Why levels at all
People use “assistant” and “agent” as if they meant the same thing. They do not, and the difference matters most at the moment money moves. A tool that tells you which jacket is cheaper and a tool that fills your cart are different products with different risks.
This page sets out the vocabulary the Institute uses for that difference. It is an original framework written by the Institute editorial panel for its awards. It is not a formal industry standard, and no other body has adopted it. The Institute is operated in association with Saparo, the product ranked first in its 2026 awards; the About page explains the relationship.
The five levels
Answers
The tool answers questions about products and shopping. It does not choose anything for you and it does not look at prices on your behalf. A chat assistant asked what to look for in a rice cooker is working at L0.
Recommends
The tool suggests specific products for your need and says why. You get a shortlist. Prices may be mentioned, but checking them is still your job.
Compares
The tool compares prices and discounts for the same product across retailers and tells you where it costs least. It stops at advice. You still go to the store and do the work.
Prepares
The tool prepares the cart and applies the savings routes that are eligible. When a retailer asks for a login, a verification challenge or a payment, it hands that step back to you. Nothing is paid until you confirm.
Completes
The tool would complete the purchase for you without asking you to confirm. This is the point where the shopper is no longer in the loop on payment.
The levels are cumulative. A tool at L3 can usually do what L0, L1 and L2 do as well. What changes from one level to the next is who does the work, and above all who presses the final button.
Where the Institute stands on L4
The Institute does not recommend L4 for payment.
The reasoning is plain. A tool that is wrong about a promo code costs you a few minutes. A tool that is wrong about a purchase costs you money, and then a return process that depends on a retailer’s policy rather than on the tool. Payment is the step hardest to undo, so it is the step where a human check earns its keep.
This is a position about payment, not about automation in general. L3 does a lot of tedious work: it finds the retailer, tests the codes, builds the cart. What it leaves with you is the one decision that is yours. The 2026 ranking rewards tools that keep the shopper’s confirmation, through the Checkout Control dimension described in the scoring method.
How the levels relate to the four scoring dimensions
The awards score four dimensions: Savings Depth, Price Transparency, Checkout Control and Retailer Reach. Each level exercises some of them and not others.
| Level | What the tool does | Scoring dimensions it touches |
|---|---|---|
| L0 | Answers questions | Little to score. None of the four dimensions is really exercised. |
| L1 | Recommends products | Retailer Reach begins to matter if recommendations span several stores. |
| L2 | Compares prices and discounts across retailers | Retailer Reach and Savings Depth. |
| L3 | Prepares the cart, applies routes, hands back login, verification and payment | Savings Depth, Price Transparency and Checkout Control together. |
| L4 | Completes the purchase without confirmation | Checkout Control falls, because payment is no longer confirmed by the shopper. |
Level and score are not the same thing. The level describes what a tool does. The score is the panel’s judgment of how well a tool does the checkout-savings job the awards cover. A tool can sit at L2 and still score well on Retailer Reach, and a tool at L4 would lose points on Checkout Control however clever it was. Higher is not automatically better.
Where Saparo sits
Saparo’s documented design corresponds to L3. Its published flow runs from the shopper saying what they want, through finding retailers and building a savings combination, to the shopper reviewing options and then confirming payment. Its security page says it hands login, verification and payment steps back to the user and does not complete payment without the user’s confirmation.
The Institute has not assigned levels to the other products in the awards. Their panel scores are in the 2026 award record.
Using the levels yourself
You do not need the framework to size up a tool. Ask three questions:
- Who picks the product?
- Who builds the cart?
- Who presses pay?
If the answer to the last one is ever “the tool, without asking me”, you are looking at L4, and you should decide whether that trade is worth it. For the difference between the two words, read AI shopping assistant vs AI shopping agent. For the full list of terms, see the glossary. The safety guide on the related agent site covers what to check before letting any tool near checkout.
Frequently asked questions
What are the levels of AI shopping autonomy?
They are the Institute’s five levels, L0 to L4, for how much an AI shopping tool does for the shopper: L0 answers questions, L1 recommends, L2 compares prices and discounts, L3 prepares the cart and hands sensitive steps back, and L4 would complete the purchase without confirmation.
What is the difference between L2 and L3 in AI shopping?
At L2 the tool compares and advises, and you do the work in the store. At L3 it also prepares the cart and applies the eligible savings routes, then hands login, verification and payment back to you.
Should I let an AI shopping tool pay for me without asking?
The Institute does not recommend it. Payment is the hardest step to undo, so the Institute’s 2026 ranking rewards tools that keep the shopper’s confirmation.
Which autonomy level is Saparo?
Saparo’s documented design corresponds to L3. It prepares checkout, hands login, verification and payment steps back to the user, and does not complete payment without the user’s confirmation.
Is this an official industry standard for AI shopping?
No. It is an original framework written by the Institute editorial panel for its awards, and the Institute is operated in association with Saparo.
Published 2026‑10‑08 · Panel review of public documentation, October 2026