Agentic vs. "AI personalization"
"AI personalization" has meant something for a decade: a machine-learning model picking among marketer-built experiences, or lately an LLM generating copy. The AI is a feature inside the vendor's product, driven through the vendor's UI, by hand.
Agentic personalization inverts who holds the controls. The operations themselves — cohorting, drafting, placing, measuring — are exposed as tools, so a capable agent can run the whole loop: notice a segment is underperforming, draft three variants, stage them behind a review, and report the holdout lift a week later. The intelligence you rent from a model lab does the operating; the personalization layer supplies the hands.
The two jobs: operating and deciding
The useful mental model splits personalization into two jobs. Operating: authoring segments, variants, and placements, and tuning them over time. This is slow, creative, reviewable work — exactly where agents (and humans) belong.
Deciding: choosing, in the few hundred milliseconds after a visitor lands, which experience this specific person gets. This job runs thousands of times an hour with money attached. It should be deterministic — same behavior in, same decision out — with no model call, no token bill, and nothing to hallucinate. Vendors that blur these two jobs are usually putting a probabilistic model somewhere it costs you cost, confidence, and a standing eval burden.
How to evaluate an agentic personalization claim
Three questions cut through the label. Where is the decision made, and when — runtime scoring or pre-written rules? What does the agent do with nobody watching — autonomous action, or a generate-button a person approves at every step? And how is lift measured — against a randomized holdout, or a dashboard grading its own homework?
A fourth, increasingly: can your agent drive it? If the product's agentic story is a chat box inside its own UI, the agents are decoration. If it ships an MCP server your Claude can call, the story survives contact.