Salesforce open model AI strategy aims for trust, automation

## Overview

This episode of the Targeting AI podcast features Jes Gohland Rajan, Senior Vice President of Salesforce AI, discussing Salesforce's generative AI strategy, its open model ecosystem approach, and the Einstein Trust Layer. The conversation covers the shift from 2023's experimentation phase—dominated by ChatGPT exploration—to 2024's focus on real enterprise implementations of generative AI with measurable business value.

## Key Topics

**2023-2024 Market Shift** 2023 saw extensive prototyping and proof-of-concepts as enterprises experimented with generative AI. Rajan notes the focus has shifted to identifying and implementing the highest-value use cases: making service personnel more efficient, enabling salespeople to be significantly more productive, and helping marketing managers create better-converting campaigns. Implementation now prioritizes safety, security, and trust alongside measuring actual business value.

**Generative AI Applications** Rajan identifies two primary categories. Content creation and conversational automation address scenarios like automating customer service responses and enabling marketing professionals to generate multilingual campaigns instantly. Co-pilots represent the second category—assistants that orchestrate actions on behalf of users, functioning like hiring a knowledgeable intern that can execute tasks with appropriate guardrails.

**Open Model Ecosystem Strategy** Salesforce rejects single-model dependency, advocating instead for an open ecosystem. Rajan notes that Salesforce's team occupies the top three positions on Hugging Face's LLM leaderboard. The company works with multiple providers (OpenAI, Anthropic) while building proprietary models like xgen for code generation. This approach enables task-specific fine-tuning at a fraction of commercial model costs, allowing customers flexibility to bring their own models or select from Salesforce's offerings.

**Einstein Trust Layer** Launched approximately one year prior, the trust layer operates across Salesforce's AI stack, addressing both data protection and result accuracy. Components include secure data retrieval, dynamic grounding for customer data integration, sensitive data masking, prompt guardrails, zero data retention policies with third-party providers, toxicity detection, and comprehensive audit trails. Critically, the layer remains model-agnostic, functioning consistently across the open ecosystem through an LLM Gateway abstraction.

**Responsible AI and Bias Mitigation** Salesforce deploys detection models tailored to CRM use cases and company ethical guidelines to identify and prevent biased, harmful, or inappropriate content. Rajan acknowledges this challenge intensifies with an open model ecosystem and identifies bias mitigation as an active research area. The company develops and shares open-source toxicity models with the community.

**Predictive and Generative AI Integration** The discussion illustrates convergence of both approaches. An account executive using a co-pilot might request a sales forecast (predictive), then use that output to generate an email (generative), with interactions feeling natural despite underlying model complexity.

**Investment and Acquisition Strategy** Salesforce Ventures announced a $500 million generative AI fund targeting core technology providers and ecosystem companies. Acquisitions focus on specific portfolio gaps and talent acquisition needs. The company emphasizes ecosystem partnership alongside internal development.

**Future Evolution** Salesforce plans continued development of human-centric assistance and "in-the-flow" generative capabilities that enhance existing workflows without changing interfaces. The traditional "AI Cloud" branding transitions to "Einstein One Platform," integrating AI, data, and CRM. Significant focus on developer reskilling ensures the ecosystem can build and maintain co-pilot applications.

## Takeaways

- Enterprise generative AI adoption is moving from exploration to implementation, with focus on measurable business value and risk mitigation. - Open model ecosystems enable enterprises to optimize tool selection for specific tasks rather than relying on single providers. - Trust requires both data security/privacy protections and accuracy/hallucination controls; a dedicated trust layer becomes essential when orchestrating user actions. - Responsible AI—particularly bias detection—remains an ongoing challenge, especially across diverse model providers. - Talent shortages in generative AI development persist; reskilling existing developer bases is critical for ecosystem scaling.