Franny Hsiao, Salesforce: Scaling Enterprise AI

Why It Matters

Without robust data pipelines and governance, AI projects stall, costing enterprises time and money; Salesforce’s roadmap offers a replicable blueprint for scaling trustworthy, production‑grade AI across industries.

Key Takeaways

• Pilots fail due to isolated “pristine island” data setups
• End‑to‑end observability and guardrails enable production scaling
• Progressive streaming reduces perceived latency for large models
• Edge AI delivers offline intelligence for field service teams
• Open standards ensure multi‑agent interoperability and data consistency

Pulse Analysis

Enterprises often launch AI pilots in controlled, curated datasets that give a false sense of success. Hsiao stresses that the real challenge lies in building a production‑grade data infrastructure with built‑in governance, normalization, and transformation capabilities. By treating data as a continuous asset rather than a sandbox, organisations avoid the “pristine island” trap and create a foundation for reliable, scalable AI.

Latency perception is another critical hurdle. Salesforce’s Agentforce Streaming delivers responses incrementally, letting users see progress while heavy reasoning continues in the background. This approach, combined with design cues such as progress bars, keeps users engaged and builds trust. Simultaneously, edge AI solutions empower field technicians to operate offline, capturing images or error codes and receiving instant on‑device insights, then syncing data once connectivity returns—driving productivity in low‑signal environments.

Interoperability and accountability round out the scaling equation. Open protocols like MCP and A2A, alongside the Open Semantic Interchange, prevent vendor lock‑in and enable agents from different systems to communicate meaningfully. Human‑in‑the‑loop safeguards at high‑stakes gateways ensure critical actions receive expert review, while Salesforce’s Session Tracing Data Model provides granular visibility for monitoring and optimization. Looking ahead, the industry’s bottleneck will shift from model size to making enterprise data “agent‑ready,” emphasizing searchable, context‑aware architectures that power hyper‑personalized experiences.