key Takeaways

• Successful enterprise AI enablement hinges on fostering a shared language and common expectations for human-agent collaboration, not just tool adoption.
• The Salesforce TPIL team developed the Proficiency Level (PL) Framework, focusing on mindset and behavior shifts rather than specific tools, to guide engineers towards AI-native practices.
• The PL Framework categorizes progression into four levels: AI-Assisted, AI-Validating, AI-Orchestrating, and AI-Native.
• Empirical measurement of AI adoption should focus on observable behavior change, such as evolving questions and workflow adaptations, rather than attendance or tool usage metrics.
• Key principles for building AI-native engineering organizations include scaling shared expectations, prioritizing learning journeys over assessments, measuring behavior change, and cultivating a strong learning culture.