Salesforce AI CEO Clara Shih discusses how Salesforce is rapidly integrating generative AI across its product suite, how early adopters like Gucci are transforming customer service into a revenue center, and what it takes to scale AI in the enterprise.
Salesforce has moved faster than most incumbents in shipping AI products, launching Einstein GPT applications in summer 2023 and going general availability (GA) in early 2024 with Einstein Copilot (a natural language assistant grounded in customer data) and Copilot Studio (a prompt builder and model builder for customization).
The company restructured its AI teams: initially decentralized across product clouds, it created a shared AI platform team to build core infrastructure like the Einstein Trust Layer, model gateway, and prompt builder, while individual clouds (sales, service, Slack) build domain-specific AI actions on top of that platform.
Gucci: From Cost Center to Revenue Center
Gucci was an early adopter of Salesforce’s AI, visiting Salesforce’s chief scientist in November 2021 to see early large language model demos, and became a co-creation partner for Service GPT.
Service representatives solving a customer issue (e.g., a broken belt buckle) are now augmented in real time by AI that notices the customer has been browsing handbags online and coaches the rep to make a sales recommendation.
This has reduced average handle time and increased conversion rates by double digits, turning a traditional cost center into a revenue center, and empowering service reps to become brand storytellers and salespeople.
Engineering Trust into AI Products
Salesforce addresses trust at three levels:
Technology: The Einstein Trust Layer includes data masking, grounding with Data Cloud to reduce hallucinations, citations, audit trails, prompt defense, and zero data retention.
Policy: AI bots must self-identify as AI when interacting with consumers.
Stakeholder engagement: Published open-source guiding principles around accuracy, honesty, and empowerment, shared with regulators and the industry.
Sensitive data fields (name, gender, zip code) are proactively surfaced for masking to prevent bias in both predictive and generative AI models.
Salesforce’s Unique Data Advantages
Salesforce identifies four types of data that make it uniquely powerful for AI:
Structured CRM data: The traditional customer records and sales pipeline data.
Unstructured data: Knowledge articles, Slack conversation transcripts, contact center voice calls, chats, and emails, now supported with vector search and hybrid reranking in Data Cloud.
Metadata layer: Originally built for multi-tenancy 25 years ago, it now gives AI critical context about which data objects, tables, and functions to use.
Feedback/outcome data: Salesforce captures whether opportunities advanced, deals closed, or marketing campaigns converted—essentially a reward function for AI models, whether predictive or generative.
Salesforce also has zero-ETL partnerships with major data lake providers (BigQuery, Redshift, Snowflake, Databricks) to bring in enterprise-wide data.
Model Strategy: Open Architecture and Vertical Benchmarks
Salesforce takes an open architecture approach: customers can use models hosted on Salesforce’s service, bring their own models, or use third-party models through Model Builder.
Salesforce’s own research models are fine-tuned for specific domains like code generation, Salesforce flow generation, financial services sales, and high-tech sales.
The company is building vertical-specific evaluation benchmarks (e.g., pharma sales, wealth management) to determine which model performs best for each task on cost, performance, and latency.
Trends among customers: existing Google Cloud customers tend to prefer Gemini; OpenAI’s GPT-4 remains best in class for complex agentic planning and plan generation.
Barriers to Enterprise AI Adoption
Clara identifies the biggest barriers as:
Trust: Data security, privacy, and honoring internal data access rules and entitlements across departments.
Business case: Productivity gains and margin expansion must clearly outweigh costs, especially when scaling beyond pilots to entire contact centers.
Turnkey AI use cases (Service GPT, Sales GPT, Commerce GPT) that take five minutes to set up have been critical for demonstrating immediate business value before customers customize with Copilot Studio.
User education involves ethnographic research—sitting in call centers to understand employee fears about job replacement—and designing UX and onboarding that makes clear AI is a co-pilot, not an autopilot.
AI Costs and Scaling Considerations
Costs are captured in product SKUs and passed to customers; they are less of a concern during pilot phases but become significant when scaling.
Salesforce is pursuing a multi-model, multi-size architecture that routes tasks to appropriately sized models to optimize cost.
If inference were nearly free, the company would test everything on every model simultaneously without needing cost optimization strategies.
The Future of Slack as an AI Interface
Slack AI already offers conversation and channel summaries; the next step is bringing Einstein Copilot into Slack so account teams can summarize all customer data (support tickets, marketing engagement, web activity) before meetings.
Clara sees Slack as a natural interface for “team plus AI” workflows—where groups of people collaborate alongside an AI co-pilot that can chime in with suggestions visible to everyone.
This mirrors emerging consumer trends (e.g., Character.AI group chats) and could extend to coding teams tagging LLMs for input or sales teams getting AI suggestions on how to close deals.
Customer Support Automation
The percentage of support questions AI can answer varies by industry, but the bigger opportunity is workflow orchestration—AI can initiate actions like processing returns and sending shipping labels, not just answer questions.
Governance is critical: admins must explicitly authorize which flows and Apex code the copilot can access, starting with low-risk lookups before enabling higher-stakes actions like issuing refunds.
Incumbents vs. Startups in AI
Clara sees room for both incumbents and startups to create enormous value, drawing parallels to previous technology waves (cloud, mobile) where both new winners emerged and incumbents reinforced their positions.
Salesforce Ventures is investing in AI startups it believes will become multi-billion or even trillion-dollar companies, and encourages startups to build on the Salesforce platform to leverage its data cloud and UI.
Personal Reflections and What’s Ahead
Clara founded Hearsay Systems (a social media and predictive AI company for financial services) and says if she were starting it today, LLMs would make signal mining from text messages and social posts far more accurate than the regex-based NLP they originally used.
She believes every department needs to rewrite job descriptions and that functional leads must set expectations for how AI changes roles—similar to how workers in the 1990s had to learn email and internet search.
On overhyped vs. underhyped: she thinks everything in AI is hyped but that we haven’t seen anything yet—comparing the current moment to 1998 but moving much faster.
Working with Marc Benioff: she describes him as inspiring, fun, and remarkably able to grasp diverse domains from sales and marketing to engineering and AI at a fundamental level.
Host Debrief: Key Takeaways
Permissioning is a major unsolved problem: Enterprises need granular control over which workflows AI can access, and Salesforce’s existing infrastructure gives it a significant advantage over startups trying to build this from scratch.
Vertical-specific evaluation will matter more than general benchmarks: Salesforce’s industry-specific teams building domain-specific eval sets reflects a broader trend toward specialized assessment over generic leaderboard rankings.
Slack as a multiplayer AI interface: The hosts found Clara’s point about Slack being ideal for team-plus-AI workflows compelling, noting that most AI interaction today is single-player but will evolve into collaborative, multiplayer settings.
Salesforce’s speed of execution: The company has shipped AI products rapidly and maintains a strong “always be launching” culture attributed to Marc Benioff’s leadership.
The hosts also discussed their recent investments in Abridge (AI that summarizes doctor-patient visits for electronic health records, eliminating “pajama time” charting) and Ideogram (a text-to-image model excelling at images with text, built by former Google imaging researchers), as well as Google’s mixed news week with the open-sourcing of Gemma and controversies around biases in its image generation models.
