Delivering on AI has become both a buzzword and a pressure point for organisations globally. There is also a gap between its adoption and its ability to deliver value. The McKinsey State of Organisations 2026 report surveyed executives across 15 countries and 16 industries, and it found that 88% of companies were investing in AI, but 81% were not seeing any meaningful gains from this AI[1]. This is against the backdrop of aggressive AI spend, with IDC’s April 2026 update saying that global AI infrastructure spending reached $89.9 billion in the last quarter of 2025[2]. There remain barriers to progress, and one of the most persistent is data. 

“Data and governance readiness are the primary blockers to successful AI integration and adoption,” explains Mandla Mbonambi, CEO of Africonology. “Without trusted data foundations and an architecture that supports AI and its governance, financial institutions will continue to struggle with limitations around their AI capabilities. They will also battle to see the returns the technology promises.” 

Financial institutions are dealing with data spread and weight. Customer information is scattered across core banking systems, cloud platforms and regional data centres, while every AI initiative begins with the slow and expensive work of stitching this together. The traditional route to unified data is extraction and duplication, and in financial services, that route runs directly into regulation. Data residency laws mean engineers must build compliant pipelines to copy data between environments and then maintain two versions of the truth. Time and trust both suffer as a result. 

This is where architecture decides the outcome. One methodology is taking a different approach by flipping the traditional model and querying data where it sits rather than moving it. This approach creates a single access layer across on-premises and cloud systems, preserving data residency because the data never leaves its jurisdiction, maintaining a single version of the truth. However, federation alone doesn’t make this data usable for AI. The layer that matters is the one that organises the data into certified, reusable products that companies can find, trust and understand. 

When an AI agent sits on top of that layer, it has the schema, the language and the trusted data it needs to answer questions accurately. As the local Starburst partner, Mbonambi explains the business value it brings: “The context layer allows the business to organise the data that comes from hundreds of data sources into a very small number of data products. Then companies have the schema, the language and the data they can trust.”

Now, companies can feel the difference in how they can answer questions. Legacy business intelligence environments can respond to the questions they were built for and nothing more, but when AI is integrated properly, this story changes. A compliance officer can pull a regulator’s request in minutes, spot that most loan declines share a single reason code and interrogate the data behind it. All without waiting on a data engineering queue. Governance concerns can be answered within the same architecture, and suddenly institutions can delve into data in new ways that highlight friction points, bottlenecks, opportunities and risks. 

“Every query is traceable to the person who asked it, costs can be attributed to individual teams, and access controls can operate at a row and column level, so a junior analyst asking the same question as a compliance officer can only see what their role permits,” says Mbonambi. “This is a mindset change that gives companies a fighting edge when it comes to AI.” 

Benefitting from AI adoption is a foundational problem that goes beyond the model and the system and deep into the architecture of the organisation. Financial institutions need to fix the data layer first using tools that connect data where it lives, organise it into trusted products and build governance in from the start rather than bolting it on later. 

For South African banks, the stakes are sharper than the global averages suggest. The sector carries decades of legacy infrastructure, operates under tight data residency and privacy obligations, and faces a customer base moving faster than most core systems were designed to serve. The institutions that treat AI as a data problem before a model problem will move from pilot to production in months. Those who keep funding proofs of concept on broken foundations will keep paying for capabilities they never see. 

"The race won't be won by the bank with the biggest AI budget," concludes Mbonambi. "It will be won by the one whose data is ready to answer the questions it has to ask."