Most South African organisations we engage with say they are exploring artificial intelligence (AI) but far fewer can explain how their existing data architecture will sustain it. The current AI conversation tends to focus on models, use cases and productivity gains. Boards want to know how generative AI will reshape operations. Executives ask which processes to automate first. Technology teams experiment with pilots and proofs of concept.
Yet in many environments, the core data estate remains fragmented, duplicated across systems, inconsistently governed and slow to reconcile. That is where most AI initiatives quietly encounter difficulty and in some cases just fail.
But, AI does not fail because the model is insufficiently sophisticated. It fails because the data beneath it is incomplete, delayed, contradictory or poorly contextualised. When intelligent systems rely on untrusted foundations, their outputs become unreliable. In regulated or high-stakes sectors like finance and healthcare, that is not a minor inconvenience, it is a huge risk.
If organisations are serious about AI, they need to go back to basics.
From tools to infrastructure
For years, AI was treated as an overlay. A chatbot here. A predictive model there. Analytics layered onto operational systems after the fact.
That approach enabled experimentation without requiring structural change. Data could be copied into a lake, processed in batches, and reconciled later. Governance often followed behind the technology rather than being designed into it.
Layered systems, data exports and batch processing are not architectural failures. They are a practical response to systems built at different times, for different purposes, under different constraints. In many environments, they remain necessary. What has changed is the expectation placed on AI.
Generative and agentic AI systems are now expected to work across structured and unstructured data transactions, documents, images, events and to assist with decisions that increasingly require both precision and speed. Rather than eliminating layered architectures, organisations must consolidate their data capabilities into a multimodal platform — one capable of supporting structured and unstructured data and operating across both batch and real-time processing.
The objective is not simplification for its own sake. It is context.
The Significance of Context Engineering
The effectiveness of generative and agentic AI is determined less by model sophistication than by whether the right data is available, in the right form, at the right time.
This is a matter of context engineering. Most organisations operate across multiple technology layers. Core transactional systems serve one function. Analytics platforms serve another. Integration tools connect them. Cloud services extend capability, while legacy environments continue to run critical processes.
Individually, these components may perform well. Collectively, they require deliberate coordination. When AI systems are introduced without that coordination, they inherit latency, inconsistent definitions and fragmented identity resolution. At pilot stage, these limitations can be masked through manual curation and constrained scope. At scale, they become structural.
Addressing this does not require dismantling layered systems. It requires interoperability standards, consistent identity resolution, and the ability to process events as they occur — while retaining batch capabilities where appropriate. It involves unifying structured and unstructured data without sacrificing performance. It demands that governance, auditability and access controls are embedded at platform level rather than retrofitted at application level.
AI that influences real-world decisions must operate across a trusted, multimodal data foundation. Infrastructure, in this context, is the disciplined orchestration of data so that AI systems can function with speed, resilience and defensible context.
Trust as an architectural property
As AI systems begin to influence operational decisions, trust becomes central. While trust is often discussed in ethical or regulatory terms, it is first and foremost an engineering question. Can the system explain how it arrived at a recommendation? Is there an audit trail? Are data sources known and validated? Are roles and permissions enforced consistently?
In environments with strict data protection requirements and hybrid estates, the stakes are higher. Systems must maintain integrity across cloud and on-premises infrastructure. They must support secure access while preserving performance. They must function reliably even where connectivity is constrained.
Trust cannot be added after deployment. It must be designed into the way data is stored, processed and orchestrated from the outset.
A quieter race
The broader conversation around AI has become preoccupied with model comparisons and benchmark scores, as though performance differentials alone determine value, while far less attention is given to whether the underlying systems are coherent, resilient, and well-governed enough to carry those capabilities into production.
Organisations that align transactional systems, analytics and AI around a unified data foundation will be able to scale intelligent capabilities steadily. Those that continue to layer AI onto fragmented estates will remain in pilot mode, revisiting the same integration and data quality challenges with each new initiative.
The difference will not always be obvious at first. Both groups may produce impressive demonstrations. The distinction emerges over time, in reliability, in governance, and in the ability to move from isolated projects to enterprise-wide adoption. Ask yourself: will my data architecture sustain intelligent systems as part of everyday operations?
If AI is to become embedded in core operations rather than remain confined to experimentation, the effort must start with system design, data integrity and architectural discipline, since algorithms on their own are insufficient; it is the structure around them that determines whether they can be trusted, governed and sustained over time.
By Reven Singh, Sales Engineer, InterSystems South Africa
