Can you trust AI? It’s a question a lot of people are asking. And many say “No.”
But why? Aren’t most enterprises testing AI models, while tens of billions of dollars are being spent on supporting infrastructure?
Seems like a disconnect: IT teams are pressured to move AI projects from trial to production, while senior management has concerns about the initial value and ROI of the effort.
It really isn’t a contradiction; both things are true. We’re at a point where adoption of AI is running ahead of the controls and reliability required by senior executives who control budgets. That’s why there’s growing concern among the C-suite community who question whether AI is a potentially risky experiment rather than a well-managed business asset.
There’s no question AI technology has enormous potential benefits. What’s often lacking is an architecture that ensures AI outcomes can be accurately replicated, validated, and audited.
And this leads to skepticism, doubt, and rising demands for proof – not trust.
AI Raises Critical Questions
Like any new technology, the early stages of AI development involved a small core of autonomous and highly specialized engineering and development teams. But as AI moves from experimentation to production, managers from other disciplines and departments (including sales, legal, finance, and the C suite) become involved. They’re asking challenging questions, like:
- Has this passed governance review?
- Where’s the audit trail?
- If you can’t duplicate this result, how do we know it’s accurate?
After talking with customers, partners, and industry observers, we believe the single biggest contributor to the AI trust gap is uncertainty. Agentic AI is unfamiliar in many ways, and managers are learning some hard truths, such as:
- Humans track progress incrementally; agents plow ahead, unrestrained and naive
- Humans can explain how they reached their conclusions; agents just provide results, which aren’t always verifiable.
- Agents do not operate with an inherent understanding of data hygiene, governance rules, or compliance boundaries.
- Legacy security controls were not designed to govern how data is transformed, inferred, and propagated through AI workflows.
- Agents will use whatever data they can access, not just the data they should use.
The Foundation of Trust
Trust has always been built through collaboration. But that doesn’t necessarily happen with agentic AI. In fact, the hallmark of agentic AI is autonomy, the opposite of collaboration. And the more autonomy, the less insight into sources and methods – and less trust.
Consider this for context: if someone on your team delivered a recommendation without clear evidence, source references, or a rationale, you would not simply accept it. The same standard must apply to AI. When an AI output cannot be reliably traced back to its sources, rules, and decision path, it is minimized as a hallucination. In reality, it reflects a breakdown in transparency and control.
Let’s be clear what we’re talking about. We’re not discussing AI-fueled applications that automate mind-numbing routine tasks like sorting customer comments, tracking inventory or answering simple questions. We’re talking about complex applications – like CRM, ERP, financial modeling, or healthcare management – that draw data from multiple locations, span organizational and geographic boundaries, and are subject to legal or regulatory oversight.
These are applications that require the highest level of accuracy and verifiability – in other words, trustworthiness.
The problem is not that the outcome of agentic AI projects can’t be explained and justified with the appropriate technological underpinnings. The challenge is that many AI applications are quickly rolled out without the business rules and operating context that are necessary for trustworthy answers.
The closer agentic AI applications come to full production, the more scrutiny must be applied to critical data-security and governance issues. But agentic AI operates so differently from traditional IT applications that security and governance procedures must be completely rethought.
Reimaging Data Security and Governance for AI
Agentic workflows aren’t just another category of application. They operate in an entirely different environment – of LLMs, RAG, MAS*, HITL**, and more. It’s one that legacy security measures weren’t designed for.
Traditional data-security models were developed for static systems and predictable access patterns based on user’s permissions. They assume that data resides in known locations and that users access it through well-defined interfaces. Plus, they presume controls can be enforced at the perimeter, application, or infrastructure layer. For decades, this model worked because data interactions were largely linear and bounded.
AI completely upends this model.
Modern AI systems introduce new data-interaction models, such as multi-step reasoning, continuous retrieval, and agentic workflows. Data used in AI models is transformed, combined, and analyzed rather than simply accessed. If AI teams aren’t rigorous in their oversight, AI models will indiscriminately ingest everything they’re given, regardless of provenance, and pass through any inaccuracies or false assumptions.
Further complicating data security, AI models look to ingest sensitive information flows through prompts, embeddings, intermediate representations, tools, and outputs. And this often occurs across multiple systems (and even geographies) in a single interaction.
Legacy controls designed to surround systems cannot reliably govern these interactions. Guardrails applied after the fact, prompt-based restrictions, and access-only policies were never built to track how data is used, transformed, or inferred across AI workflows.
This is not a failure of implementation — it is an architectural mismatch between legacy security assumptions and modern AI behavior.
In this new environment, trust cannot be established through outdated procedures. It must be earned by implementing new measures and disciplines. So, what does this solution look like?
The Need for Data-Centric Security
Earlier, we identified uncertainty as the main source of the AI trust gap. So, it follows that an essential way to build trust is to increase certainty.
And certainty starts with data, the heart of any AI project. The best model won’t produce trustworthy results if the information it ingests is faulty, incomplete, or unsecured. Results can only be trusted when there is certainty that:
- the information in the model has been properly sourced and secured
- governance rules are strictly followed
- humans are in the loop at all critical stages
- semantic guardrails are in place to prevent errors, omissions, and malicious intent.
And here’s another contributor to certainty: consistency. As every step of a model is executed, teams must be assured that data-hygiene procedures are followed and governance closely monitored. If checks and audits must be repeated – say, when data moves across borders or between companies – progress is slowed.
Establishing trust involves many people and many steps. But in AI environments, data rarely stays in one place or one form. It is retrieved, transformed, combined, and passed across models, tools, and workflows. That is why governance cannot depend solely on perimeter controls or application-level rules.
A practical starting point is to embed governance standards and data-security measures into the data itself. Techniques such as masking, tokenization, and encryption help reduce direct exposure of sensitive information while preserving protection as data moves through retrieval, reasoning, and generation, so policies can be enforced consistently and deterministically at runtime.
Only data-centric security can safely support wide-ranging AI workflows while ensuring sensitive information is never exposed. This is the foundation for AI to operate consistently, at scale, under real-world constraints.
To be blunt, trust is not an aspirational goal. It is an operational necessity that demands deterministic protection and enforceable governance of all data and interactions used by every AI model. Your AI success depends on it.
* MAS (Multi-Agent Systems): Systems composed of multiple interacting intelligent agents that collaborate or coordinate to perform tasks.
** HITL (Human-in-the-Loop): An approach where human input, oversight, or intervention is integrated into automated or AI-driven processes to improve accuracy and governance.