# What is agentic personalization?

> Agentic personalization is personalization run by an LLM agent. Given a goal, a set of tools, and guardrails, it plans and carries out the steps itself, rather than executing a flow a person designed. The label is used loosely, so treat it as a claim to test.

_Last updated 2026-11-01. From The Personalization Guide, https://personalization.guide_

Agentic personalization is personalization run by an LLM agent that is given a goal, a set of
tools, and guardrails, and then plans and carries out the steps to reach that goal, rather than
executing a sequence a person designed in advance. The agent decides what to do; a person decides
what the goal is and what the agent is not allowed to do.

## How it works

The agent starts from a goal stated in plain terms, such as reducing churn among trial users or
raising repeat purchases in a category. Around it sit three things:

- **Tools** it can call: an audience builder, a content generator, send APIs for the channels, and
  analytics to read what happened.
- **Guardrails**: contact policies so it does not over-message, brand and compliance rules, budget
  limits, and consent and purpose constraints on the data it may use.
- **A loop**: it proposes or takes an action, observes the outcome, and adjusts.

Within that frame the agent sets its own intermediate steps: which audiences to build, what content
to generate, on which channel and at what time to send, and how to change course from the results.
The defining feature is that the plan is the agent's, not a flow a person wrote out in advance.

## Why it is, in principle, a real shift

Earlier personalization executes a flow a person built: the rules, the journey, the model scoring
candidates inside a fixed pipeline. Agentic personalization moves the person from author of the
steps to author of the goal and the guardrails, and lets the system plan the steps between. If it
works as described, it changes who does the planning, which is a genuine change in kind rather than
degree.

## The label problem

In the market, "agentic" covers a wide range. At one end is a real planner that decomposes a goal,
takes actions with tools, observes outcomes, and replans. At the other is a conventional
machine-learning model, or even a rules engine, with a chat interface bolted on, where nothing
actually plans. Vendors use the word loosely, so any reported autonomy is a vendor claim until it is
shown.

## How to test the claim

The useful question is whether the system sets its own steps toward a goal, or merely scores options
inside a flow a person designed. In practice, ask:

- Does it decide the sequence of actions, or follow a fixed journey?
- Can it choose to do something its designer did not script?
- What are its tools, and can it genuinely call them?
- Where are the guardrails, and what stops it acting outside them?

Until these are clear, treat "agentic" as a claim rather than a capability.

## What to watch

Agentic personalization sharpens three older concerns. Auditability: when the agent both plans and
generates, there needs to be a record of what it showed and why. Measurement: a policy that keeps
changing itself is hard to measure, which is the argument for policy-level holdouts rather than
one-shot tests ([how personalization is measured](/measurement)). Manipulation: an agent optimizing
to persuade a specific individual raises the line between personalization and manipulation, which the
[critics](/ethics) page takes up. It also overlaps [generative
personalization](/library/techniques/generative-personalization), because the agent usually generates
the content it sends, and generative UI when it assembles the interface itself (see the sibling
[Generative UI Guide](https://generativeui.guide)).