The debate over whether AI can be creative is less philosophical than it appears. It is an economic question about where value accrues. Current systems excel at interpolation across vast training corpora, producing outputs that recombine known patterns. What they lack is intuition, the compressed, tacit judgment that lets an engineer smell a design flaw before running the numbers, or an architect reject a technically valid solution because it will not age well. This distinction matters because it draws the line between tasks that commoditize and tasks that retain scarcity value.
Globally, the implication is a bifurcation of technical labor. Routine synthesis, boilerplate generation, first-draft design exploration, and pattern-matching diagnostics are collapsing in price. Meanwhile, the judgment layer, deciding which problem to solve, when a model's confident output is quietly wrong, and how to weigh tradeoffs the training data never captured, is becoming the durable moat. Firms that treat AI as a replacement for judgment rather than an amplifier of it will ship faster and fail harder, accumulating technical debt at machine speed. The winners will be those who redesign workflows so human intuition is applied at the highest-leverage decision points rather than spread thin across low-value tasks.
For Japanese enterprises and SIers, this reframing is both a warning and an opening. The traditional SIer model has monetized labor volume: large teams executing well-specified requirements. That model is directly exposed, because specified, repeatable work is exactly what AI compresses. But Japanese engineering culture has long prized the tacit, hard-to-document knowledge that accumulates in domain-heavy systems like manufacturing lines, financial back-office logic, and public infrastructure. That intuition is not in any training set. It is the asset.
The strategic move for SIers is to stop selling headcount and start selling judgment. That means restructuring engagements around outcomes and architectural decisions rather than person-months, and pairing AI-accelerated delivery with senior engineers whose value is precisely their intuition about failure modes, regulatory nuance, and long-term maintainability. RPA-centric practices face the sharpest pressure here, since rule-based automation is the easiest layer for AI agents to absorb and extend.
For domestic development teams, the practical discipline is calibration: knowing when to trust AI output and when to distrust it. Building that muscle, through code review norms, verification gates, and deliberate cultivation of senior judgment, will separate teams that use AI as leverage from those that use it as a crutch. The scarce resource in the coming cycle is not compute or model access. It is the human intuition that decides what the machines should be pointed at.
