The 18-month myth: AI cannot automate jobs until companies do
The countdown narrative is back. In a recent interview, Mustafa Suleyman argued that most computer-based professional tasks could be automated within 12 to 18 months. This is a bold claim and easy to believe if you only watch model demonstrations. Here is the controversial counter: The timeline is wrong, not because the models are weak, but because organisations are slow. Everyone is confusing capability with adoption. Models can improve at an exponential-looking rate, whereas institutions almost never do. Companies move through procurement, governance, integration, training, and change management, only to discover that the edge cases were never documented. The bottleneck is not intelligence. It is implementation.
Most white-collar work is also not ‘sitting at a computer’. It is the invisible glue between systems: Exceptions, approvals, hand-offs, judgement calls, liability, and audit trails. This glue is what makes work expensive, and what makes automation hard. You cannot replace an accounts payable clerk with an agent if invoices arrive in five formats, vendor masters are inconsistent, approvals happen via e-mail, and the record of truth lives in someone’s head. Artificial intelligence (AI) can draft, classify and suggest, but true automation requires machine-readable work: Clean data, defined controls, and a reliable path from input to decision to audit. That is why the phrase “retrofit AI into any organisation” should make executives smile. Most companies cannot quickly retrofit far simpler technology. Anyone who has lived through a simple workflow redesign knows the pattern: Data cleansing, policy fights, scope creep, and months lost to the messy 10% that drives 90% of the risk.
In regulated industries, the friction is multiplied. If a workflow touches client money, privacy, or fiduciary duty, “good enough” is not good enough. You need governance: What the model can access, how it is monitored, how errors are escalated, and who is accountable when it fails. Automation is not a software purchase; it is a new operating model with new liability. Productivity is even messier than the hype suggests: A rigorous field experiment by METR found that experienced software developers sometimes took longer with AI tools because they spent time verifying, correcting, and integrating outputs. That does not disprove AI; it proves that efficiency depends on workflow fit, not just raw capability.
So, what happens first, if not mass unemployment in 18 months?
Task compression: The hours spent on drafting, searching, summarising, first-pass analysis, routine reporting, and basic client communication shrink. Teams do not disappear overnight; they re-shape. The biggest hit is often at the entry level: Fewer junior analysts and fewer apprenticeship roles that used to be justified by grunt work. The labour market adjusts quietly long before it adjusts loudly through mass layoffs.
And here is the investor implication that people keep missing: If adoption is the bottleneck, the winners are not only the frontier model builders. The winners are the businesses selling the plumbing that turns old institutions into AI-ready institutions: Data engineering, integration layers, identity and access management, cybersecurity, model monitoring, audit trails, and process redesign. In other words, the boring enterprise stack that makes automation safe, accountable, and scalable.
Expect the AI boom to show up first as spending on integration and governance, not sudden headcount collapses. This matters in South Africa more than most, as many firms are still digitising basics and modernising legacy cores. In this environment, AI does not replace the back office overnight; it sits on top of it, constrained by it. The sequence is to digitise, standardise, implement, and then automate. Skip the first steps, and you do not get a jobless boom; you get a compliance incident.
The point is simple: AI may be racing ahead. Organisations are not. The next 18 months will be noisy, experimental, and occasionally brutal at the margins. But the “white-collar extinction event” is not a date on a calendar. It is an organisational decade disguised as a technological year.
