Half of South Africa's AI practitioners say their company has no budget to train staff in AI next year. In DayOne's first State of AI Delivery ZA 26 report, 49% of the 282 people surveyed said their organisation has not budgeted for AI upskilling in the next 12 months, and another 18% were not sure. Only 33% said there is a budget for it.

Where companies do spend on AI, tools come first. 84% of respondents said their AI budget covers tools and licences, against 38% who said it covers training and upskilling, 26% internal headcount and 14% external partners or vendors. Respondents could choose more than one, so these figures show how many companies spend on each item, not how the money is split.

Training budgets stay thin even as companies get more experienced with AI. Among companies still exploring AI, 67% have no training budget. At no stage of maturity do even half have one: 44% of companies scaling AI have a budget, and among the most mature companies, 40% have one while 49% do not. Where budgets exist, respondents described courses, vendor certifications and, often, simply more access to tools. Time set aside to learn, such as hackathons or learning days, came up less often, and governance training barely came up at all. That matters because the skills gaps holding companies back are practical.

Asked about the biggest skills gap blocking AI delivery, 29% named prompt engineering, meaning how to give AI clear instructions, and 22% named evaluation and testing, checking whether what AI produces is right. AI governance or ethics followed at 18% and change management at 17%. Deep technical skills barely featured: data engineering at 5% and machine learning operations at 4%. Prompt engineering was the top gap at every stage of maturity.

The report's sponsors were blunt. "The money goes to tools, and the people using them are left to work it out," said Rob McEwan of Velocity Labs. Nemanja Stabic of Digiata said companies putting money into tools rather than people are "buying capability they cannot yet absorb".

OfferZen said the fix is giving teams time and training, "not just licences and hoping for the best", and has published a practical AI framework for teams. The 282 respondents were surveyed between 2 and 16 October, and 66% work at companies of 50 people or fewer.

Companies are paying for AI tools and leaving their people to figure them out alone, which is how expensive licences end up underused, a pattern we saw in which tools practitioners actually choose.

The good news is that the skills gaps at the top of the list do not need big budgets to close. Writing clear instructions for AI and checking its output are skills people learn by practising on real work, with someone showing them what good looks like.

An afternoon a week set aside for the team to learn, share what works and review each other's AI output would cost less than many licences, and would make those licences far more useful. That is the thread from last Monday's look at small business AI too: the tool is the easy part, the people are the work.

You might also like our piece on women's ChatGPT use in SA, AI agent swarms in Minecraft, and POPIA and AI at work.

Get more SA tech and business news and subscribe to The Open Letter.