As enterprises accelerate the adoption of Artificial Intelligence (AI) and agentic systems, the conversation around data protection is becoming more urgent.
Yet, according to Dariel CTO Wayne Yan, organisations should resist the temptation to treat AI as an entirely new governance problem. Instead, businesses should evolve and extend the same information security principles they already rely on today.
“Fundamentally, data loss prevention protocols are essential rules designed to ensure that sensitive data either has authority or not to traverse organisational boundaries,” says Yan. “Those boundaries may be defined by geography, networks, systems, or even departments within the organisation. The type of rules embodied in data security policy are really not different for AI-oriented solutions. These are essentially the same policies one would apply under any integration-driven solution.”
Yan explains that while AI introduces new layers of complexity, the underlying governance challenge remains familiar. Enterprises already manage ecosystems of partners, SaaS providers, cloud platforms, and integrations. AI solutions simply extend those ecosystems further.
Understanding the AI stack
To govern AI responsibly, organisations first need to understand the building blocks of modern agentic systems. According to Yan, these typically include:
- A large language model (LLM), which may be proprietary, open-weight, or self-hosted.
- Training data used to build the model, often a combination of public and private information.
- Context layers that provide domain-specific enterprise knowledge.
- Context-bound rules that shape how answers are generated.
- APIs and integrations that allow agents to perform business actions.
“The real governance question is how private domain data traverses’ enterprise boundaries,” Yan explains. “For example, if the LLM is hosted in a foreign jurisdiction, should it receive context that is not legally allowed to leave a legislated boundary?”
He notes that businesses do have options, including self-hosting models or limiting how external hosted services are used. But these decisions require careful governance, particularly when organisations operate in regulated industries or across multiple jurisdictions.
“This is not unlike the governance models businesses already apply to traditional cloud or SaaS solutions,” he says. “The responsibility of the CIO remains the same: understand your legislative requirements, define your information boundaries, and manage integration parties responsibly.”
The trust dilemma
While data governance remains critical, Yan believes the larger unresolved issue lies in the trustworthiness of AI-driven decision making. “The pressing question organisations need to answer is: to what extent can they trust the decision-making capability of AI?” he says.
Unlike traditional software systems, large language models are probabilistic rather than deterministic. This means outputs may appear plausible and convincing without necessarily being factually correct.
“Correctness is not guaranteed,” Yan explains. “Semantic plausibility is not equivalent to factual accuracy. An answer can look right while still being wrong.”
He points to scenarios where AI agents may guide customers toward financial products, insurance policies, or automated onboarding processes. In these environments, questions around accountability quickly emerge.
“If an AI-driven broker agent provides advice that is not in the customer’s best interests, who becomes liable?” Yan asks. “Is it the LLM provider, the developer who assembled the solution, or the organisation that provided the contextual data? Businesses need to understand that these risks extend far beyond traditional data loss prevention concerns.”
Responsible AI implementation
According to Yan, organisations should avoid rewriting their information security policies entirely. Instead, they should evolve and augment existing frameworks to accommodate AI-enabled systems.
“The common problem is that information wants to be set free,” he says. “The common solution is that effort must be applied to implement information security that confines information for permissible and lawful use. This pattern transcends technology.”
He adds that agentic systems should operate under tightly governed permissions, much like human users within enterprise environments.
“Agents designed for specific intentions should only be allowed to execute actions aligned with their mandated intent,” Yan says. “The endpoints within the private domain must still be guarded no differently than before.”
For Dariel, the future of AI adoption lies in disciplined execution, governance, and responsible implementation rather than unchecked experimentation.
“Innovation without governance creates risk,” Yan concludes. “Businesses absolutely should adopt AI technologies, but they need to tread responsibly. The customer trust relationship must never be jeopardised in pursuit of automation or convenience.” Ends.