Why AI Projects Fail in South Africa: the foundations problem behind the spend

AI projects in South Africa mostly fail for reasons that have nothing to do with the model. Accenture South Africa reports that 89% of local executives say legacy infrastructure limits their agility and adds to technical debt, while only 24% are actively addressing it. The tools work. The data, processes and systems underneath them do not.

That gap is the story of South African AI in 2026, and it is now being said out loud by the people selling the transformation. Below are the figures, what they mean if you run a smaller business rather than an enterprise, and the order of work that actually gets a project over the line.

What the latest South African figures say

Writing in July 2026, Kgomotso Lebele, Country Managing Director for Accenture in South Africa, set out how far local AI ambition has outrun its foundations:

• More than 80% of South African enterprises increased ICT budgets in recent years.
• IT spend grew by 11% between 2022 and 2025, while AI investment rose by 80% over the same period.
• By 2025, technology spend had outpaced revenue growth by roughly 115 percentage points.
• 89% of executives acknowledge that legacy infrastructure limits agility and contributes to growing technical debt, but only 24% are actively addressing it.
• Managing that debt consumes between 20% and 40% of IT budgets.

Read those together and the picture is blunt. Spending on AI went up by 80% while spending on the systems it has to run on went up by 11%. Nearly nine in ten leaders can see the problem. Fewer than one in four are doing anything about it.

Why spending more does not fix it

The instinct when an AI pilot disappoints is to buy a better tool. It rarely works, because the failure is almost never in the model. It is in what the model is standing on.

An AI system is only ever as good as the inputs it can reach and the process it plugs into. If your customer history lives in three places and one person's phone, no assistant can answer a question about a customer. If nobody has written down how a quote actually gets approved, no automation can run the approval. The tool inherits your operation. A mess automated is still a mess, only faster and more confidently wrong.

This is also why AI budgets can rise 80% and produce disappointment. The money went to the visible layer. The invisible layer, where the data quality and the process definitions live, got the leftovers.

The four patterns behind failed AI projects

1. No clear job. The project starts from "we should use AI" rather than from a specific, repeated, expensive task. Without a defined job, there is nothing to measure and no point at which anyone can say it worked.
2. Data the system cannot reach. The information the tool needs is spread across chat threads, spreadsheets, inboxes and someone's memory. Nobody solves this by adding another tool on top.
3. An undocumented process. If a task is done slightly differently by each person, there is no process to automate, only a habit. Writing it down is the actual work, and it is unglamorous enough that it gets skipped.
4. No owner and no measure. The pilot has a champion but no one accountable for the outcome and no baseline number to compare against, so it quietly ends without ever being declared a failure.

None of those are AI problems. Every one of them exists before the first tool is bought, and every one of them is cheaper to fix than to work around.

What to do instead: map the work before you buy the tool

The order matters more than the technology choice.

1. Pick one expensive, repeated task. Something that happens weekly, takes real hours, and has a visible cost when it goes wrong.
2. Write down how it is done today. Every step, every handoff, every place someone has to look something up. This is where most of the waste turns up, usually before any AI is involved.
3. Fix the data that task depends on. Get it into one place that a system can read. This is the step that gets skipped and the step that determines whether anything downstream works.
4. Then automate the narrow thing. One task, one owner, one number to beat.
5. Measure against the baseline you took in step 2. If you did not take a baseline, you cannot prove a result, and unproven results do not get funded twice.

We have written before about what to realistically expect in the first 90 days, and about what AI automation actually costs in South Africa. Both come back to the same point the Accenture figures make from the enterprise end: the return depends on the foundation, not the fashion.

What this means for a smaller business

The enterprise numbers describe organisations with large legacy estates. If you run a business of ten or fifty people, the good news is real: you do not have decades of accumulated systems to unpick. Your foundations problem is smaller, more visible, and fixable in weeks rather than years.

The bad news is the same trap in miniature. Buying an AI tool because a competitor mentioned one, without ever defining the job it must do, produces exactly the same stalled pilot on a smaller budget. If you are weighing outside help, our guide to choosing an AI consultant in South Africa sets out what to ask for, and the piece on agentic AI for small business covers where the technology genuinely is now.

Frequently Asked Questions

Why do most AI projects fail?
Because the failure sits underneath the model. Data the system cannot reach, processes nobody has written down, no clearly defined job and no baseline measurement will stall a project regardless of which tool is chosen.

Is South African AI spending actually growing?
Yes, sharply. Accenture South Africa reports AI investment rose 80% between 2022 and 2025 while overall IT spend grew 11%, and that technology spend outpaced revenue growth by roughly 115 percentage points by 2025.

What is technical debt and why does it matter for AI?
It is the accumulated cost of systems that were never modernised. Accenture South Africa puts the cost of managing it at 20% to 40% of IT budgets, money that is not available to build on, and it is what AI tools have to run against.

Do small businesses have the same problem?
In a smaller and more fixable form. There is far less legacy estate, but the same pattern of buying a tool before defining the job produces the same stalled pilot. The advantage is that a small operation can map its work in weeks.

What should be fixed first?
The data the target task depends on. Get it into one place a system can read, after writing down how the task is done today. Automation applied to an undocumented process just makes the inconsistency faster.

Sources

• Kgomotso Lebele, Country Managing Director, Accenture South Africa, "South Africa's AI push is running ahead of its digital foundations", SA Instrumentation and Control, July 2026: more than 80% of South African enterprises increased ICT budgets; IT spend grew 11% between 2022 and 2025 with AI investment up 80%; technology spend outpaced revenue growth by roughly 115 percentage points by 2025; 89% of executives acknowledge legacy infrastructure limits agility while only 24% are actively addressing it; managing technical debt consumes between 20% and 40% of IT budgets.

Map the work first

Every failed AI project we have looked at was decided before the software was chosen. A Business Autopsy maps your actual processes, where the data sits and which tasks are worth automating first, so the tool arrives after the decision rather than instead of it. Book a discovery call to talk it through.