No PriorsNo Priors Ep. 43 | With Clara Shih, CEO of Salesforce AI
CHAPTERS
- 0:00 – 1:07
Clara Shih’s path to leading Salesforce AI
Elad introduces Clara Shih’s background spanning entrepreneurship (Hearsay Social) and executive leadership at Salesforce. Clara explains how earlier NLP/ML work and her leadership in Service Cloud positioned her to take on the CEO role for Salesforce AI.
- •Clara’s career context: Hearsay Social, Service Cloud, broader Salesforce leadership scope
- •Early exposure to NLP/ML before today’s LLM wave
- •Service Cloud as a natural proving ground for AI in customer service
- 1:07 – 3:19
From Service Cloud experiments to an AI platform mandate
Clara describes the transition from early prototypes (with customers like Gucci) to a company-wide push after ChatGPT’s launch. She explains why Salesforce centralized efforts into shared services to move faster across clouds and the ecosystem.
- •Early prototyping cadence and learning loops with design partners
- •ChatGPT as an inflection point: “apply LLMs to every cloud”
- •Rationale for a common platform: fine-tuning, prompt tooling, trust layer, gateway
- •Formalizing Salesforce AI as a dedicated role to coordinate cross-org execution
- 3:19 – 5:32
Internal models vs. open ecosystem: Salesforce’s ‘open architecture’ approach
Elad asks how much model development is in-house versus external. Clara outlines a customer-choice strategy that supports enterprises with their own models, SMBs who want turnkey solutions, and customers who want Salesforce to select the best model per task.
- •Customer diversity drives a flexible model strategy (enterprise to SMB)
- •Mix of in-house models, customer fine-tuning, and third-party model options
- •Examples: CodeGen powering developer experiences; domain/industry fine-tuning
- •Support for partners and “bring your own API keys” purchasing/usage models
- 5:32 – 6:32
Shipping GPT features in every cloud: what’s live vs. what’s platform
Clara distinguishes between packaged, out-of-the-box GPT features (Service GPT, Sales GPT, etc.) and the underlying platform being built in parallel. She gives a concrete example: service reply recommendations grounded in knowledge and prior cases.
- •Out-of-the-box GPT features are prompt templates designed around workflow bottlenecks
- •Service reply recommendations as a flagship GA use case
- •Grounding responses in knowledge articles and similar historical cases
- •Two-track strategy: ship product value now while building reusable platform layers
- 6:32 – 8:32
Copilot, Copilot Studio, and the road to agents
Clara explains Salesforce’s Copilot as a cross-cloud natural language interface and Copilot Studio as the customization and extension layer. She breaks down the three core components—Prompt Builder, Action Builder, and Einstein Studio—and how they enable agent-like capabilities.
- •Copilot spans all Salesforce clouds plus Slack as an interface surface
- •Prompt Builder: customize templates, tone, grounding data, and model selection
- •Action Builder: turn workflows/integrations into agent actions with permissions and guardrails
- •Einstein Studio: bring-your-own models, training/fine-tuning using Salesforce data graph
- 8:32 – 10:28
Enterprise adoption reality: production wins, pilots, and the data prerequisite
Elad probes real adoption levels, and Clara describes a spectrum from full production rollouts to experimentation. She emphasizes that many organizations must first unify and connect data before generative AI can be reliable and scalable.
- •Some customers have operationalized GPT features across contact centers
- •Most enterprises remain in experimentation due to data fragmentation
- •Data Cloud as an enabler for connecting internal lakes and systems
- •Zero-ETL partnerships (e.g., BigQuery, Databricks, Snowflake) to simplify data access
- 10:28 – 13:00
When will enterprise AI scale? ‘Second or third inning’ and where to start
Clara estimates enterprise AI adoption is still early, with a few strong proof points but broad readiness gaps. She and Elad discuss an adoption progression and why customer service is often an easier first bite-size use case.
- •Enterprise adoption is early but value is increasingly proven
- •Data readiness is the dominant constraint for most large companies
- •Adoption often progresses: data cleanup → internal tools → customer-facing rollout
- •Customer service can start sooner by unifying knowledge sources for RAG
- 13:00 – 14:13
Cross-team enablement: educating product and engineering for an agent-driven future
Sarah asks how Clara drives AI understanding across Salesforce’s broad product org. Clara describes a collaborative, bottom-up ideation culture and predicts agents will reshape how software is specified and built.
- •AI roadmap ideas come from across PM and engineering teams
- •Education is continuous as capabilities evolve quickly
- •Agents may replace hard-coded branching with dynamic decisioning
- •Shift in building software: less screen-by-screen specification, more goal-driven design
- 14:13 – 18:36
‘AI is the new UI’: Slack, Generative Canvas, and next-gen enterprise UX
Clara expands on AI-driven user experiences, from conversational interfaces (often in Slack) to dynamic UI assembly via a prototype called Generative Canvas. She explains how AI could surface the right components, visualizations, and record updates without pre-wiring every page.
- •Conversational UX as a primary interface layer—often via Slack
- •Generative Canvas prototype dynamically surfaces relevant UI components
- •Pulls from Tableau/Reporting and can drill into records based on conversation
- •AI-driven UX is promising but still “raw,” requiring iterative customer testing
- 18:36 – 20:59
Data structuring and governance: making unstructured data usable (and safe)
Sarah asks about handling sensitive enterprise data, and Clara distinguishes between types of unstructured data. She explains why transcripts and chats need preprocessing, and how valuable structured attributes can be extracted back into CRM fields.
- •Not all unstructured data is alike: curated docs vs. noisy transcripts/chats
- •Transcript-like data often wasn’t intended for broad reuse—needs mining and preprocessing
- •Vectorization/embeddings for retrieval, with additional cleaning steps for relevance and safety
- •Extracting structured signals (preferences, customer attributes) from conversations into CRM fields
- 20:59 – 22:28
The future enterprise stack: deterministic systems vs. stochastic AI workflows
Elad asks what changes in enterprise software to expect over the next few years. Clara predicts a split: some workflows remain deterministic (finance, auth, healthcare), while many others shift toward AI-driven execution where teams specify intent and outcomes rather than exact steps.
- •Analogy to early cloud adoption: not everything migrates the same way
- •Deterministic workflows remain critical for high-stakes, repeatable processes
- •PM/engineering/design roles shift from defining “how” to defining “what/why/goal”
- •AI handles the stochastic “how” dynamically as agents mature
- 22:28 – 25:23
Pricing and unit economics: balancing compute COGS with clear customer ROI
Sarah raises the challenge of AI compute costs and pricing models that customers can understand. Clara emphasizes the need to cover costs while keeping pricing simple, with ROI proof (handle time reduction, conversion uplift) as the ultimate anchor.
- •AI pricing is hard: real compute COGS and variable usage patterns
- •Avoiding overly complex token calculators for customers
- •Value proof is essential: productivity, handle time, conversion improvements
- •Example of massive productivity shift in media (Runway’s VFX team scale reduction)
- 25:23 – 27:14
Where startups should play: models, vertical solutions, tooling, and new apps
Clara closes by outlining startup opportunities across layers of the stack. She highlights domain-specific applications, tooling gaps (e.g., for differing org needs), and the importance of aligning with data graphs to make applications relevant in real enterprise environments.
- •Opportunities span foundation models, vertical/domain specialists, and tooling layers
- •Tooling remains underbuilt and highly variable by organization type
- •Many new AI-native applications still need to be invented and tested
- •Successful apps will align tightly with enterprise data graphs and systems of record
