A significant proportion (85%) of EMEA decision-makers have already limited or redesigned AI use cases because of rising token consumption, while 82% say data residency, sovereignty or jurisdiction-specific requirements influence where AI workloads are deployed.
These are among the findings of the State of Enterprise AI 2026: The New Rules of Scale report, based on a survey commissioned by Cloudera and conducted by Wakefield Research in September 2026. The survey covered 1 500 global IT leaders and decision-makers, including 525 in EMEA.
Governance was the most frequently cited barrier among EMEA respondents, at 75%, compared with 45% who identified token and inference costs. Only 28% said AI was advancing faster than their controls could accommodate.
Warren Olivier, regional VP for Africa at Cloudera, said governance requirements often slowed implementation because each new use case required approval of data sources, access to outputs and procedures for dealing with errors.
“Organisations take those warnings seriously, and therefore the governance process slows them down. Each new use case needs someone to approve where the data comes from, who sees the output and what happens when the output is wrong.”
These processes are often handled manually, project by project, and may need to be repeated for each cloud or data centre used by a workload. Nearly half (49%) of respondents said governance and compliance costs had increased faster than anticipated. A further 58% said moving workloads between cloud and on-premises platforms required costly application redesigns.
Cloudera's separate Great AI Re-Architecture Survey, conducted in June 2026, found that 83% of South African organisations had moved at least some AI workloads away from public cloud – the highest rate recorded in that study – and that 47% had moved an AI workload because of higher or more volatile costs.
Separating data from AI models and the computing environments in which they run can allow organisations to change models or computing resources without rebuilding applications, provided the underlying data and governance arrangements remain in place. Governance rules that apply across environments can also follow workloads between platforms.
AI deployment projects withdrawn
Globally, 98% of organisations had withdrawn an AI deployment because of hallucinations, errors or unexpected model behaviour in production.
Olivier said this statistic partly reflected organisations' improved ability to monitor AI deployments, but pointed to other risks.
“You can only pull back on a deployment you’re watching. A team that sees a model misbehave and switches it off has therefore done its job. The worrying part is how close that number is to everyone. A pilot runs on clean, curated data and a narrow task, but production gives the model years of messy records and employees who use it in ways nobody explicitly tested. A hallucination rate that may have seemed tolerable in a demo can suddenly become a serious liability if the output has not been verified for credibility,” said Olivier.
The report found that, globally, 49% of respondents would strengthen governance from the outset if they could restart their AI programmes. A further 46% would invest in data quality earlier, 45% would modernise their infrastructure sooner and 45% would invest more in employee training.
The financial returns from AI remain limited for many organisations. Only 23% of EMEA respondents said AI was generating returns greater than its costs, while 57% said they were breaking even.
AI budget allocation
South African practitioners use AI every day, but only a third of organisations have built it into their systems – and budgets are still going to tools rather than the people who have to make them work.
This is according to the State of AI Delivery in South Africa report released by tech professionals network DayOne, in partnership with OfferZen, Altron AI Factory, Pepperplane and Velocity Labs, and based on 282 responses from AI practitioners and decision-makers across SA.
The report found that although 87% of respondents said they used AI tools several times a day, only 35% said their organisation had integrated AI into a system or agent set-up.
Two-thirds (67%) of practitioners at organisations with mature AI operations said AI had exceeded expectations. Among organisations still exploring AI, the figure was 11%.
Liesel Bester, founder of DayOne, said: “Almost everyone is using AI, so usage is no longer the thing that separates the organisations getting results from the ones that aren't. The difference is whether it's been built into how the organisation works. A team that has wired AI into its systems gets compounding value out of it. A team of individuals each prompting a chatbot gets a productivity bump, and not much more."
Few AI experiments reach production
According to 72% of respondents, South African organisations attempted three or more AI initiatives in the past year. However, 59% said their organisations had only one or two initiatives in production, while 44% had tried an AI initiative and subsequently abandoned it.
Respondents at mature organisations were more likely to report abandoning an AI initiative than those at less mature organisations.
"The mature organisations aren't failing more, they're deciding faster," said Bester. "If a pilot isn't showing the promise they expected, they kill it and move on, or double down on the thing that is working. The less experienced organisations often don't yet know what a pilot that's going nowhere looks like, so it just sits there."
AI budgets favour tools over training
The report found a gap between leadership's stated AI ambitions and available budgets. Overall, 64% of respondents reported a mismatch, rising to 84% at larger organisations.
Where ambitions and resources were aligned, 61% said AI had exceeded expectations. Among organisations with a significant gap, the figure was 33%.
The survey also found that 84% of organisations allocated AI budgets to tools and licences, while only 38% spent money on training. Nearly half (49%) had no dedicated AI upskilling budget for the following year.
Respondents identified prompt engineering (29%) and evaluation and testing (22%) as the skills their organisations lacked most.
"Organisations are buying the tools and assuming the skills will follow," said Bester. "They don't. The gaps people are reporting are the hands-on ones: how to brief the tool properly and how to check whether what it gave you is any good. That's training, not procurement."
Unclear goals remained the most frequently cited reason AI failed to meet expectations. A further 58% of practitioners said they would set their expectations differently if starting their AI programmes again.
Francisco Mateo-Sidron, head of EMEA at Cloudera, said: “Scale is exposing a new reality for enterprise AI: what works today may not make economic or operational sense tomorrow. Costs change, models evolve, regulations shift and the right place to run a workload changes with them. Leaders need an AI strategy built for that reality, one that gives them the freedom to adapt without giving up control of their data, governance or spend.”
Regulatory requirements also continued to change, Olivier said. “South Africa's draft national AI policy was withdrawn this year, and a revised version is still being drafted. In the meantime, local organisations must govern their AI use in accordance with the Protection of Personal Information Act and the requirements of their respective sector regulators.”
Share
