Businesses shift to private AI models to retain data control

Businesses are rethinking how much data they hand over to public AI models, shifting instead to smaller, private systems built on their own infrastructure to keep control of their information.

“Organisations have learnt that when they push proprietary data into a public model hosted in another jurisdiction, they surrender a measure of control over their own intellectual property,” said Warren Olivier, regional vice-president for Africa at Cloudera, an enterprise data and AI platform company that helps organisations manage data across hybrid and multi-cloud environments.

Olivier said the shift toward smaller, private models is one of the clearest trends Cloudera is seeing, with enterprise AI moving from early experimentation with large public models to a phase defined by control.

“When that model is updated without warning, the applications they carefully built on top of it can break overnight. Private AI reverses that equation,” he said, adding that organisations are increasingly running models closer to where their data sits rather than moving data to external platforms.

He said smaller, open-source models have made this shift more practical, allowing companies to train systems on their own data and run them on more predictable infrastructure than pay-per-use AI services, with some approaches cutting inference costs by up to 75%. Hybrid platforms such as Cloudera support this by letting different models, including public, commercial and open-source systems, run across environments on governed data without leaving the organisation.

The shift comes as many organisations continue to struggle with the quality and accessibility of their own data. According to Cloudera’s 2026 data readiness index, which surveyed 1,270 IT leaders globally, 84% of respondents were confident in the accuracy of their data. However, only 18% said it was fully governed, and 79% reported that data-driven projects failed because they could not access information across different systems.

Olivier said Cloudera’s approach centred on giving organisations a single view of their data across on-premises, cloud and edge environments, applying consistent security and governance and enabling AI to run where the data already resides rather than requiring large-scale migration.

He said South African banks, telecommunications companies and other large organisations are already adopting these approaches as they move AI into production. He added that this is happening as organisations work to meet regulatory requirements such as the Protection of Personal Information Act (Popia).

According to Cloudera, AI adoption is now widespread, with 96% of organisations reporting integration into core business processes and 85% saying they have a clear data strategy. But widespread adoption has not translated into consistent value.

“Both are true, and that is precisely the problem,” Olivier said, noting that the survey found that nearly four in five IT leaders say their AI initiatives are constrained by limited access to data. When asked why AI investments fall short, 22% of respondents cited data quality as the leading factor, ahead of cost overruns and poor workflow integration.

The priority now is ensuring that the economic value AI generates in South Africa stays here

A survey by Gartner, an American research and advisory firm, has predicted that 60% of AI projects lacking AI-ready data will be abandoned through 2026.

The result, Olivier said, is that AI is becoming inevitable for organisations, but meaningful value is not.

Those seeing consistent returns were not necessarily the ones running the most pilots but those that prioritised the less visible work of building strong data foundations. Others, he said, were effectively running “expensive demonstrations”.

Fragmented data remains the single biggest obstacle to scaling AI, Olivier said. According to the Europe, Middle East and Africa results of Cloudera’s data readiness index, only 34% of organisations report that their data sources are fully integrated across hybrid environments.

Olivier said this fragmentation creates latency, drives up dollar-denominated egress fees at a time of rand volatility and leaves AI systems making decisions on incomplete information.

Governance, he said, should be treated as an enabler rather than an obstacle, allowing organisations to move data across environments without being locked into a single vendor.

Olivier said Cloudera’s regional focus is on sectors where data control is most critical, including financial services, telecommunications and the public sector. Key issues include data sovereignty under Popia, dollar-denominated cloud costs and moving organisations from stalled pilots to AI in production.

“South Africa is not waiting for the AI era to arrive,” Olivier said, adding that the country’s organisations are early adopters of the technology. He said the priority now is ensuring that the economic value AI generates in the country stays here.