Amex's Intent Intelligence: Intent as a Token — Judgment Still Missing

American Express published its Intent Intelligence Framework on October 13, 2026 — and it's the first card-network design that treats a customer's intent as a machine-checkable object instead of a checkout click.

The short version: capture and authenticate the customer's purchase intent before the agent buys, carry it as an Intent ID + Proof of Intent Token, and scope the payment credential to exactly that intent. A purchase that doesn't match what was approved can be stopped at authorization — not disputed after settlement.

How it works (from Amex's release)

1. Create Intent API records the instructions and constraints — merchant, maximum budget, preferences. The customer authenticates and signs off.
2. The system issues an Intent ID and a Proof of Intent Token carrying those constraints.
3. The token obtains a secure digital payment credential scoped to the intent. The agent never touches the actual card number.
4. The ID and token act as controls downstream — validation, fraud prevention, authorization decisioning, dispute resolution.

It's a core service of the ACE Developer Kit (launched April 2026). Amex says it has not seen an agentic-dispute increase yet — this is proactive. Its Trendex survey: 66% of surveyed U.S. consumers likely to use agents to research prices, 61% for recommendations. Amex also name-checked engaging EMVCo and supporting the x402 Foundation's standards work.

What the token can't do

It checks a purchase against approved constraints. It does not score whether the agent's choice inside a broad intent was wise — and it faithfully enforces an intent that was captured wrong and rubber-stamped. Amex's own autonomy example (agents reordering household staples inside price/size constraints) is exactly where that gap opens.

We tested the gap live

This morning (Oct 14, ~09:24 EDT) we ran two Intent-Intelligence-shaped instructions through our public decision gate at scriptmasterlabs.com/api/harness/decide (bands: ≥0.80 auto-act, 0.50–0.79 advisory, <0.50 escalate; decider local-heuristic-v1, calibrated=false):

• Mismatch: authenticated intent for white Nike sneakers, size 9, max $120, Nike.com — agent tries black Adidas at $145 from an unlisted seller → 0.59, advisory/hold
• Exact match: usual paper towels, $24, usual merchant, inside the authenticated intent → 0.59, advisory/hold

Identical scores. The honest lesson cuts both ways: a gate fed intent as prose cannot discriminate. Amex carries intent as structured fields bound to a token — and structured constraints are mechanically checkable in a way prose never will be.

The layers compose: the token answers "does this match what was approved?", a per-payment confidence score answers "should this fire at all?", and a sealed decision receipt is the dispute file. Our Verdict Protocol was re-verified today: LIVE_SETTLEMENT_VERIFIED, /decide $0.25 USDC, /seal $0.01 USDC, Base, x402 v2.

Full canonical with the dated receipts table, the enforce/cannot-enforce table, Claim Receipts and a 5-step DIY: https://scriptmasterlabs.com/amex-intent-intelligence-framework

ScriptMasterLabs is a Service-Disabled Veteran-Owned Small Business. Coverage-based, not hands-on: Amex published a framework and developer-kit service; we could not verify a public testable endpoint for it today.