Regulated enterprises are currently navigating a high-stakes contradiction between AI adoption goals and fundamental data security requirements. A new survey from Fortanix, conducted by Atomik Research, indicates that while a vast majority of AI leaders recognize the inherent risks of processing sensitive information in public environments, a significant portion continues to utilize these services regardless. This tension is creating measurable friction in enterprise digital transformation, as security and compliance concerns move from theoretical risks to active blockers for large-scale AI deployment. The findings highlight a growing demand for infrastructure that can provide cryptographic assurance during the critical phase of data processing, signaling a potential shift in how highly regulated industries architect their long-term AI ecosystems.
The Compliance Gap in Public Cloud AI Deployment
The survey of 300 AI leaders across banking, healthcare, government, and insurance reveals a stark disconnect between perceived risk and operational reality. While 93% of respondents acknowledge that traditional encryption fails to protect data during active processing, 61% of these organizations still transmit sensitive data to public cloud AI services. This exposure is not viewed lightly by leadership; 93% of those surveyed expressed concern regarding the potential exposure of proprietary prompts, models, or intellectual property to AI infrastructure providers, with nearly 40% describing themselves as “very concerned.”
This lack of confidence is directly impacting project timelines and deployment capabilities. According to the Fortanix report, 59% of organizations have delayed an AI project within the past year due to privacy or security concerns, while 41% have completely blocked a deployment. Notably, regulatory compliance has emerged as the primary obstacle to utilizing sensitive data for AI, cited by 46% of respondents. This places compliance ahead of direct risks like data exfiltration or exposure. The data suggests that the inability to meet stringent regulatory standards is currently a more significant bottleneck for enterprise AI than the technical threat of a data breach itself.
Shifting Architectures Toward Sovereign and Hybrid AI
To mitigate these risks, enterprises are actively seeking to move AI workloads closer to their controlled data environments. The research indicates a strong movement toward localized control, with 64% of organizations having already deployed closed-weight frontier AI models on-premises. The appetite for secure, localized options is even higher, as 97% of respondents stated they are at least somewhat likely to adopt leading AI models if those models are made available securely on-premises. This trend is closely linked to the concept of digital sovereignty, with 95% of leaders stating that sovereign AI is either important or critically important to their overarching AI strategy.
Looking ahead, the enterprise landscape appears to be moving away from a pure hyperscaler cloud model for sensitive workloads. AI leaders expect hybrid environments to serve as their primary inference location over the next 24 months. This preference for hybrid setups places them ahead of other deployment models, including fully on-premises, pure hyperscaler cloud, or "neoclouds" and AI factories. This shift is driven by a lack of confidence in current inference security; only 44% of AI leaders report being “very confident” that data is protected during the inference process, while 52% remain only “somewhat confident.” To bridge this gap, 38% of respondents identified advanced infrastructure security, specifically Confidential Computing, as the primary solution to increase their deployment confidence.
Key Takeaways
- 61% of regulated enterprises continue to send sensitive data to public cloud AI services despite 93% expressing concerns about exposing proprietary intellectual property.
- Regulatory compliance is now the leading obstacle to AI deployment on sensitive data, cited by 46% of surveyed AI leaders.
- 95% of AI leaders identify sovereign AI as a critically important or important component of their enterprise AI strategy.
TechInsyte's Take
In our view, the Fortanix findings illustrate a critical "trust deficit" that threatens to stall the ROI of enterprise AI investments. The fact that 41% of organizations have completely blocked AI deployments due to security concerns suggests that the current public cloud AI paradigm is fundamentally misaligned with the risk profiles of highly regulated sectors. We see this not merely as a technical hurdle, but as a strategic pivot point. The overwhelming preference for hybrid environments and sovereign AI indicates that the industry is moving toward a "decentralized intelligence" model. For C-suite leaders, the takeaway is clear: the era of blindly uploading sensitive datasets to third-party LLMs is ending. Success in the next 24 months will likely depend on the ability to implement hardware-enforced security, such as Confidential Computing, to provide the verifiable proof of protection that current cloud models lack.
Questions & Answers
How is the lack of data protection during processing affecting AI project timelines?
The lack of protection is causing significant delays; 59% of organizations have postponed AI projects in the last year due to privacy or security concerns, and 41% have entirely halted deployments because of security risks.
What is the primary barrier preventing the use of sensitive data in AI models?
Regulatory compliance has overtaken direct data exfiltration risks as the single largest obstacle, with 46% of AI leaders citing compliance as the main challenge to deploying AI on sensitive data.
What deployment model do AI leaders expect to dominate the next two years?
AI leaders expect hybrid environments to become the primary location for AI inference over the next 24 months, outperforming pure hyperscaler cloud and fully on-premises models.
Which technology is most highly regarded as a solution for increasing AI deployment confidence?
Advanced infrastructure security, specifically Confidential Computing, was cited by 38% of respondents as the top method to increase confidence in deploying AI on sensitive data.
Source: Fortanix
