Artificial intelligence has entered a new phase: It’s no longer just chatbots answering questions. It’s moved towards autonomous agents that take action.
Agentic AI systems can perceive conditions, decide what to do next, and execute tasks across enterprise systems. This move from insight to action is what makes AI transformational – but it’s also what makes it risky. When AI moves from answering to doing, the trust bar is raised significantly.
Our global survey of 850 enterprise leaders reveals a harsh reality: Organizations are investing heavily in AI, but most are not yet equipped to make it trustworthy. The gap is wide between what AI requires and what enterprises can deliver.
This is not an AI problem. It’s a data problem. In my experience working with enterprises adopting AI, the point of failure is not the model or how it was trained. It’s when AI is tested or deployed with real-world production data.
The 3 gaps undermining AI trust
There are three critical data-related gaps that consistently undermine AI initiatives: accessing live data, identifying the right data, and complying with the guardrails.
1. The live data gap
Agentic AI depends on real-time awareness. It needs to respond to changing business conditions as they happen, yet most enterprise data architectures were designed for reporting and analytics, not operational decision-making.
The result? A fundamental imbalance. According to the survey, 66% of organizations believe AI data must be real-time or near real-time to be trustworthy, but many current architectures cannot deliver that responsiveness.
This challenges operational use cases like fraud detection, customer engagement, or responding to outages. The outcome is a growing disconnect between how quickly AI can act and how slowly enterprise data environments can react.
2. The ‘right data’ gap
Even when data is available and completely clean and accurate, that doesn’t mean it’s the right data for the task at hand. The survey found that 63% of organizations struggle to identify and prepare trustworthy data for AI initiatives. The challenge is not only technical. It is contextual.
Modern enterprises operate across hundreds of systems. On average, enterprises draw from more than 400 data sources for their AI initiatives, while nearly 20% rely on more than a thousand. In that environment, determining which data is authoritative for a given situation, and how it should be interpreted, is far from straightforward.
In these environments, the same business concept can mean different things depending on the system involved. A ‘customer’ in a CRM platform may not match a ‘customer’ in a billing platform or customer support system.
For example, sales tools may track transactions and interactions, support systems may hold service histories, and social monitoring platforms may capture customer sentiment. All of that data may be accurate on its own, but AI still needs to determine which source is authoritative in a specific situation.
This is where many AI initiatives fail, not because the models are wrong, but because the context is.
3. The guardrail gap
If agentic AI is going to take action, it must do so within clearly defined boundaries. Governance remains one of the biggest challenges as the survey showed 67% of organizations struggle with AI data security and access controls.
The report makes an important distinction: Guardrails for AI must go beyond model safety. They are about enforcing policy, compliance, and access controls consistently for data from every system an AI agent can reach.
Most organizations struggle to maintain consistent governance and security across highly distributed environments, particularly when AI requires direct access to live data from hundreds of systems. Without these controls, AI doesn’t just produce bad answers; it can take the wrong actions.
Why traditional architectures fall short
Many organizations have invested heavily in modern data platforms, cloud infrastructure, data lakehouses and data catalogs. These are essential foundations, but they are not enough for operational AI. Nearly 60% of organizations still struggle to optimize performance for AI workloads, regardless of whether the organization adopted a modern data platform.
Traditional approaches rely on copying and consolidating data into centralized systems that work for analytics but break down for operational AI. Agentic AI requires live data access, contextual decision-making, consistent governance, and scalable performance across distributed environments.
Trying to solve these issues individually within every source system does not scale. Instead, organizations need a centralized layer where access policies, governance, and data controls can be consistently applied.
The reality: AI runs on distributed data
One of the clearest findings from the report is that AI does not operate in a neatly consolidated data environment. The curated, consolidated data sets used during model training deliver little resemblance to the messy, distributed nature of data in real-world operational environments. In production, AI systems must operate across this fragmented, hybrid, multi-source landscape.
This complexity is only increasing as organizations are expanding their agentic use cases, often necessitating access to new categories of data, including unstructured and third-party data. AI increasingly relies on both structured and unstructured data spread across multiple systems.
For example, a customer-facing agent may need access to transaction history, support records, and call transcripts to make informed decisions. As agentic use cases expand, organizations must connect these sources securely, consistently, and quickly.
Closing the trust gap
To build trustworthy AI, organizations need more than powerful models. They need a modern data foundation designed for operational intelligence.
That foundation should provide the following:
- Live data access to reflect real-world conditions in real time
- Semantic consistency for a shared understanding of data across systems
- End-to-end guardrails to enforce data governance, security, and compliance across all systems
- Cost and performance optimization so AI can scale economically across many agents and data sources
In other words, organizations need to rethink the data layer. It is no longer a place where data is stored, but a capability that delivers governed, contextual, real-time access to data wherever it resides.
A shift in thinking
The most important takeaway from this report is that closing the AI trust gap requires a new way of thinking about data management. Many enterprises are pursuing AI transformation while relying on decades-old data management principles, and it is holding them back.
For decades, the industry has relied on the idea of centralizing data, bringing it to powerful compute engines through data pipelines and ETL jobs. From enterprise data warehouses in the 1990s, to big data clusters in the 2010s, to modern cloud-native platforms today, the core principle has remained the same: Move the data to where the intelligence resides.
Agentic AI flips that model. Instead of moving data to the intelligence, the intelligence must reach the data wherever it lives, while maintaining governance, context, and performance in real time. For organizations that fail to make this shift, the consequences can be significant.
Gartner predicts that more than 40% of agentic AI projects will be canceled by 2027 due to cost, risk, and lack of business value. Organizations that rethink their data architecture to deliver live, trusted, and governed data will be far better positioned to succeed with AI.
Be First to Comment