AI transformation is a capability problem, not a technology problem.

The models are not the problem. By 2025, the tools are good enough for most real work. The problem is that most organisations cannot absorb them.

In its 2025 State of AI in Business report, MIT's NANDA initiative studied 300 public AI deployments, interviewed about 150 leaders, and surveyed 350 staff. The headline travelled fast: roughly 95% of enterprise generative AI pilots delivered no measurable impact on profit, and only about 5% created real value. What matters is the reason. MIT did not find a technology failure. It found what it called a learning gap, the inability of an organisation to fold the tool into how work actually happens.

That is the whole thesis in a line. AI transformation is a capability problem, not a technology problem.

The supporting evidence is consistent. The same research found that buying or partnering for AI succeeded around two-thirds of the time, while building in-house succeeded roughly a third as often. Pilots stalled because the tools could not retain feedback or adapt to context, and because the people around them were never brought along. Meanwhile, in more than nine out of ten firms, staff were quietly using their own AI tools to get work done, a shadow economy that existed precisely because the official programme had not built the capability to use the sanctioned one.

Why technology-first fails

A technology-first programme buys a platform, runs a pilot, and waits for value. It treats AI as something you install. But the value does not live in the model. It lives in the moment a person decides to use it on a real task, trusts the output enough to act, and knows when not to. That decision is a capability, and capability has to be built. Skip it, and you get an expensive tool nobody uses, and a pilot that looks good in a boardroom and collapses in the field.

What capability-first looks like

Capability-first inverts the order. You start with the people and the work. You learn how each role actually spends its day, where the time goes, and what someone would hand off if they could. You teach the skill before you reach for a tool, and you build a tool only where the skill has shown it is needed. You time the task before and after, so the value is a number, not a promise.

This is slower to start and far more durable. The few who succeed are not the ones with the best model. They are the ones who treated AI as an organisational learning problem, who gave line managers the job of driving adoption rather than a central lab, and who chose to integrate deeply rather than impress quickly.

The South African point

South Africa makes the case sharply. The country leads the continent on AI infrastructure and readiness, yet the binding constraint is not compute. When the Financial Sector Conduct Authority and the Prudential Authority surveyed more than 2,100 institutions in late 2025, the single biggest barrier to AI adoption was not technology or budget. It was a shortage of skills. The same pattern shows up in the research on South African small and medium businesses, where the obstacles are skills, weak digital foundations, and resistance to change, not the availability of the technology.

The lesson is the one MIT found globally. The model is not your problem. Your capacity to use it is. Build that first, and the technology becomes almost incidental.

Sources: MIT NANDA, The GenAI Divide: State of AI in Business (2025); FSCA & Prudential Authority, Artificial Intelligence in the South African Financial Sector (Nov 2025).