How enterprises approach AI, glimpsed through the AI training market

How enterprises approach AI (LLMs), glimpsed through the AI training market — three shifts in four years:

2022–2023 — The wake-up call: "Build our own model."

ChatGPT (late 2022) made AI suddenly practical, and enterprises woke up to the possibility. The first reflex was to own it: the courses on offer were math, statistics, RNNs and CNNs, Python, ML frameworks — the goal was to train an in-house model, because security meant keeping it inside.

2024–2025 — "Just use the SOTA models."

The gap to frontier models had grown too wide to close. What companies bought shifted to building on top — APIs, multi-agent systems, and classic software engineering: loops, conditionals, workflows, orchestration.

Now — AX (AI Transformation).
AX has become the headline 2026 strategy you now see across Korea's largest conglomerates. Frontier model makers are moving into the application layer themselves, so building your own app gets harder to keep competitive when the model creator ships a better-tuned one. The emerging stance: become a fast, fluent consumer of AI-native tools, applied to the work you already do.

Here's the part worth noting. What mattered to learn didn't just change — it moved from what a course can teach to what only your own work can.

Phase 1 was math and statistics.
Phase 2 was programming — how to build it.
Phase 3 is domain knowledge — something you can't download: a clear answer to "what is the work we actually do, and where exactly does it hurt?"

The scarcest skill in AX isn't technical anymore. It's understanding your own job well enough to point AI at the right problem.

Curious whether others are seeing the same shift in how their organizations frame AI skilling.