Unsuccessful artificial intelligence pilot projects observed in South African organizations are not a result of AI being a poor business technology. On the contrary, they reflect a deeper problem that AI reveals, and it is the usage bill that makes this reflection increasingly obvious.
In South African headquarters this quarter, the concept of AI, which impressed the executive committee six months ago, is being shut down or postponed. Although the demonstration looked magnificent, the project was never implemented into the working environment. Post-analysis will likely blame the model, the vendor, the hype, or the AI technology itself.
However, this instinct is almost always wrong. When an AI project fails in a large company, the intellectual component itself is rarely to blame. The main problem is that it is much easier to create a controlled and impressive demonstration than to scale it to production software. Projects often collapse due to unforeseen infrastructure and computing costs, as well as a lack of true enterprise-level readiness.
Teams often view these initiatives as technology-focused experiments, where success is measured by technical accuracy rather than tangible business value, completely ignoring human change management and the peculiarities of the real workflow required for daily employee adoption. Experimental enthusiasm and momentum then give way to harsh reality. This reality check includes both a sharp awakening and a bitter pill: more broadly, the project failure reflected deeper shortcomings in the business's readiness to work with AI.
The pilot project acted as a mirror to the organization, and the organization did not approve of what it saw, blaming the mirror (AI), which was then broken. The scale of this problem is not merely anecdotal. The NANDA initiative from MIT tested over 300 implementations in large companies for its 2025 study, 'The GenAI Divide,' and found that approximately 95% did not have 'measurable impact on net profit.' Only about one in twenty reached production with real value. MIT explains most of this through weak workflow integration and systems that never learn.
From personal experience working in the financial sector, telecommunications, and public sector in Africa, there is a more fundamental thread underlying everything: fragmented, unregulated, poorly integrated data. The unpleasant fact is that most organizations do not realize this is their problem because they are confident in the data they cannot actually access. According to the 2026 Data Readiness Index by Cloudera, 89% of IT leaders in the EMEA region claimed to have a full overview of where their data is located, yet only 26% stated that this data was fully governed. The gap between what leaders believe they can see and what they can actually control is precisely the reflection in the mirror being discussed.
Similarly, Gartner predicted in 2025 that this paradox would destroy 60% of corporate AI projects by the end of 2026 due to a lack of AI-ready data. Faced with a stalled pilot, there is a temptation to smash the mirror and blame the hype around AI for overpromising a useless tool, and then quickly scale back AI ambitions. In other cases, they might replace the model, simply spend more money, or add an agent. This seems like progress, but ultimately changes nothing because the defect being reflected is upstream. There is also an unpleasant economic twist: brighter AI costs much more to operate (and still won't automatically make the picture in the reflection any better).
This is where tokenization stops being a minor expense and becomes the whole story. Generative AI is billed by tokens. You pay for the volume of text and content entering and exiting, not for whether the resulting answer was worth anything. With a clean, governed foundation, this is affordable. On a messy foundation, it is a slow leak that frighteningly quickly turns into a flood.
Fragmented data forces more context to be inserted into every prompt to compensate. Untrusted data pushes teams to route everything through the largest and most expensive model, as cheaper models cannot be relied upon with unverified input. Unreliable responses cause retries, and every retry is another paid call. Agents also make cost growth exponential, not linear. An AI agent doesn't ask once and stop. It reasons, rereads, calls tools, and operates cyclically. Engineers at Anthropic found that one agent consumes about four times more tokens than a standard chat, and a multi-agent system about fifteen times more before something goes wrong. Point this machine at a weak data foundation, and a bad outcome will not be the only consequence. It produces a bad solution at speed, and then generously bills you for the privilege.
In its latest forecast, Gartner now predicts that over 40% of agent AI projects will be canceled by the end of 2027, with rising cost being the primary reason cited. Thus, the token bill is not a separate issue from the data issue; it is essentially a metric that prices it. An organization may be able to discuss data governance in abstraction for years, but it cannot argue with the invoices.
The local picture confirms this for me. In the results of our own EMEA index study, 42% of leaders admitted that complex access requirements were their main obstacle to using visible data, and only about a third have fully integrated data sources across different environments. Add to this the pressure characteristic of this market—POPIA obligations, limited budgets, skills shortages—and conducting expensive demonstrations on shaky foundations becomes an expense that most South African organizations cannot sustain for long.
The solution is simple, and that is why it works. Before buying more intelligence, make the underlying data accessible, integrated, and governed so that governance follows the data, rather than residing within a single vendor's platform, and so that the Protection of Personal Information Act (POPIA) leaves you an audit trail, not liability. Then, bring AI to this governed data, maintaining open standards, and let the cost become something you can plan for, not something that plans you. And none of this is as exciting as launching a model or launching an agent. But the organizations achieving real, day-to-day value from AI in this country are not those running the most pilots with the largest resources. They are those who first fixed the plumbing, regardless of everything else.
You cannot blame the mirror for what it reflects, and a fancier mirror will not change anything. This means: a successful AI project starts with the humble parts. It starts with data—always.
