Scaling enterprise AI demands a fundamental shift beyond selecting the right models. The journey from a promising pilot to a robust, production-ready AI asset hinges on tackling the complex challenges of data engineering and governance. As generative AI prototypes become increasingly accessible, the real hurdles emerge in transforming these experimental tools into reliable business drivers.
Franny Hsiao, EMEA Leader of AI Architects at Salesforce, recently shared insights into why many AI initiatives falter and how organizations can build resilient systems capable of navigating the complexities of real-world deployment. Her perspective, shared ahead of the AI & Big Data Global 2026 event in London, highlights critical architectural considerations.
### The “Pristine Island” Pitfall in Enterprise AI Scaling
A primary cause of AI pilot failure lies in the environment where these systems are developed. Prototypes are often built in highly controlled settings, creating a deceptive sense of security that evaporates when confronted with the demands of enterprise-scale operations.
“The most common architectural oversight that hinders the scaling of AI pilots is the failure to establish a production-grade data infrastructure with end-to-end governance built in from the outset,” Hsiao explained.
She elaborated on the tendency for pilots to originate on “pristine islands”—environments utilizing small, carefully curated datasets and streamlined workflows. This approach, however, neglects the often-unwieldy reality of enterprise data, which requires intricate integration, normalization, and transformation to manage the sheer volume and variability encountered in real-world scenarios.
When companies attempt to scale these “island-based” pilots without addressing the underlying data complexities, the systems inevitably falter. Hsiao warns that “the resulting data gaps and performance issues, such as inference latency, render AI systems unusable—and, more importantly, untrustworthy.”
Organizations that successfully navigate this challenge, according to Hsiao, are those that “bake end-to-end observability and guardrails into the entire lifecycle.” This holistic approach provides crucial “visibility and control over the effectiveness of AI systems and how users are adopting the new technology.”
### Engineering for Perceived Responsiveness
As enterprises deploy sophisticated reasoning models, such as Salesforce’s ‘Atlas Reasoning Engine,’ a critical trade-off emerges between the depth of the AI’s analytical process and the user’s patience. Extensive computation can lead to noticeable latency.
Salesforce addresses this by prioritizing “perceived responsiveness through Agentforce Streaming,” Hsiao noted. This innovative approach enables the progressive delivery of AI-generated responses, even while the underlying reasoning engine undertakes heavy computational tasks. It’s a powerful strategy for mitigating the perception of latency, which is a common impediment to production AI deployment.
Transparency also plays a vital role in managing user expectations during the scaling of enterprise AI. Hsiao emphasized the strategic use of design elements as trust mechanisms. By displaying progress indicators that illustrate the reasoning steps or the tools being utilized, alongside visual cues like spinners and progress bars to signify loading states, organizations can not only maintain user engagement but also enhance the perception of responsiveness and foster trust.
“This visibility, combined with strategic model selection—such as opting for smaller models that require less computation and thus deliver faster response times—and explicit length constraints, ensures the system feels deliberate and responsive,” she added.
### Offline Intelligence at the Edge
For industries with significant field operations, such as utilities or logistics, a reliance on constant cloud connectivity is often impractical. “For many of our enterprise customers, the most significant practical driver is offline functionality,” Hsiao stated.
She highlighted a growing trend toward on-device intelligence, particularly in field services, where workflows must continue seamlessly regardless of network signal strength.
“A technician can photograph a faulty part, error code, or serial number while offline. An on-device LLM can then identify the asset or error and instantly provide guided troubleshooting steps from a cached knowledge base,” Hsiao explained.
Data synchronization occurs automatically once connectivity is restored. “Once a connection is re-established, the system handles the ‘heavy lifting’ of syncing that data back to the cloud to maintain a single source of truth. This ensures that work proceeds uninterrupted, even in the most disconnected environments.”
Hsiao anticipates continued innovation in edge AI, driven by benefits such as “ultra-low latency, enhanced privacy and data security, energy efficiency, and cost savings.”
### High-Stakes Gateways for Accountability
Autonomous agents are not designed for a “set it and forget it” approach. When scaling enterprise AI deployments, robust governance requires clearly defined checkpoints for human verification of actions. Hsiao frames this not as a limitation, but as “architecting for accountability and continuous learning.”
Salesforce mandates a “human-in-the-loop” system for specific operational areas, which Hsiao categorizes as “high-stakes gateways.”
“This includes specific action categories, such as any ‘CUD’ (Creating, Uploading, or Deleting) actions, as well as verified contact and customer contact actions,” Hsiao stated. “We also default to human confirmation for critical decision-making or any action that could potentially be exploited through prompt manipulation.”
This structured approach fosters a feedback loop where “agents learn from human expertise,” cultivating a system of “collaborative intelligence” rather than unchecked automation.
Building trust in an AI agent necessitates visibility into its operational processes. Salesforce has developed a “Session Tracing Data Model (STDM)” to provide this critical insight. It captures “turn-by-turn logs” that offer granular detail into the agent’s decision-making logic.
“This provides us with granular step-by-step visibility that captures every interaction, including user queries, planner steps, tool calls, inputs/outputs, retrieved data, responses, timing, and error logs,” Hsiao elaborated.
This comprehensive data enables organizations to perform ‘Agent Analytics’ for adoption metrics, ‘Agent Optimization’ for in-depth performance analysis, and ‘Health Monitoring’ for tracking uptime and latency.
“Agentforce observability serves as the central command for all Agentforce agents, offering unified visibility, monitoring, and optimization,” Hsiao summarized.
### Standardizing Agent Communication
As businesses increasingly deploy agents from various vendors, these systems require a shared protocol to facilitate effective collaboration. “For multi-agent orchestration to function, agents cannot operate in isolation; they need a common language,” Hsiao argued.
She outlined two crucial layers of standardization: orchestration and meaning. For orchestration, Salesforce is adopting open-source standards such as MCP (Model Context Protocol) and A2A (Agent to Agent Protocol).
“We believe open-source standards are non-negotiable; they prevent vendor lock-in, enable interoperability, and accelerate innovation.”
However, effective communication is undermined if agents interpret data differently. To address data fragmentation, Salesforce co-founded OSI (Open Semantic Interchange) with the aim of unifying semantics, ensuring that an agent in one system can “truly understand the intent of an agent in another.”
### The Future Enterprise AI Scaling Bottleneck: Agent-Ready Data
Looking ahead, the primary challenge in enterprise AI scaling is poised to shift from model capabilities to data accessibility. Many organizations continue to grapple with legacy, fragmented infrastructure, where “searchability and reusability” remain significant hurdles.
Hsiao predicts that the next major obstacle—and its corresponding solution—will involve making enterprise data “‘agent-ready’ through searchable, context-aware architectures that replace traditional, rigid ETL pipelines.” This evolution is essential to deliver “hyper-personalized and transformed user experiences because agents can consistently access the appropriate context.”
“Ultimately,” Hsiao concluded, “the coming year is not about the race for bigger, newer models; it’s about building the orchestration and data infrastructure that allows production-grade agentic systems to thrive.”
Original article, Author: Samuel Thompson. If you wish to reprint this article, please indicate the source:https://aicnbc.com/16695.html
