A viral essay by Belgian engineer Frederick Van Brabant has racked up 571 points on Hacker News in nine hours, and the argument lands hard in Australian boardrooms still weighing AI rollouts: AI rarely makes a slow process faster. The bottleneck, Van Brabant writes, sits upstream, in unclear requirements, missing context and the quiet work of getting the brief right.
Faster execution is the easy part. Knowing what to build, and for whom, is where the time actually goes. For Australian mid-market operators piloting AI, that flips the usual ROI story on its head.
What the essay actually argues
Van Brabant’s piece is short and pointed. He uses software development as the worked example: a request like ‘send an email to the user once a sale is completed’ sounds trivial, but in practice it spawns weeks of clarification before a single line of code gets written. Which user? Which sale? What if the user has unsubscribed? What about partial refunds? The coding is the small part; the thinking is the big one.
AI tools, he argues, only feel faster because they hide that upstream work. The model gets fed a perfectly specified prompt, then writes the code in seconds. Strip the specification step out of the comparison, and the productivity gain shrinks fast.
Why the comparison with developers is unfair
The essay’s sharpest point is about the AI-versus-developer benchmark. Companies measuring ‘AI 10x faster than a junior dev’ usually feed the AI a tight, well-formed brief and feed the developer a messy ticket. Give the developer the same crisp specification, and the productivity gap collapses.
Van Brabant references Frederick Brooks’ The Mythical Man-Month (1975) and Eli Goldratt’s The Goal to make the case that ‘predictable, high-quality inputs’ are what speed up any process, with or without AI. The pattern repeats outside software. A customer service bot deployed without a tidy knowledge base produces the same mess a human agent would, just faster and at scale.

Why this lands in Australian boardrooms
Australian mid-market and enterprise leaders have spent the past 18 months hearing the same pitch: hire an AI vendor, watch productivity jump 30 to 50 per cent. The pitch has produced plenty of pilots, fewer scaled deployments and a fair bit of disappointment.
Van Brabant’s essay names the missing piece. Most pilots stall not because the model is wrong but because the upstream work, the clean data, the documented process, the agreed definition of done, was never finished. Our AI strategy practice routinely starts there: less prompt engineering, more requirement engineering. The hard part is not asking Claude or ChatGPT a better question. It is knowing the answer you needed in the first place.
The upstream work most pilots skip
In practice, three jobs do most of the heavy lifting in an AI rollout, and none of them involve the model. The first is mapping the process as it actually runs, including the exceptions and the workaround spreadsheets nobody documents. The second is cleaning the inputs the AI will read: customer records, contracts, support tickets, the bits a model needs in order to answer the way a senior staff member would. The third is deciding what ‘good’ looks like, in plain English, before the first prompt is written.
Skip any of those three and the AI inherits the mess. We’ve made the same point about chatbots in particular: most fail because they ship with the same superficial thinking that gave us the homepage carousel, a feature added because it was technically possible rather than genuinely useful.
Where AI does actually speed things up
None of this means the productivity claims are pure spin. Van Brabant concedes AI tools genuinely help on well-scoped tasks, particularly code with a tight test suite and clear inputs. In customer service, AI handles peak overflow without hiring. In finance, document parsing chews through compliance work that used to take days. The pattern is consistent: AI is fast where the brief is clean and the failure modes are bounded. It is slow, or actively damaging, where the process is fuzzy and the data is dirty.
For Australian operators considering running models locally to keep data on-shore, our OpenClaw guide covers the practical setup. The teams getting real productivity gains from AI in 2026 are the ones who spent the prior year cleaning data, documenting processes and writing down what good looks like. The honest answer is that there is no shortcut. The unglamorous upstream work is still the work.
