Amex Intent Intelligence Framework: Enterprise Guide to Agentic Commerce Governance
Amex's Intent Intelligence Framework stops AI agents from making unauthorized purchases. What IT and finance leaders need to know about agentic commerce controls.
TLDR: American Express published its Intent Intelligence Framework on October 13, 2026 — a system that captures what you told your AI agent to buy, turns it into a Proof of Intent Token, and blocks checkout when the agent’s cart doesn’t match. If your employees carry corporate Amex cards, this is the first payment-network guardrail that speaks your language. But it’s a framework, not a product: nothing enforces it until your agents, your merchants, and your expense systems adopt it. Start by updating your corporate card policy for agent-initiated spend now, before the disputes start.
Why This Matters Now
Your employees’ AI assistants can already spend money. ChatGPT’s shopping integration went live with Visa rails in June. Amex is processing “thousands” of agentic transactions a week. And Forter — the fraud-prevention vendor half the enterprise e-commerce world runs on — reported the same day as the Amex announcement that agentic orders across its network quadrupled since August while automated fraud doubled in two months.
But here’s the number that should keep you up at night: Riskified’s Q1 2026 Agentic Commerce Pulse found 61.5% of consumers have used AI for product discovery, yet 55% aren’t comfortable letting agents actually buy — and 50.8% say the AI platform is responsible when an unauthorized purchase happens. That liability gap is where your finance team lives. When an employee’s agent orders the wrong $4,000 laptop configuration on a corporate card, who eats it? The employee? Procurement? The AI vendor? Amex just published its answer. You need your own.
How the Intent Intelligence Framework Works
The framework has four stages. Each one maps to an API surface in Amex’s ACE (Agentic Commerce Experiences) Developer Toolkit, which launched in April 2026.
| Stage | What happens | Enterprise analogy |
|---|---|---|
| 1. Create Intent | The Create Intent API captures purchase instructions: merchant name, max budget, product preferences | Purchase requisition in Coupa or Ariba |
| 2. Authenticate + confirm | Customer verifies identity and signs off on the captured intent; intent is editable and cancellable | Manager approval workflow |
| 3. Proof of Intent Token | System generates an Intent ID plus a token encoding the spend constraints; token is exchanged for a single-use digital payment credential | Virtual card number with embedded policy |
| 4. Enforce at checkout | Merchant / issuer checks the purchase against the token; mismatches (wrong brand, over budget, wrong merchant) get declined | Three-way match (PO / receipt / invoice) |
The critical design choice: intent is captured and authenticated before the agent transacts, and it travels with the payment as a first-class credential. Amex VP of Agentic Partnerships Hannah Bozian told TechCrunch the differentiator is pulling intent “earlier into the transaction flow” so risk gets mitigated “before the risk occurs.” Her example: you ask for white Nike sneakers in size 9, the agent submits Adidas — the transaction never clears.
The agent never touches your real card number. It gets a tokenized credential scoped to the authorized intent, similar in spirit to Stripe Issuing’s single-use virtual cards for agent checkout.
Earned insight: I’ve run corporate card programs where a single shared card number lived in five different SaaS billing portals, and reconciling mystery charges took my team half a day a month. Agent-initiated spend multiplies that problem by every employee with a chatbot. The Intent ID is the first mechanism I’ve seen that gives finance an audit key — which agent, authorized by whom, for what — at the network level instead of in a spreadsheet after the fact. That’s the part worth paying attention to, not the token plumbing.
The Competitive Landscape: Every Network Is Building This
Amex didn’t act in a vacuum. Every payment network shipped agentic commerce infrastructure in the first half of 2026, and they took different architectural bets. If you’re standardizing corporate spend, you need to know which rails your cards run on.
| Dimension | Amex Intent Intelligence | Visa Intelligent Commerce | Mastercard Agent Pay | Stripe Agent Toolkit | Forter Agentic Commerce Engine |
|---|---|---|---|---|---|
| Announced | Oct 13, 2026 | Apr 2026 (+ Connect, Apr 8) | Apr 2026 (+ AP4M, Jun 10) | 2025–2026 (ongoing) | Oct 13, 2026 |
| Core mechanism | Proof of Intent Token + Intent ID | Tokenization, auth, commerce signals APIs | Agentic Tokens + Verifiable Intent | Virtual cards (Issuing) + agent toolkit + Radar | Merchant-side agent verification + orchestration |
| Where intent lives | Issuer side, pre-transaction | Shared APIs across ecosystem | Identity + intent + behavioral + fraud insights (expanded Sep 30) | Merchant/developer integration | Merchant side, real-time traffic decisioning |
| Card scope | Amex network | Visa + non-Visa cards via Connect | Cards, bank accounts, stablecoins | Any (processor-level) | Network-agnostic (merchant-side) |
| Enterprise angle | Corporate card + T&E natural fit | ChatGPT checkout integration (Jun 10) | 30+ partners incl. Stripe, Coinbase, Adyen | Developer-first, MCP-native | No-code merchant integration |
| Standards posture | Chairs EMVCo agentic working group; framework offered as industry standard | EMVCo participant | EMVCo participant | Open Agentic Commerce Protocol (ACP) with OpenAI | Fraud-signal contributor |
| Status | Framework published; partner discussions ongoing | Live pilots (Aldar/UAE real-estate fees, AWS, Mesh) | Live; AP4M for machine-to-machine micro-payments | Live | Live today |
But the standards story is the one that matters most. EMVCo — the body owned by Amex, Visa, Mastercard, Discover, JCB, and UnionPay — published its draft “Agentic Payments Framework for Specifications v1.0” on September 1, with the comment period closing September 30. Its centerpiece is a neutral “Intent Services” layer: shared infrastructure where any participant can register, retrieve, and update consumer-authorized intent. Amex chairs that working group. So today’s announcement is best read as Amex shipping its implementation of a standard it helped write — and inviting the industry to converge on intent-first architecture.
Warning: Don’t mistake “industry standard” for “works everywhere.” EMVCo’s framework is still a draft. Amex’s implementation is Amex-network only. Visa’s Connect supports multi-network cards but through Visa APIs. If your card program mixes Amex corporate cards with Visa procurement cards — and most do — you have two different intent architectures to track, with no interoperability layer shipping yet.
What Works About the Amex Approach
Intent travels with the payment, not beside it
Most fraud controls evaluate a transaction after the agent submits it — amount, merchant, velocity, device fingerprint. The problem Sardine’s risk team flagged is that agent traffic breaks behavioral signals built for humans: there’s no typing cadence, no mouse movement, no familiar device. Amex sidesteps the signal problem entirely by binding the authorization constraint to the credential itself. The check isn’t “does this look fraudulent” — it’s “does this match what was approved.” That’s a fundamentally better question.
The dispute-resolution hook is real
Amex explicitly positions the Intent ID and token as inputs to dispute resolution. Today, when an employee disputes a corporate card charge, the investigation is archaeology: receipts, emails, memory. With an intent log, the issuer can see exactly what was authorized versus what was charged. Deloitte’s legal-risk analysis warned that agent-initiated purchases will “test the limits of traditional fraud models” through unauthorized chargebacks. An intent receipt doesn’t eliminate disputes, but it compresses a two-week investigation into a database lookup.
Pre-transaction posture beats post-transaction cleanup
Forter’s data makes the urgency concrete: agentic orders up 4x since August, automated fraud doubled in two months. Every other vendor in this space is building better detection. Amex is building prevention — declining the Adidas order before it posts rather than refunding it after. For finance teams, prevented transactions don’t generate reconciliation work. Refunded ones do.
Intent Intelligence Strengths:
- Intent bound to the payment credential, not evaluated after the fact
- Intent ID creates a network-level audit trail (agent, authorizer, constraints, timestamp)
- Editable and cancellable intents keep the cardholder in control
- Offered as an industry framework, not a proprietary lock-in — Amex chairs the EMVCo working group writing the shared spec
- Builds on April’s ACE toolkit and Agent Purchase Protection commitment, so it’s an iteration, not a day-one bet
Intent Intelligence Weaknesses:
- Framework, not enforcement — nothing happens until agents, merchants, and issuers implement it
- Amex-network only; mixed card programs need parallel architectures for Visa/Mastercard
- No pricing, SLA, or liability terms published — the “who pays when it fails” question is unanswered
- Thousands of weekly transactions is a pilot-scale signal, not production proof
- Doesn’t address the procurement-system layer (Coupa, Ariba) where enterprise spend policy actually lives
Where It Struggles: The Enterprise Gaps Nobody’s Covering
Your procurement system isn’t in the loop
Here’s the catch: enterprise spend doesn’t flow through consumer authorization patterns. When an employee buys software, the approval chain runs through Coupa or SAP Ariba — budget owner, category manager, sometimes legal — before a card gets charged. Amex’s framework authenticates the cardholder’s intent. It has no concept of a procurement policy, a preferred vendor list, or a spend category code. An agent that faithfully buys exactly what the employee asked for can still violate every procurement rule your company has. The framework prevents agent error. It doesn’t prevent policy violation. Nobody in the day-one coverage made this distinction, and it’s the one your CPO will ask about first.
Liability terms are conspicuously absent
The Yahoo Finance liability-gap analysis from October 6 — one week before this announcement — nailed the core issue: merchants are being asked to absorb the economic risk of agent mistakes they didn’t cause. Amex’s April announcement extended “Agent Purchase Protection” to registered-agent transactions, which is a concrete liability backstop competitors haven’t matched. But the Intent Intelligence announcement contains zero language about who pays when the framework itself fails — when intent was captured correctly and the wrong purchase still cleared. Until liability terms ship, treat this as a risk-reduction tool, not a risk-transfer one.
The consumer/enterprise split is unaddressed
Everything in the announcement — Trendex survey stats, the Nike-vs-Adidas example, paper-towel reorders — is consumer framing. But Amex’s card base skews heavily corporate: business travel, T&E, procurement cards. Corporate cards have different authorization patterns (delegated spend, per-diem caps, MCC restrictions), different dispute processes, and different regulatory exposure. The framework will need an enterprise profile — delegated authority, spend-category constraints, multi-level approval — before a controller can bless it. That profile doesn’t exist yet.
Earned insight: In every corporate card rollout I’ve been part of, the failure mode was never the technology — it was the policy gap between what the card could do and what the company allowed. Virtual cards with per-transaction limits existed for years before finance teams actually configured them. Intent tokens will follow the same curve: the capability ships in 2026, but most companies won’t have an agent-spend policy until their first five-figure surprise charge, probably sometime in 2027. Write the policy now and you’re ahead of 90% of your peers.
For enterprise buyers, the ecosystem’s strengths:
- Multiple rail options (Amex, Visa, Mastercard, Stripe) means no single-vendor dependency for agent payments
- EMVCo draft framework signals genuine industry convergence, not a format war
- Forter’s merchant-side verification covers the acquiring side Amex doesn’t touch
- Tokenized credentials eliminate the “agent holds the real card number” nightmare scenario
Ecosystem weaknesses:
- No interoperability between network implementations yet — mixed card programs face fragmentation
- Merchant adoption is the bottleneck; frameworks don’t matter if checkout doesn’t check the token
- Fraud is scaling faster than defenses (Forter: automated fraud doubled in two months)
- Consumer trust lags capability (55% uncomfortable with agent purchasing per Riskified)
Pricing Reality
There’s no price list for the Intent Intelligence Framework — it’s a published specification, not a SKU. But “free framework” doesn’t mean free adoption. Here’s what it actually costs an enterprise to act on this announcement, verified October 13, 2026:
| Cost layer | What you pay | Verified source |
|---|---|---|
| Framework itself | $0 — published spec, no license fee | Amex newsroom / TechCrunch Oct 13 |
| ACE Developer Toolkit integration | Engineering time: expect 2–4 sprints for agent-intent capture in a custom agent; prebuilt integrations via partners TBD | Amex ACE developer docs (Apr 2026 launch) |
| Merchant-side verification (Forter) | Custom enterprise pricing; Forter typically prices on transaction volume processed | Forter pricing page (custom quotes only, verified Oct 2026) |
| Dispute cost avoided | $15–$25 per chargeback in network fees alone, plus 5–10 hours of analyst labor per corporate dispute | Industry-standard chargeback fee schedules |
| Policy work (the real cost) | 20–40 hours of finance + legal + IT time to write an agent-spend policy, update cardholder agreements, and brief approvers | Practitioner estimate |
The math that matters: a single wrong agent purchase on a corporate card — say a $4,000 laptop misconfiguration — costs more in investigation labor than the entire policy-writing exercise. The framework is free. The absence of a policy is what’s expensive.
Tip: Don’t wait for vendor integrations. The highest-ROI move this quarter is a one-page addendum to your corporate card policy: (1) employees must disclose agent-initiated purchases over a threshold you set, (2) agents may only use designated virtual card numbers, never the physical corporate card, and (3) disputed agent transactions get investigated against the agent’s conversation log. That covers 80% of the risk with zero vendor dependency.
Who Should Act Now — and Who Can Wait
Good fit — act this quarter:
- Companies running Amex corporate card programs. You’re the native audience. Talk to your Amex rep about ACE pilot participation and get your developers the Intent API docs.
- Finance teams seeing early agent-spend. If employees are already expensing AI-assisted purchases, the dispute wave Forter measured (4x volume, 2x fraud) is coming to your ledger.
- Regulated industries (financial services, healthcare, government contractors). Your auditors will ask about agent-initiated spend controls within 12 months. An intent-based framework maps cleanly to existing separation-of-duties requirements.
- IT teams standardizing on a single agent platform. If you’ve centralized on ChatGPT Enterprise, Copilot, or Claude with corporate cards attached, you can implement intent capture at one integration point.
Not a good fit — monitor, don’t deploy:
- Visa/Mastercard-only card programs. Amex’s framework doesn’t cover your rails. Track Visa Intelligent Commerce and Mastercard Agent Pay instead — and push your network rep on EMVCo interoperability timelines.
- Companies without agent adoption. If nobody’s buying through AI assistants yet, a framework integration is premature. Write the one-page policy addendum and move on.
- Heavy Coupa/Ariba shops expecting procurement integration. The framework doesn’t speak procurement. Wait for the enterprise profile or a Coupa/Ariba connector before investing engineering time.
Bottom Line
Amex’s Intent Intelligence Framework is the most enterprise-relevant agentic commerce announcement of 2026 — not because the technology is revolutionary (tokenized credentials and pre-auth constraints are established patterns), but because it’s the first payment-network move that gives finance teams an audit key for agent spend: who authorized what, through which agent, under which constraints. Combined with Forter’s same-day merchant-side launch and the EMVCo draft nearing finalization, the intent-first architecture is winning. That’s the direction all four networks are converging on.
But the announcement leaves the two questions enterprise buyers actually need answered — liability terms and procurement-system integration — completely untouched. Treat Intent Intelligence as a strong foundation signal, not a deployable control. Your move in the next 30 days isn’t a vendor integration. It’s writing the agent-spend policy addendum your card program needs before the first surprise charge hits: designate virtual card numbers for agent use, set disclosure thresholds, and require conversation-log retention for disputed transactions. The frameworks will mature. Your policy should already exist when they do.
FAQ
What is the Amex Intent Intelligence Framework?
Published October 13, 2026, it’s American Express’s specification for making AI-agent purchases match what the customer actually authorized. The flow: a Create Intent API captures purchase instructions (merchant, budget, preferences), the customer authenticates and signs off, the system generates an Intent ID plus a Proof of Intent Token, and that token is exchanged for a single-use payment credential the agent uses at checkout. If the agent’s cart doesn’t match the token constraints, the transaction is declined. Amex hopes it becomes an industry standard and chairs the EMVCo working group writing the shared spec.
How does the Proof of Intent Token actually work?
The token encodes the customer’s high-level purchase constraints — merchant name, maximum budget, product preferences — captured at authorization time. When the agent reaches checkout, it presents the token instead of the real card number and receives a tokenized payment credential scoped to those constraints. Downstream, the Intent ID and token serve as controls for fraud prevention, authorization decisions, and dispute resolution. Think of it as a virtual card number with the purchase policy cryptographically attached, rather than enforced by a separate system.
Does Intent Intelligence work with Visa or Mastercard?
No. It’s Amex-network native, built on Amex’s ACE Developer Toolkit. Visa has its own Intelligent Commerce APIs (which notably support both Visa and non-Visa cards through Intelligent Commerce Connect), and Mastercard has Agent Pay with Agentic Tokens plus Verifiable Intent. The long-term interoperability bet is EMVCo’s draft Agentic Payments Framework v1.0 — whose comment period closed September 30, 2026 — which proposes a neutral Intent Services layer all networks could share. Until that ships as a final standard, mixed card programs need to track each network’s implementation separately.
What does it cost to adopt?
The framework itself is a free published specification with no license fee. Real costs are integration engineering (2–4 sprints for custom agent implementations), merchant-side tooling like Forter’s Agentic Commerce Engine (custom enterprise pricing by volume), and — biggest of all — the internal policy work: 20–40 hours across finance, legal, and IT to write agent-spend rules and update cardholder agreements. Offset against that: each avoided corporate card dispute saves $15–$25 in network fees plus 5–10 hours of analyst investigation time.
How does this affect corporate cards and T&E policy?
Directly, though Amex framed the launch in consumer terms. Corporate Amex programs are the natural first enterprise adopters. But the framework authenticates cardholder intent, not procurement policy — it won’t enforce preferred vendors, spend categories, or multi-level approvals from Coupa or SAP Ariba. Finance leaders should treat it as a complement to (not replacement for) procurement controls, and should immediately add an agent-spend addendum to cardholder agreements covering disclosure thresholds, designated virtual card numbers, and conversation-log retention.
Who is liable when an AI agent makes a wrong purchase?
This is the unanswered question. Amex’s April 2026 announcement extended Agent Purchase Protection to registered-agent transactions, but the October Intent Intelligence publication contains no liability terms for framework failures. The October 6 Yahoo Finance liability-gap analysis — published a week before the announcement — argued merchants are currently absorbing risk they didn’t create. Until Amex, Visa, or Mastercard publish explicit liability allocation for intent-framework transactions, enterprises should assume the cardholder (i.e., your company) retains dispute risk and should budget accordingly.
How is this different from Visa Intelligent Commerce or Mastercard Agent Pay?
Architecturally, Amex pulls intent capture earliest — before the transaction, as a pre-condition of the credential. Visa’s approach centers on shared APIs (tokenization, authentication, commerce signals) with broad ecosystem pilots including ChatGPT checkout. Mastercard splits the problem: Agent Pay for consumer-agent transactions plus Agent Pay for Machines (June 2026, 30+ partners) for autonomous machine-to-machine micro-payments across cards, bank accounts, and stablecoins. All three are converging on EMVCo’s intent-layer concept. The practical difference today is which network your cards run on.
What should IT leaders do in the next 30 days?
Four things, none requiring vendor integration: (1) Add a one-page agent-spend addendum to your corporate card policy — disclosure thresholds, designated virtual cards for agent use, conversation-log retention. (2) Ask your Amex/Visa/Mastercard rep for their agentic commerce roadmap and pilot options. (3) Audit whether employees are already making agent-assisted purchases on corporate cards — check expense reports for AI-tool subscriptions paired with unusual merchants. (4) Brief your procurement team that network-level intent frameworks don’t enforce procurement policy, and scope the integration gap with Coupa or Ariba. Vendor deployments come later; governance comes now.
Related Articles
Discussion