Card Fraud Defense Rests on One Assumption. Agentic Commerce Voids It.
Every control we built to stop card fraud quietly assumes a human is present at checkout. Agents remove the human. The answer is not sharper detection. It is authorization decided before the transaction executes.
Eight years inside credit card authorization taught me one uncomfortable thing. The entire fraud stack is a presence detector. Device fingerprinting, behavioral biometrics, velocity rules, the step-up prompt that asks you to confirm a code. Every layer exists to infer that a specific human is on the other end of the session. We do not verify intent. We verify presence, and we assume the rest follows.
That assumption held because a human was always, eventually, there. Card-not-present fraud was the hard case, and the industry answered it by getting better at spotting the person. 3-D Secure, risk scoring, one-time passcodes, models trained on how real people move and hesitate. After Covid, BNPL bolted a thinner layer of credit onto the same checkout with weaker identity checks, and fraud found it within months. Every fix in thirty years has been a better presence detector. Nobody questioned the premise, because the premise was always true.
It is about to stop being true.
What agentic commerce voids
Agentic commerce removes the human from checkout by design. The pitch is simple, and people want it: let my agent handle it, I do not have the time. Headless commerce, where software represents the buyer end to end, is arriving faster than the fraud stack is ready for. When the buyer is an agent, presence signals do not degrade gracefully. They vanish, or worse, they get emulated perfectly. An agent is a clean, patient, tireless actor that looks exactly like an ideal customer until the moment it is not. Detection built to find the anomalous human has nothing to find when there is no human and no anomaly.
Run it forward and the economics of the card break. Fraud losses climb. Chargebacks climb. Merchants who already push card fees onto consumers push harder, or they start refusing card-not-present traffic that originates from an agent. The rails that made frictionless commerce possible add friction back, or back away from the segment. Commerce drifts toward cash and closed-loop wallets, a worse outcome for everyone who spent two decades building the open card economy.
The Presence Assumption
I call the flaw the Presence Assumption. It is the belief, baked into every fraud control we run, that authenticating the human is the same as authorizing the transaction. It never was. It only looked that way while a human was guaranteed to show up.
The answer is not a better guess about who is typing. The answer is to authorize the transaction itself, at the moment it happens, against identity and context resolved live across the systems that already hold the truth, with a confidence score, before the action executes. Not an audit assembled afterward. Not a model scoring the odds a real person is present. A decision about whether this specific transaction should proceed, made against premises the system sourced rather than the ones the agent asserted.
This is federation over mastery applied to commerce. There is no master profile of the buyer to forge and no canonical record to spoof. Identity and standing are resolved in flight, at transaction time, from the authorities that already know. The question moves from who is present to whether this action, by this actor, on this target, should execute right now. Presence becomes an input. It stops being the verdict.
Where the industry has to spend next
The implication is a reordering of where the industry spends. Merchants should stop pouring the entire fraud budget into sharper presence detection and demand a decision layer that authorizes the agent's transaction on its own merits. Networks should accept that the amount and the authority have to travel together, so the authorized amount and the settled amount are one object rather than a hope. Regulators should require that an agent-driven purchase names an accountable party and can prove a decision was made before execution, not reconstructed from logs a quarter later. The winners will be whoever moves authorization upstream of presence.
Card fraud was never solved. It was deferred, behind an assumption we never had to defend. Agentic commerce is the bill coming due.
You cannot detect a human who was never there. You can only authorize the transaction that is.
Cross-link: this piece applies the argument from N° 020 (The Authorization Gap) to card-not-present commerce and coins "the Presence Assumption" as the specific version of the flaw that thirty years of card fraud defense has been standing on. Direct extension of N° 029 (Detection Is Downstream of Authorization), which argued that classification after the action is not authorization at the moment of action; presence detection is a specific case of that pattern. Companion in spirit to N° 030 (A Signature Is Not a Premise) on the same failure mode applied to signed records, and to N° 018 (The Question No One Signed) on who is accountable when the human is not the actor. BNPL fraud growth rate and weak-identity-check pattern verified against LexisNexis Risk Solutions and Juniper Research BNPL fraud reports (2025 to 2026). Card-not-present controls (3-D Secure, one-time passcodes, device fingerprinting, behavioral biometrics) are established industry practice with EMV 3DS specification as the standard reference. First-person voice in this essay is a deliberate departure from house convention, warranted by the payments-industry credential the argument leans on.