Levels of Autonomy
A six-level framework for how much an AI shopping agent decides and does alone, from assistive recommendations to delegated purchasing within mandates.
In short: “Agentic commerce” covers everything from an AI that suggests products to one that buys on a schedule without asking. This page defines six levels, from L0 (no agent) to L5 (fully delegated purchasing within mandates), describes the human-in-the-loop patterns that separate them, and gives real examples at each level as of October 2026.
Why levels matter
Autonomy determines almost everything downstream: what payment credential is needed, what evidence of consent exists if a purchase is disputed, how much merchants must trust the agent, and what users expect. A single product can operate at several levels; Google’s AI Mode, for example, offers both assisted checkout and price-triggered purchases. The levels describe a given transaction, not a company.
This framework is descriptive and specific to this wiki. It borrows the idea of graded autonomy from driving automation but is not an industry standard.
The six levels
| Level | Name | Agent does | Human does | Typical credential |
|---|---|---|---|---|
| L0 | Manual | Nothing | Everything | Card or wallet at checkout |
| L1 | Assistive recommendation | Answers questions, suggests products | Searches, chooses, and buys on merchant site | None (agent never touches payment) |
| L2 | Guided selection with handoff | Researches, compares, builds a shortlist or cart, links out | Chooses and completes checkout on merchant site | None, or a pre-filled cart link |
| L3 | Assisted checkout | Runs checkout end to end | Reviews the final cart and confirms each purchase | Delegated or network token, per-purchase approval |
| L4 | Conditional delegation | Monitors and buys when user-set conditions are met | Defines conditions in advance; may be notified | Token plus a pre-authorized mandate for a specific item or rule |
| L5 | Fully delegated purchasing within mandates | Decides what, where, and when to buy within a standing budget or policy | Sets the mandate and reviews afterward | Agent-bound token with spending controls, standing mandate |
L1: Assistive recommendation
The agent is an informed adviser. It explains, compares, and suggests, but every action happens on the merchant’s own site.
Examples: general AI assistants answering “which robot vacuum handles pet hair”; Amazon’s Rufus at its February 2024 launch; most brand chatbots.
L2: Guided selection with handoff
The agent narrows the field and sends the user to buy, often with a link straight to the product or a pre-filled cart. The merchant’s checkout, account, and payment method are used.
Examples: ChatGPT shopping research (November 2025), which produces a buyer’s guide with links; ChatGPT after OpenAI ended Instant Checkout in March 2026 and routed purchases to retailer sites and apps; Perplexity’s shopping answers that link to merchants.
L3: Assisted checkout
The agent completes the purchase inside its own interface, but the user explicitly approves the final cart and price. This is the dominant pattern for in-assistant checkout as of October 2026.
Examples: OpenAI’s Instant Checkout (September 2025 to March 2026); Microsoft Copilot Checkout (January 2026); Google agentic checkout in AI Mode and Gemini; Meta’s Muse, which returns to the user for approval before completing a purchase.
In AP2 terms this is a “human present” transaction: the user approves the exact items and price, producing closed Checkout and Payment Mandates (called a cart mandate before AP2 v0.2).
L4: Conditional delegation
The user approves a specific rule in advance (“buy this jacket in size M if it drops below $150”), and the agent executes later without asking again.
Examples: Amazon Rufus auto-buy (November 2025), which places an order within a short window after a target price is met using the customer’s defaults; Google’s price-tracking agentic checkout, which can be set to buy when a price drops below a threshold; scheduled reorders.
In AP2 terms this is “human not present”: the agent acts on signed open mandates (an intent mandate before AP2 v0.2) without a fresh approval from the user. Liability allocation for this case is still being worked out by networks and issuers.
L5: Fully delegated purchasing within mandates
The agent decides what to buy, from whom, and when, inside a standing policy: a monthly budget, approved categories or merchants, quality rules. The user reviews after the fact.
Examples: largely emerging in consumer retail as of October 2026. Closest are enterprise and machine-to-machine cases: agents paying for APIs and compute over x402 or Stripe’s Machine Payments Protocol, and procurement agents operating within corporate spending policies. The June 2026 Visa and OpenAI collaboration described user-set spending caps, merchant restrictions, and approval requirements, which are the controls an L5 consumer product would need.
Human-in-the-loop patterns
The line between levels is drawn by where and how a human intervenes. Common patterns:
- Confirm-to-buy. A final review screen showing items, total, shipping, and merchant (L3).
- Threshold escalation. The agent acts alone below a value and asks above it (L4 and L5).
- Allow-lists and block-lists. Approved merchants, brands, or categories.
- Time-boxed mandates. Authorization expires after a set period or a single use.
- Step-up authentication. The issuer or wallet requires a passkey or biometric at approval time, binding the confirmation to the real user.
- Notify-after. The user is told after each purchase and can cancel within a window.
- Pause and resume. Long-running agents (such as Muse) work in the background and stop when they need a decision.
How autonomy maps to infrastructure
| Requirement | L1–L2 | L3 | L4–L5 |
|---|---|---|---|
| Structured product data | Helpful | Required | Required |
| Merchant checkout integration | Not needed | Protocol or browser automation | Protocol strongly preferred |
| Payment credential held by agent | No | Scoped token, per purchase | Agent-bound token with standing limits |
| Proof of user intent | Not needed | Cart approval | Signed mandate or equivalent |
| Agent identity verification | Optional | Expected | Expected |
| Dispute evidence | Standard | Approval record | Mandate plus agent logs |
See payments and trust and agent identity and fraud for the mechanisms.
Practical guidance
- Users should match autonomy to stakes: L4 for repeat purchases of known items, L3 for anything new or expensive.
- Merchants can serve L1 and L2 agents with good product data alone; L3 and above require checkout integration. See for brands.
- Agent builders should make the current level visible to users and log the authorization behind every purchase. See for developers.
Sources
- Amazon announces Rufus — Amazon
- Amazon Rufus can now buy items for you when they go on sale — Shopifreaks
- Google Shopping launches agentic checkout and more AI shopping tools — Google
- Microsoft debuts Copilot Checkout — GeekWire
- Meta launches Muse AI agent with Stripe Link checkout — The Paypers
- Google launches payments protocol for AI commerce — Digital Commerce 360
- Visa Partners with OpenAI to Power the Next Generation of AI Commerce — Visa
- Everything we announced at Sessions 2026 — Stripe
- OpenAI Ends Instant Checkout in ChatGPT — The Keyword
- OpenAI debuts ChatGPT Shopping Research — SiliconANGLE