The half of the problem most tools skip
Most "AI personalization" tooling competes on the supply-chain half: generating variants faster, keeping brand voice consistent, routing assets to channels. That work is upstream and batch — it happens on your schedule, in bulk, before any visitor arrives. The decision is the other half: runtime and per-visitor, in the few hundred milliseconds after someone lands, with one shot to get it right. A warehouse full of beautifully produced variants converts nobody if nothing is choosing between them intelligently in the moment.
Properties of a real decision layer
Four tests. It decides at runtime, per visitor — not by executing pre-written audience rules on a schedule. It works on anonymous traffic, since most visitors are anonymous and a decision layer that needs a resolved identity forfeits the majority of its decisions. It is deterministic — same behavior in, same experience out, no model improvising in a path that runs thousands of times an hour. And it measures itself against an always-on holdout, so the lift it produces is proven rather than asserted.
The evaluation question that separates decision layers from supply-chain tools: where does the decision happen, and when? "We generate the variants and you wire them into your existing system" means the deciding is still your problem.