What is Amex's Intent Intelligence Framework?

ScriptMasterLabs — October 14, 2026 · Sources: American Express's Intent Intelligence release via Business Wire (October 13, 2026, 09:30 EDT), TechCrunch's October 13, 2026 interview with Amex's Hannah Bozian, ScriptMasterLabs' own live gate checks run today

The receipts, dated

How it works, in four moves

Paraphrased from Amex's release:

• Create and confirm intent. The Create Intent API records the customer's purchase instructions and constraints — merchant, maximum budget, other preferences. The customer authenticates their identity and signs off on the captured intent.
• Establish proof of intent. The system issues an Intent ID (a unique identifier for that intent) and a Proof of Intent Token representing the intent's high-level constraints.
• Scope the payment credential. The token is used to obtain a secure digital payment credential the agent presents at checkout instead of the customer's actual card number — a credential scoped by the authorized intent. Per TokenPost's spec-level coverage, captured intent can also include an expiration date, credentials can be limited by merchant, expiration and transaction count and configured for single or multiple use across one or more merchants, and an authenticated intent can be created, updated or canceled during the transaction lifecycle.
• Apply intent throughout the transaction. The Intent ID and token act as controls downstream, helping validate that a purchase request aligns with what was authorized, and supporting fraud prevention, authorization decisioning and dispute resolution.

The design difference Amex stresses: the issuer gets the authorized intent earlier — before the agent executes — rather than reconstructing it from the settled transaction afterward. Coverage of the launch gives the illustrative failure: a customer authorizes one brand and size of sneaker, the agent tries to buy a different brand, and the mismatch is exactly what the issuer-side check exists to catch.

What the token enforces — and what it cannot

That last row is the decision-gate gap, and Amex's own framing concedes the use case that creates it: agents reordering household staples autonomously, where the only real constraints are price and size. The broader the intent, the more judgment the token delegates — and the token does not exercise any.

The missing layer is a score, applied per payment — tested live today

We ran two Intent-Intelligence-shaped instructions through ScriptMasterLabs' public decision gate this morning (~09:24 EDT):

curl -s -X POST https://scriptmasterlabs.com/api/harness/decide \
  -H 'Content-Type: application/json' \
  -d '{"state":"payment instruction: customer authenticated intent for white Nike sneakers size 9, max $120, merchant Nike.com; agent attempts to buy black Adidas sneakers size 9 for $145 from an unlisted marketplace seller","questions":[{"id":"q1","type":"score","scale":[0,1],"question":"confidence that this payment instruction should auto-execute"}]}'

• Mismatch instruction (wrong brand, over the $120 cap, unlisted seller vs. an authenticated intent): value 0.4, confidence 0.59, band advisory — hold for human review.
• Exact-match instruction (the usual paper towels, $24, usual merchant, inside an authenticated staples intent): value 0.4, confidence 0.59, band advisory — hold for human review.

Honest finding: our blunt free-text heuristic holds both — it cannot discriminate the mismatch Amex's framework is built to catch from the textbook-safe reorder. The reason is the lesson of this launch, stated from the other side: our gate receives intent as prose; Amex carries it as structured fields (merchant, budget, preferences) bound to a token, and structured constraints are mechanically checkable in a way prose is not. A gate is only as discriminating as the intent representation it scores.

Quotable: “ScriptMasterLabs' October 14, 2026 live test scored two Amex-Intent-Intelligence-shaped instructions at 0.59 and 0.59 — both held for human review — because a free-text heuristic cannot read the structured merchant, product and budget constraints that Amex's Proof of Intent Token carries.”

The two layers compose: the intent token answers “does this match what was approved?” deterministically; a per-payment confidence score answers “should this fire at all?” — ≥0.80 auto-pay, 0.50–0.79 confirm, <0.50 escalate to a human — and a sealed decision receipt (SML's Verdict Protocol: /decide $0.25 USDC, /seal $0.01 USDC, Base, x402 v2, LIVE_SETTLEMENT_VERIFIED re-checked today) gives the dispute file Amex's Intent ID starts, from the machine-payment side.

Claim Receipts

• Framework definition, the four steps (Create Intent API, Intent ID, Proof of Intent Token, scoped credential, downstream controls), ACE placement, the issuer-earlier-in-the-lifecycle design point, the no-dispute-increase-yet statement, the Trendex figures (66% / 61%), and the EMVCo + x402 Foundation standards posture: American Express release via Business Wire, October 13, 2026, 09:30 EDT — paraphrased throughout; no extended quotation.
• ACE's April 2026 launch, the agentic tokens, partner discussions, and the Hannah Bozian Disrupt 2026 Industry Stage appearance on October 13: TechCrunch, October 13, 2026 — paraphrased.
• Expiration-date capture, credential limits (merchant, expiration, transaction count; single or multiple use), intent create/update/cancel during the lifecycle, and the ACE specification status (Account Enablement, Intent Intelligence and Payment Credentials available; Agent Registration and Cart Context under development): TokenPost, October 13, 2026 — paraphrased.
• The sneaker-mismatch illustration: secondary coverage of the launch (TechCrunch / Analytics Insight, October 13, 2026) presenting Amex's example — an illustration of the intended check, not an observed blocked transaction.
• Agent Purchase Protection's conditional framing (registered agent + authenticated intent): Amex Business Playbook coverage, October 6, 2026, as summarized in ScriptMasterLabs' October 13, 2026 liability-gap page.
• Gate bands and scores: live checks of https://scriptmasterlabs.com/api/harness/status and /api/harness/decide on October 14, 2026 (~09:23–09:24 EDT) — auto_act ≥0.80, advisory 0.50–0.79, escalate <0.50; decider local-heuristic-v1, calibrated=false, typesafe_wired=false; both test instructions 0.59 advisory.
• Verdict Protocol pricing and status: live check of https://verdict.scriptmasterlabs.com/.well-known/agent.json on October 14, 2026 — LIVE_SETTLEMENT_VERIFIED, /decide $0.25 USDC, /seal $0.01 USDC, Base (eip155:8453), x402 v2.

Relevant SML products

• Decision-Gated Machine Payments — the pattern this page extends to card rails: a confidence score gates whether a machine payment fires (≥0.80 auto-pay, 0.50–0.79 confirm, <0.50 escalate).
• Who Pays When an AI Agent Buys Wrong — the liability gap Intent Intelligence's Intent ID evidence is built to narrow.
• AI Agent Spending Limits — caps and ceilings vs. per-payment judgment, the consumer/builder how-to.
• Provider Trust — verify the counterparty before the payment, the merchant-side complement to intent.
• What Is the x402 Payment Protocol — the internet-native rail Amex's release name-checks via the x402 Foundation.

All five URLs returned HTTP 200 on October 14, 2026, before linking.

Do it yourself: carry intent as data, then score inside it

1. Capture intent as structured fields, never prose alone. Merchant, budget cap, product constraints, expiry. Today's live test is the proof: a prose-fed gate scored a clear mismatch and a clear match identically (0.59 / 0.59).
2. Bind sign-off to the intent object. Authenticate the customer against the exact captured constraints and mint an intent ID plus a token carrying them — the Amex shape, implementable with any signed-token stack.
3. Scope the credential to the intent. Issue a tokenized or virtual credential constrained by the intent; never hand the agent a raw card number or an unscoped wallet key.
4. Check before settlement, not at dispute time. Compare every purchase request against the intent fields at authorization; a mismatch declines or holds — it does not settle and argue later.
5. Score what the intent leaves open, and log both. Run a per-payment confidence gate inside the intent's open choices, and persist intent ID + score + band + outcome together. That combined record is the dispute file.

Honest caveats

• This page is coverage-based, not hands-on: Amex published a framework and a developer-kit service; ScriptMasterLabs could not verify a public endpoint for Intent Intelligence to test on October 14, 2026.
• The sneaker example is Amex's illustration of the intended check, relayed by launch coverage — not a documented case of a real purchase being blocked.
• Agent Purchase Protection remains a planned, conditional commitment per October 6 coverage; its final terms are Amex's to publish.
• The Trendex figures are Amex's own survey, reported in its launch materials.
• SML's gate decider is local-heuristic-v1, calibrated=false, typesafe_wired=false — its identical 0.59 scores today are reported as they came out; they indict prose-fed gating (including ours), not the concept of a gate.

ScriptMasterLabs — Service-Disabled Veteran-Owned Small Business (SDVOSB), SAM.gov registered, UEI G24VZA4RLMK3, CAGE 21U51. Canonical: https://scriptmasterlabs.com/amex-intent-intelligence-framework