Scaling artificial intelligence across enterprise operations presents significant hurdles, often extending beyond mere model capability to encompass complex data engineering and robust governance. Ahead of her appearance at AI & Big Data Global 2026 in London, Franny Hsiao, EMEA Leader of AI Architects at Salesforce, shed light on why many AI projects fail to reach production and how organizations can construct systems engineered for real-world demands.
The ‘Pristine Island’ Pitfall
Many AI deployments encounter difficulties because they are initially developed in isolated, controlled environments. These ‘pristine islands’ provide a deceptive sense of progress, only for the initiatives to collapse under the pressure of actual enterprise scale.
Hsiao identifies the most common architectural misstep as neglecting to establish a robust, production-ready data infrastructure with integrated, end-to-end governance from the outset. Pilots typically rely on small, meticulously curated datasets and simplified workflows. This approach, however, disregards the intricate reality of corporate data, which demands extensive integration, standardization, and transformation to manage its substantial volume and inherent variability.
Attempts to expand these isolated pilot projects without addressing the underlying data complexities inevitably lead to system failures. Hsiao warns that the ensuing data deficiencies and performance issues, such as delayed inference, render AI systems impractical and, more critically, untrustworthy. Organizations successfully navigating this chasm, she notes, are those that embed comprehensive observability and protective guardrails throughout the entire AI lifecycle, ensuring transparency into system effectiveness and user adoption.
Designing for Immediate Perception
As businesses adopt advanced reasoning models, a fundamental tension emerges between the computational intensity of these models and users' patience. Extensive processing often introduces noticeable latency.
Salesforce addresses this dilemma by prioritizing ‘perceived responsiveness,’ a strategy implemented through Agentforce Streaming. This method allows for the progressive delivery of AI-generated responses, even while the core reasoning engine conducts intensive background computations. It serves as an extremely effective technique for mitigating perceived delays, a frequent impediment to production AI.
Transparency further contributes to managing user expectations when scaling enterprise AI. Hsiao highlights the role of thoughtful design in fostering trust. By displaying progress indicators that illustrate reasoning steps or tool usage, alongside visual cues like spinners and progress bars, organizations not only sustain user engagement but also enhance perceived speed and build confidence. This visibility, combined with judicious model selection – opting for smaller models to reduce computations and accelerate response times – and explicit output length constraints, ensures the system feels deliberate and highly responsive.
Offline Intelligence at the Edge
For sectors with extensive field operations, such as utilities or logistics, continuous cloud connectivity is often impractical. According to Hsiao, the most significant practical requirement for many enterprise clients is robust offline capability.
Hsiao points to a growing trend towards on-device intelligence, particularly vital in field services where tasks must proceed irrespective of network signal. She provides an example: a technician in a disconnected environment can photograph a faulty component, error code, or serial number. An on-device large language model can then instantly identify the item or error and offer guided troubleshooting steps drawn from a cached knowledge base.
Upon regaining connectivity, the system automatically synchronizes this data back to the cloud, maintaining a unified source of truth. This approach guarantees that critical work continues even in the most remote settings. Hsiao anticipates continued advancements in edge AI, driven by benefits such as minimal latency, enhanced data privacy and security, energy efficiency, and cost reductions.
Controlled Autonomy: High-Stakes Gateways
Autonomous agents are not deployed and forgotten; scaling enterprise AI requires meticulous governance that defines precisely when human verification of an action is essential. Hsiao describes this not as a dependency, but as ‘architecting for accountability and continuous learning.’
Salesforce enforces a ‘human-in-the-loop’ for specific areas termed ‘high-stakes gateways.’ These include explicit action categories such as creating, uploading, or deleting (CUD) data, as well as verified contact and customer interaction activities. Furthermore, human confirmation is the default for critical decision-making or any action vulnerable to manipulation via prompt engineering.
This structured oversight establishes a feedback loop where agents gain insights from human expertise, cultivating a system of collaborative intelligence rather than unfettered automation. To foster trust in agents, their operational logic must be transparent. Salesforce’s Session Tracing Data Model (STDM) provides this clarity by capturing turn-by-turn logs that offer detailed insight into an agent’s reasoning processes. This granular, step-by-step visibility records every interaction, including user queries, planning stages, tool invocations, inputs/outputs, retrieved information, responses, timing metrics, and any errors. Such data empowers organizations to conduct ‘Agent Analytics’ for adoption metrics, ‘Agent Optimization’ for performance deep-dives, and ‘Health Monitoring’ for tracking uptime and latency. Hsiao summarizes Agentforce observability as the ‘single mission control for all Agentforce agents for unified visibility, monitoring, and optimization.’
Standardizing Agent Interaction
As enterprises integrate agents from various vendors, these systems necessitate a common communication protocol for effective collaboration. Hsiao asserts that ‘for multi-agent orchestration to work, agents cannot exist in isolation; they require a common language.’
Hsiao outlines a two-tiered standardization approach: orchestration and semantic meaning. For orchestration, Salesforce is adopting open-source standards like the Model Context Protocol (MCP) and Agent to Agent Protocol (A2A). The company believes open-source standards are vital for preventing vendor lock-in, enabling interoperability, and accelerating innovation.
However, communication is futile if agents interpret data differently. To resolve fragmented data understanding, Salesforce co-founded the Open Semantic Interchange (OSI) initiative. This aims to unify semantics, ensuring an agent in one system genuinely comprehends the intent of an agent in another.
The Future Bottleneck: Agent-Ready Data
Looking ahead, the primary challenge for enterprise AI will transition from refining model capabilities to ensuring data accessibility. Many organizations continue to grapple with legacy, fragmented infrastructure, where data searchability and reusability remain difficult.
Hsiao predicts the next significant hurdle—and its solution—will involve transforming enterprise data to become ‘agent-ready.’ This entails creating searchable, context-aware architectures that supersede traditional, inflexible ETL pipelines. This paradigm shift is crucial for enabling hyper-personalized and profoundly transformative user experiences, as agents will consistently have access to the appropriate context.
Ultimately, the focus for the coming year is not on the pursuit of larger, newer models, but rather on constructing the orchestration and data infrastructure that empowers production-grade agentic systems to flourish, Hsiao concludes.
Salesforce, a prominent sponsor of the upcoming AI & Big Data Global in London, will feature multiple speakers, including Franny Hsiao, who will share further insights during the event.
This article is a rewritten summary based on publicly available reporting. For the original story, visit the source.
Source: AI News
