Haider Aziz, Managing Director, META at VAST Data, has penned a compelling op-ed exclusively for tahawultech.com, in which he highlights that vector databases designed to serve as the connective tissue between LLMs and real-time enterprise knowledge are no longer fit for purpose.
Across the Gulf, we’re witnessing a once-in-a-generation transformation. From Abu Dhabi’s partnership with OpenAI on the Stargate initiative to the continued momentum of homegrown large language models like Falcon and Jais, and more recently the news that an AI system will become an official advisory member of the UAE Cabinet from 2026, the region isn’t just adopting
AI, it’s actively shaping its future.
Government ministries have appointed AI leads, sovereign funds are underwriting compute capacity at national scale, and public-private collaborations are redefining what’s possible across energy, healthcare, and finance. In this context, AI infrastructure isn’t just a technical consideration. It’s becoming national infrastructure.
But for all the progress in model development and GPU deployment, one critical component is quietly falling behind: the retrieval layer that underpins modern AI systems. More specifically, the vector databases meant to serve as the connective tissue between large language models and real-time enterprise knowledge. As it stands, they just aren’t fit for purpose.
From Prototypes to Production
In the early days of Retrieval-Augmented Generation (RAG), use cases were modest. Enterprises indexed a few internal PDFs, connected them to an LLM, and deployed chatbots or semantic search tools with impressive early results. These pilots proved the potential of retrieval to ground AI outputs in enterprise-specific context.
But today, the demands are very different.
Organisations across sectors, from government to financial services to energy, are building AI agents to operate across petabytes of sensitive, constantly changing data. The Middle East, in fact, may have an advantage here: with less legacy infrastructure than many mature markets, there’s greater freedom to build from first principles. But even in forward-leaning environments, retrieval at scale introduces real architectural complexity.
And globally, this is where the stress begins to show.
The shift from controlled pilots to production-scale systems has exposed the limitations of retrieval stacks not designed for real-time indexing, fine-grained access control, or trillion-vector search performance. What worked in a lab no longer works in the enterprise.
We’ve proven that retrieval can work in contained, curated environments: the toy problem. But building systems that work across live, complex, permissioned enterprise data? That’s the real challenge.
When Vector Systems Become the Bottleneck
Most vector databases were not designed to handle the size or complexity of modern enterprise data. In many current setups, data is embedded and indexed in batches—sometimes hourly, sometimes daily. This creates the impression of a responsive system, but in reality, it often serves answers based on outdated information.
For industries like energy, finance, or healthcare, where conditions and decisions can shift by the second, this delay matters. A lot.
Access control is another area where today’s retrieval systems fall short. Enterprise environments often have strict policies around data permissions and auditing. Yet many retrieval pipelines rely on separate systems to handle those permissions. This means access rules are not always enforced at the moment of the query, and that creates a risk of exposing information that a user shouldn’t see.
Then there is the issue of scale. While some global AI developers have built systems capable of indexing the public internet, most enterprises already manage hundreds of petabytes of internal data. Few of today’s vector tools were built to support that kind of volume, especially when you also need speed, accuracy, and secure access.
These challenges all come into play as organisations shift from pilot projects to full production AI systems. When AI moves from the edge of the business to the centre, the expectation is not just that it works, but that it thinks. And thinking, in this context, depends on what the AI can reach.
If the AI can only access a small, handpicked set of documents, its answers will be limited. Worse, when questions go beyond that narrow dataset, the model starts to guess, filling in gaps with incorrect or misleading information. This erodes confidence.
By contrast, when every contract, meeting transcript, design spec, and operational log is embedded and searchable, the AI can draw from the full depth of institutional knowledge. Only with this kind of access can it uncover new insights, connect patterns, and offer guidance that feels relevant and trustworthy. But access alone isn’t enough.
The system also needs to know who is asking. Every AI response must be tied to the identity of the user making the request. That user should only see what they are authorised to see, nothing more. This is why it is so important for vector-based retrieval to follow the same access control policies as the source systems. It must enforce those rules automatically, in real time, for every single query.
That is the foundation of trust. And it is one of the most critical elements in making AI enterprise-ready.
Regional Ambition Requires Architectural Rethinking
The UAE’s AI strategy is not incremental. From the work of G42 across genomics, space, and sovereign clouds, to the AI-driven automation efforts in NEOM and Saudi Aramco, regional leaders are setting the bar high. These aren’t controlled experiments, they are production-scale deployments that demand architectural maturity.
To meet this ambition, we need to move beyond the previous infrastructure playbook as follows:
- Retrieval must be embedded within the data platform itself, not bolted on as a secondary service.
- Vector indexing must operate in real time, not on batch timers.
- Permissioning must be enforced atomically, from the data source through to the AI output.
- Retrieval systems must be capable of managing trillions of vectors, not just curated datasets.
The good news is that this is achievable, but it requires moving past the prototype phase and rethinking how data infrastructure is built for the age of AI.
What Comes Next
In a region that is investing billions in sovereign AI capabilities, the retrieval layer cannot remain an afterthought. As more government and enterprise teams in the UAE and across the GCC shift from proof-of-concept to production deployment, the limitations of today’s vector systems will become more apparent. The question is not whether retrieval will scale, it’s whether organisations are ready to scale with it.
To get there, we need infrastructure that treats vectors not as sidecar metadata, but as core intelligence, integrated, permission-aware, real-time, and built for the scale of national ambition.
As many of you will know, in this part of the world, AI isn’t just another tool in the stack. It’s a strategic priority for the region. And, we can no longer afford to build it on brittle foundations past their prime.
