Microsoft flags 1 million-plus AI-crafted scam emails seeking nearly $50,000 ACH transfers, success unconfirmed

One executed ACH transfer out of a million emails is a conversion rate of 0.0001%, and it moves nearly 50,000 USD to the attacker ^[18]. Whether the campaign got that one hit is not in the published material. The number of replies, invoice approvals, account registrations and executed transfers is not shown ^[6].

The build order is more instructive than the volume:

1. Register similar domains and prepare accounts with third-party email delivery services ^[4]. 2. Fill a template with the target company executive's name, ServiceNow branding, and the destination bank account ^[3]. 3. Combine the executive approval text, the fake invoice and a fabricated forward thread into a single email sent to accounts payable staff ^[19]. 4. Prompt the recipient to request a PDF version or to execute the ACH transfer ^[14].

The destination account is a template field, sitting next to the executive's name.

The fabricated thread is body text, which is why the tells are structural. Microsoft's detection notes point at forward histories lacking normal forward headers or indentation, and at a sender address or Reply-To that does not match the executive display name ^[13]. The threads are described as plausible but carrying contradictions in format or content ^[21]. For anyone reading logs, the instruction is to distinguish fake forward headers in the body from actual delivery history ^[11].

Authentication does not close this. Microsoft's own guidance says successful email authentication does not guarantee the legitimacy of billing details or transfer destinations, and pairs SPF, DKIM and DMARC with impersonation detection on executive display names and lookalike domains ^[9]. The loss event is narrower than the mail flow: financial loss occurs when the transfer executes, and approval or account registration alone does not mean the transfer is complete ^[15]. That is where the recommended controls sit, as callback to pre-registered contact details and multiple approvals for new or changed transfer destinations ^[10]. A control that fires only on a changed destination is cheap to run, because vendors do not change bank details often.

The telemetry is thin. Malware execution on devices is not observed in this case ^[7], and compromise of the target company or of ServiceNow accounts has not been confirmed ^[8], so endpoint and sign-in logs start empty. Where the inducement is purely email, recipient web access may not occur at all, and the absence of matching DNS or web traffic does not rule out a reply or a transfer ^[12]. What remains is header data: From, Reply-To, source IP, authentication results, Message-ID, recipient, and the delivery or quarantine result ^[11], with domain registration dates checked through registration data such as RDAP rather than DNS logs ^[17].

The generative-AI attribution is an inference from artifacts. Microsoft gives the reason for its high severity rating as the scale of over a million emails combining executives, vendors, invoices and conversation history, with evidence of template generation consistent with generative AI usage ^[5]. For that inference to change a control decision, it would have to change either the cost per variant or the quality of the pretext beyond what a mail merge with a CRM export already produces, and the published material shows neither figure. The scale and the templated destination account are enough on their own to argue the check belongs in the payment process.