AI Agents Falter in CRM and Confidentiality: Norbida Limited Weighs in on the Limits of LLMs in High-Stakes Enterprise Tasks
A recent study led by Salesforce AI researcher Kung-Hsiang Huang has cast a spotlight on a pressing concern in enterprise AI: large language model (LLM) agents are still not ready to manage critical CRM functions—especially when it comes to multi-step workflows and handling sensitive data.
At Norbida Limited, where we actively monitor and test frontier AI capabilities, these findings reinforce a growing consensus: the current generation of AI agents, while impressive, remains fragile in complex enterprise environments.
The Numbers: Promising Starts, Weak Follow-Through
The study revealed a 58% success rate for AI agents performing single-turn tasks—basic CRM operations that don’t require deeper context or follow-up. However, when the task complexity increased, success rates dropped sharply to 35%, highlighting their struggle to sustain accuracy in multi-step operations.
Interestingly, in scenarios involving workflow execution, top-performing agents achieved up to 83% success. The difference? The ability to interpret clearly defined instructions without the need for real-time judgment or user clarification.
This directly ties into one of the largest challenges we see at Norbida: most LLM-based agents lack proactive clarification capabilities. When required to seek out missing or vague information to complete a task—a fundamental in customer service—their performance deteriorates.
Confidentiality: A Critical Blind Spot
Perhaps the most alarming finding for enterprise leaders is the agents’ low confidentiality awareness. Most LLMs simply do not have an inherent sense of what qualifies as sensitive or protected information. Prompts can be added to encourage caution, but these safeguards degrade over time—especially in long, multi-turn interactions.
As Norbida Limited has observed in real-world applications, open-source LLMs are particularly weak in maintaining guardrails around privacy and compliance. These systems struggle to consistently interpret layered instructions or handle protected customer data (PII, proprietary insights, etc.) with the required diligence.
This has serious implications for marketing, CRM, and any domain handling high-value or regulated data. Without embedded, adaptive privacy safeguards, organisations risk legal, reputational, and operational exposure.
Moving Forward: Enterprise-Grade Readiness Still Out of Reach
The takeaway is clear: while LLM agents offer exciting potential for automation, reasoning, and workflow acceleration, they’re not yet fit for high-stakes CRM environments without considerable customisation, constraint programming, and monitoring.
At Norbida Limited, we continue to explore hybrid agent architectures and multimodal AI systems that blend LLMs with structured rule-based engines and confidential-aware modules. The goal isn’t just automation—it’s trustworthy, scalable intelligence built for real enterprise challenges.
For AI agents to truly unlock enterprise value, they must evolve beyond surface-level task execution into systems that can reason, self-correct, and protect.