Industrialising AI
Eight essays on why agentic AI programmes fail, and what the ones that survive do differently.
Gartner expects around 40 per cent of enterprise applications to carry task-specific AI agents by the end of 2026. It also expects around 40 per cent of agentic AI projects to be cancelled by the end of 2027.
Those are usually presented as good news and bad news. They are not. They are one curve. Adoption is running ahead of measurement, and the organisations that will still be running agents in 2028 are the ones that built the instruments before they built the agent.
These eight essays set out what those instruments are, why capable people keep skipping them, and what South Africa's regulatory position means for anyone deploying agents in financial services or the public sector.
- 01 We have put an electric motor on the steam shaft Why your AI spend has not moved the productivity numbers, and what the electrification of factories tells us about how long it will take.
- 02 Your pilot proves nothing Almost every AI business case is written on rung one of Pearl's ladder of causation. Here is why that fails at budget time, and what a defensible one looks like.
- 03 Ninety-five per cent is not ninety-five per cent An agent that succeeds at 95 per cent of steps succeeds at 36 per cent of twenty-step tasks. Every step passes, the dashboard stays green, and only the user sees the failure.
- 04 You cannot patch a prompt injection Anything your agent reads can instruct it. That is not a bug awaiting a fix, it is a property of the architecture, and the only real control is containment.
- 05 Who signs for the agent? South Africa chose a multi-regulator model, POPIA already restricts automated decisions, and sector instruments arrive in 2027/28. That is the same window your roadmap promises the agent economy.
- 06 A gate is a condition, not a date Most of what is called AI strategy is a procurement list, and most AI roadmaps commit organisations to dates they do not control. Three tests, and a replacement.
- 07 You cannot buy a lower touch rate. You train for it. Prompt training 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.
- 08 We cannot afford three years of unmeasured AI spend. That is our advantage. Africa arrives at agentic AI after the expensive mistakes have been made by someone else. Whether that becomes an advantage depends on one decision.