As interest in enterprise AI continues to surge, many organisations are discovering that moving from promising pilot projects to reliable, large-scale deployment is far harder than expected. According to Franny Hsiao, EMEA Leader of AI Architects at Salesforce, the biggest obstacles have little to do with choosing the right model and everything to do with architecture, data, and governance.
Speaking ahead of AI & Big Data Global 2026 in London, Hsiao shared insights into why so many enterprise AI initiatives stall and what it takes to build systems that work in real-world conditions.
Why enterprise AI pilots often fail to scale
Hsiao points to what she calls the “pristine island” problem. Many AI pilots are developed in tightly controlled environments using clean, limited datasets and simplified workflows. While this makes early results look impressive, it rarely reflects the complexity of enterprise data.
“The most common architectural oversight is failing to design a production-grade data infrastructure with end-to-end governance from the start,” Hsiao explains.
Once these pilots are exposed to messy, fragmented, and high-volume enterprise data, performance issues emerge. Latency increases, data gaps appear, and trust in the system quickly erodes.
Organisations that succeed, she says, build observability and guardrails into the entire AI lifecycle. This allows teams to track performance, understand user adoption, and address issues before they undermine confidence.
Designing for speed users can feel
As enterprises deploy large reasoning models, they face a practical trade-off: deeper reasoning often means slower responses. Salesforce addresses this challenge by focusing on perceived responsiveness.
Through techniques such as progressive response streaming, AI outputs are delivered incrementally while complex computations continue in the background. This approach reduces the feeling of waiting and keeps users engaged, even when models are doing heavy work behind the scenes.
Transparency also plays a key role. Visual cues like progress indicators, loading states, and explanations of reasoning steps help manage expectations and build trust. Combined with careful model selection and response length limits, these design choices make AI systems feel faster and more deliberate.
Bringing AI to the edge for offline work
For industries such as utilities, logistics, and field services, constant cloud connectivity cannot be assumed. Hsiao highlights growing demand for on-device intelligence that works offline.
In these scenarios, a technician can capture an image of a faulty component or error code without a network connection. An on-device language model can then identify the issue and provide guidance using locally cached data. Once connectivity is restored, information syncs back to the cloud automatically, preserving a single source of truth.
Hsiao expects edge AI to continue gaining momentum due to benefits such as lower latency, stronger data privacy, energy efficiency, and reduced operational costs.
Governance through human oversight
Autonomous agents, Hsiao stresses, should never operate without clear accountability. Salesforce uses a “human-in-the-loop” approach for what it calls high-stakes gateways, including actions that create, update, or delete data, as well as sensitive customer interactions.
This framework ensures that critical decisions are reviewed by humans while allowing agents to learn from expert feedback. The result is a model of collaborative intelligence rather than unchecked automation.
To support this, Salesforce has developed detailed tracing tools that log each step an agent takes, from user prompts to tool calls and final responses. These logs enable performance analysis, monitoring, and continuous optimisation across deployed agents.
The need for shared standards between agents
As enterprises adopt AI agents from multiple vendors, interoperability becomes essential. Hsiao argues that agents must share common protocols to collaborate effectively.
Salesforce is backing open-source standards for agent orchestration while also addressing a deeper issue: semantic consistency. Without shared meaning, agents may exchange data but misunderstand intent. To solve this, Salesforce co-founded an initiative focused on aligning data semantics so agents across systems can truly understand one another.
The next bottleneck: agent-ready data
Looking ahead, Hsiao believes the biggest challenge in scaling enterprise AI will shift from model performance to data accessibility. Many organisations still rely on rigid, legacy pipelines that limit how easily data can be searched, reused, and contextualised.
The next phase, she says, will focus on making enterprise data “agent-ready” through more flexible, context-aware architectures. This shift will allow AI agents to access the right information at the right time, enabling more personalised and effective user experiences.
