Somewhere in a South African head office this quarter, an AI proof-of-concept that impressed the executive committee six months ago is being switched off or shelved.
It demoed beautifully, but it never made it into production. And the post-mortem will probably blame the model, the vendor, the hype or even AI technology itself.
That instinct is almost always wrong. When an enterprise AI project fails, it is rarely the intelligence that lets you down. The core of the problem is that it is far easier to build a controlled, impressive demo than it is to scale it into production software.
Projects usually collapse due to unexpected infrastructure and computing costs, and a lack of true enterprise-grade readiness. Teams often treat these projects as technology-first experiments, measured by technical accuracy rather than tangible business value, while entirely neglecting the human change management and real-world workflow quirks needed for everyday employee adoption.
The experimental excitement and momentum then give way to reality. That reality-check, however, includes both a rude awakening and a difficult pill to swallow: in a more holistic sense, the project’s failure reflected deeper shortcomings in AI readiness across the business.
The pilot held a mirror up to the organisation, the organisation did not like what it saw, and the mirror (AI) took the blame – then got smashed.
The scale of this is in no way anecdotal. MIT's NANDA initiative reviewed more than 300 enterprise deployments for its 2025 study, The GenAI Divide, and found that roughly 95% delivered “no measurable impact on the bottom line”. Only about one in twenty reached production with real value.
MIT puts much of that down to weak workflow integration and systems that never learn. In my experience across African financial services, telecommunications and the public sector, there is a more basic thread running beneath it all: fragmented, ungoverned, poorly integrated data.
The uncomfortable part is that most organisations do not know this is their problem, because they are confident about data they cannot actually reach. In Cloudera’s own Data Readiness Index 2026, 89% of EMEA IT leaders claimed they had complete visibility into where their data resides, yet only 26% of them said that data was fully governed.
The distance between what leaders believe they can see and what they can actually govern is exactly the reflection in the mirror we're talking about. In a similar vein, Gartner in 2025 already predicted this paradox to lay waste to 60% of corporate AI projects by the end of 2026 because they lack AI-ready data.
Faced with a stalled pilot, the temptation is to smash the mirror and blame AI-hype for overselling a tool that isn’t useful and promptly scale back AI ambitions. In other cases, they might swap the model, simply spend more, or even bolt on an agent.
It feels like progress and it ends up changing nothing, because the flaw being reflected sits upstream. There is also a nasty twist in the economics: a flashier AI is far more expensive to run (and still won’t make the picture in the reflection prettier automatically).
This is where tokenisation stops being a little billing footnote and becomes the whole story. Generative AI is metered by the token. You pay for the volume of text and content going in and coming out, not for whether the answer was worth anything. On a clean, governed foundation, that is affordable. On a messy one, it is a slow leak that turns into a flood scarily quickly.
Fragmented data forces you to stuff more context into every prompt to compensate. Data nobody trusts pushes teams to route everything to the largest, most expensive model, because the cheaper ones cannot be relied on with inputs no one has vetted. Unreliable answers trigger retries, and every retry is another metered call.
Agents also make this cost growth exponential rather than linear. An AI agent does not ask once and stop. It reasons, re-reads, calls tools and loops. Anthropic's own engineers found that a single agent burns through about four times the tokens of an ordinary chat, and a multi-agent system roughly fifteen times as many, before anything even goes wrong.
Point that machinery at a weak data foundation and a bad decision isn’t the only thing that’s going to happen. It produces a bad decision at speed, then invoices you generously for the privilege.
In its latest lookahead, Gartner now predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027, and the first reason it lists is escalating cost.
So the token bill is not some separate issue from the data one, either. Essentially, it is the meter that prices it. An organisation can probably debate data governance in the abstract for years, but it cannot really argue with the invoices.
The local picture drives this home for me, though. In the EMEA findings of our own index, 42% of leaders admitted that complicated access requirements were their primary barrier to using the data they can see, and only about a third have their data sources fully integrated across environments.
Add the pressures particular to this market – POPIA obligations, tight budgets, scarce skills – and running expensive demonstrations on shaky foundations becomes a cost most South African organisations cannot absorb for long.
The fix is unglamorous, which is precisely why it works. Before buying more intelligence, make the data beneath it accessible, integrated and governed, so that governance travels with the data instead of living inside one provider's platform, and so the Protection of Personal Information Act (POPIA) leaves you with an audit trail rather than a liability.
Then bring the AI to that governed data, hold your options open through open standards, and let cost become something you can plan for instead of something that plans for you.
And none of that is as exciting as a model rollout or an agent launch. But the organisations getting real, everyday value from AI in this country are not the ones running the most pilots with the most resources. They are the ones that fixed the plumbing first, regardless.
You can’t blame the mirror for what is reflected and a fancier one won’t change anything either. What that means is: a successful AI project starts with the unglamorous parts. It starts with the data – every time.
Share