Two forecasts have been sitting next to each other for months and almost nobody has put them on the same page.
Gartner expects around 40 per cent of enterprise applications to carry task-specific AI agents by the end of 2026, up from under 5 per cent the year before.
Gartner also expects around 40 per cent of agentic AI projects to be cancelled by the end of 2027, citing cost escalation, unclear business value and inadequate risk controls.
Those usually get presented as good news and bad news. They are not two findings. They are one curve. Adoption is running ahead of measurement, and the bill arrives about eighteen months later, by which time nobody in the building can prove the thing worked.
I have spent this year arguing that case in front of executives, and the argument kept growing past what fits in a keynote. So I have written it out properly.
Read the series here: Industrialising AI
What this is
Eight essays, roughly a thousand words each, on why agentic AI programmes fail and what the ones that survive do differently. Each stands alone. Read them in order if you want the argument to build, or jump to the one that matches your problem.
The material started as a keynote written for an industry event in Johannesburg. There is no product in it, no case study, and nothing for sale.
That was deliberate then and it is deliberate here. When you tell a room their programme is likely to be cancelled and the answer turns out to be something you happen to sell, the diagnosis stops being credible.
The eight
1. We have put an electric motor on the steam shaft
Factories electrified in the 1890s and productivity did not move until the 1920s, because owners bolted an electric motor onto the same central steam shaft instead of re-laying the factory. Read this one if your AI spend has not shown up in any number that matters.
Almost every AI business case is written on rung one of Judea Pearl's ladder of causation, which is association, where tickets went down after the agent launched and so did four other things. Read this one before your next budget review.
3. Ninety-five per cent is not ninety-five per cent
An agent that is 95 per cent reliable at every step is 36 per cent reliable across a twenty-step task, every step passes, and the dashboard stays green. Read this one if your team reports step accuracy.
4. You cannot patch a prompt injection
Anything your agent reads can instruct it, and that is a property of the architecture rather than a bug awaiting a fix. Read this one if you are the person who has to sign the risk assessment.
South Africa chose a multi-regulator model rather than a single AI Act, POPIA section 71 already restricts automated decisions, and sector instruments are expected from 2027/28, which is the same window most roadmaps promise the agent economy. Read this one if you are in financial services or the public sector.
6. A gate is a condition, not a date
Three tests you can run on your AI strategy tonight, and why the later columns of your roadmap commit you to dates set by standards bodies and insurers who have never seen your plan. Read this one if you wrote the strategy.
7. You cannot buy a lower touch rate. You train for it.
Prompting is about six per cent of AI fluency. The competency that actually moves your numbers is discernment, and it is the one nobody is buying. Read this one if you own the training budget.
8. We cannot afford three years of unmeasured AI spend. That is our advantage.
Why arriving late can mean arriving better, with four serious arguments against my own position. Read this one if you work in African technology policy or you are tired of leapfrogging talk.
If you only read two
Two and five. Essay two gives you the question that exposes an undefended business case: what did it cost us to process one ticket, one claim, one invoice before the agent existed? Essay five gives you the two dates nobody has put together, which is that South Africa's sector AI instruments and your own roadmap are aiming at the same window.
If you are technical rather than executive, read three and four instead. The arithmetic in three changes how you scope agents, and the containment rule in four is the only agent security control I know of that survives a model upgrade.
The through line
If I had to reduce nine thousand words to one instruction, it would be this.
Measure the unit before you automate it. What does one ticket, one claim, one invoice cost you today? That single number takes about a week to establish, costs nothing, and is the difference between a programme that can defend itself in 2028 and one that gets quietly cancelled because nobody could prove it did anything.
Everything else in the series is downstream of that.
What I would like back
I have tried to make each essay falsifiable rather than merely assertive, which means some of it should be arguable. Essay eight attacks its own thesis for four paragraphs on purpose.
If you think I have a rung wrong, if your organisation has actually stopped an AI project at a gate, or if you can name a South African institution that captured a baseline before its first agent, I would genuinely like to hear it. Those are the three things I could not find enough examples of while writing.
Edzai Conilias Zvobwo writes on causal inference, AI and mathematics. The Industrialising AI series is free to read and free to share.
