Agent Autonomy Levels

Levels of agent autonomy · AI agent autonomy scale · Agent autonomy scale

Definition

Agent autonomy levels are a 0-to-4 scale describing how much of a purchase an AI agent carries out itself: from answering questions to making recurring purchases on its own.

Why a scale is needed

“An AI agent bought it” can mean very different things: the agent suggested a product, built the basket, or charged the card without asking. Risks, data requirements and who is liable for a mistake all depend on which one it was. An autonomy scale separates these scenarios.

The idea is familiar from cars: the SAE J3016 standard describes six levels of driving automation, from 0 to 5. The scale below follows the same principle — the higher the level, the fewer decisions remain with the human — but it is an analogy, not a transfer of the standard.

Five levels of autonomy

Level	What the agent does	What the human does	Control mechanism
0 — answers	Answers questions about products and terms	Chooses and buys alone	Reliable data
1 — finds and compares	Builds a shortlist, compares prices and specifications	Chooses and places the order	Reliable data
2 — builds the basket	Prepares the entire order	Confirms the order and payment	Human in the loop
3 — buys within a mandate	Places and pays for the order itself	Sets limits on amount, categories, stores in advance	Payment mandate, Verifiable Intent
4 — buys regularly on its own	Decides when and what to reorder	Receives a report, changes the rules	Mandate, monitoring, access revocation

Trust mechanisms by level

At level 2 the key mechanism is the human in the loop: the agent does everything except the last irreversible step. At level 3 confirmation shifts in time — the person records permission and constraints once in a payment mandate, and the merchant and payment provider need proof that the purchase stayed within them; that is what Verifiable Intent is for.

Important: the level is defined not by how smart the model is but by how many decisions it has been trusted with. A printer that reorders its own ink formally operates at level 4, yet in Gartner’s classification it is a “bound” machine customer — it executes a fixed rule rather than choosing.

What it means for a store

Decide which levels you are prepared to accept per category: for example, consumables up to level 3, electronics and non-returnable goods at level 2 only.
For level 2, build a checkout that hands confirmation to a person without losing the basket the agent assembled.
For levels 3–4, accept and store proof of authority — you will need it for disputes and returns.
In your order analytics, separate orders via an agentic browser, where the agent walks the ordinary site, from orders via a protocol, where it calls an API: the control options differ.
Revisit the rules as standards for agent identification and payment mandates mature.

Frequently asked questions

Is this an official standard?

No. There is no generally accepted scale for shopping agents: different companies and researchers propose their own. The scale in this entry is a working one — it exists so that a team and its partners mean the same scenario when they talk. The analogy with the SAE levels of driving automation (0 to 5) helps convey the idea, but the levels do not map onto each other directly.

At what level do agents operate as of September 2026?

Publicly announced scenarios mostly sit at levels 1–2. The “Find cheaper” agent in Alice AI from Yandex, the dominant search engine in Russia, compares prices across marketplaces and stores for the user's region — that is level 1. One-click purchase via Yandex Pay and Instant Checkout in ChatGPT are designed so that a person confirms the order — level 2. Level 3 is described in protocols, for example through the Intent Mandate in AP2, but there is little public data on how widespread it is in retail.

When do you need a human in the loop, and when is a mandate enough?

The more expensive and less reversible a mistake, the lower the acceptable level. A first purchase from a new seller, expensive electronics, non-returnable goods — level 2 with human confirmation. Recurring low-value purchases from a trusted store can move to level 3: the person sets the limits in a mandate once, and the agent acts within them.

What does the scale mean for a store?

It tells you which guarantees to ask of an incoming agent. At levels 0–1 the agent needs data: catalogue, prices, availability. At level 2 it needs a reliable path to checkout that hands confirmation to a person. At levels 3–4 it needs proof of authority: a mandate, verifiable intent, identification of the agent itself.