I confess: I’m obsessed with Connections. I play it daily and religiously post my results to a group chat with friends. (Selfishly, it's one of the few games I can regularly win in a group that frequently gets "Queen Bee" or solves crosswords in under five minutes.)

But oddly enough, this very human little puzzle game explains exactly why so many enterprises are struggling to scale their AI.

The New York Times recently broke down why Large Language Models (LLMs), despite all their linguistic prowess, often stumble on Connections. The game looks deceptively simple: sixteen words that you must sort into four clear categories. Easy… until it isn’t.

Example: Daisy, Food, Mountain, Supply. Answer: Chains.

Solving the harder categories requires contextual awareness, humor, subtle distinctions, and the ability to carry reasoning across multiple steps. This is known as System 2 thinking. It’s the slow, deliberate logic our brains begrudgingly switch to when quick pattern-matching (System 1 thinking) won’t cut it.

This is exactly where AI hits a wall in enterprise implementations.

Enterprise processes are rarely just "predict the next word." They are frustratingly full of exceptions, policies, nuance, and institutional quirks—like that one specific restart process that only "Kim in Kentucky" knows how to do.

To succeed in business, AI Agents need to understand why a step matters, not just what comes next. They must obey larger, multi-step workflows and navigate the constraints or policies surrounding them. This requires high-context, rule-bound reasoning. That is hard for humans, but it is exponentially harder for AIs with limited context.

None of this is mentioned in the shiny demos where you "just drag your data into this AI widget" and—poof—the entire workflow automates itself.

Enterprises are struggling with AI at scale because the reasoning they depend on requires real investment (and time!) to make it reliable. The same cognitive load that makes Connections an enjoyable challenge is present in almost every enterprise workflow.

The only difference? The stakes are much higher than losing a purple box.