Microsoft: 1M AI Invoice Scam Emails Hit US Firms

A million emails in three days is about four messages every second, sustained, for 72 hours. That is the volume Microsoft researchers attributed to one invoice fraud campaign in early August 2026, and scale is only half the story. Every message carried a small fabricated world inside it: an executive's name in the sender field, a vendor's branding on a PDF, and a forwarded conversation that never happened. The Record first reported Microsoft's findings on September 11.

Key Takeaways

• Microsoft detected more than 1 million fraudulent invoice emails sent to corporate mailboxes between August 3 and August 5, 2026.
• Each message asked accounts payable staff to push an ACH transfer of roughly $50,000 to an attacker controlled account.
• American organizations made up about 88% of recipients, and the mail was relayed through accounts at a legitimate third party email service.
• The emails combined executive impersonation, ServiceNow branding, counterfeit invoices, and a fabricated forwarded thread into one narrative.
• Microsoft found HTML artifacts consistent with AI assisted template building but could not confirm how much content a model wrote.

What Did Microsoft Actually Find?

Microsoft found a high volume ACH payment fraud campaign that impersonated the CEO, CFO, or president of each target company and told accounts payable to release nearly $50,000. The sender display name, the reply to display name, and the signature all carried the executive's name, so the message passed a glance test three times before anyone read the body.

The messages went out through accounts at a legitimate bulk email delivery provider, not freshly registered attacker domains. That detail matters more than the AI headline. Mail relayed through a reputable sending platform inherits its IP reputation and passes SPF and DKIM for the platform's own domain, so a strict DMARC policy on your corporate domain never gets a chance to fire. The spoof lives in the human readable display name, not the envelope. It is the same weakness behind the 7.6 billion phishing threats Microsoft blocked in a single quarter: attackers stopped fighting authentication and started renting infrastructure that already passes it.

How Does the Fake Email Chain Work?

The attackers embedded a fabricated forwarded thread showing the target's executive and a ServiceNow representative apparently discussing a purchase, so the payment request arrived with its own backstory. A traditional invoice scam gives the victim one lure. This campaign gave the victim a paper trail.

Each layer answers the objection raised by the one before it:

• Executive impersonation supplies the authority to bypass normal review.
• ServiceNow branding supplies a plausible vendor a large company would genuinely pay.
• A counterfeit invoice supplies the document accounts payable expects to see attached.
• The fabricated thread supplies the approval history that would otherwise be missing.

Fake context as a payload is not new. It is the same escalation we saw in dual channel BEC, where a scam email is followed by a live phone call. What changed is cost: a phone call needs a person, a forged thread needs a template. CISA's phishing guidance still centers on urgency and unfamiliar senders. Neither cue is present here.

Why Did 88% of Targets Sit in the US?

American companies dominated the target list because ACH is the default rail for US business to business payments: fast, high limit, and hard to claw back once settled. The FBI's 2025 Internet Crime Report put business email compromise losses at $3.04 billion across 24,768 complaints, up from $2.77 billion a year earlier.

Run the arithmetic. At $50,000 per request, a hit rate of one in ten thousand across a million emails still nets about $5 million. Attackers do not need accounts payable to be gullible. They need it to be busy.

Phishing and spoofing generated roughly eight times more IC3 complaints than BEC in 2025, yet BEC produced far larger total losses, the same split we unpacked when the FBI reported $17.6 billion in cyber fraud losses for the prior year.

What Makes Microsoft Think AI Wrote These Emails?

Microsoft pointed to structural artifacts in the message HTML rather than anything in the prose: extensive HTML comments, regular section labeling, delimiter banners, and template construction that stayed uniform across an enormous sample. Those are the fingerprints of code generated to a spec, not of a human copying a template by hand.

Microsoft stated it could not independently establish how much content a model produced. Nick Tausek of Swimlane framed the trend plainly to The Record: attackers are using AI to make established fraud schemes more sophisticated and scalable. Scalable is the operative word. The tactic itself is decades old.

Most coverage treats the AI angle as the news. The sharper reading is that generation cost fell far enough to make per target customization economic at million message volume, the same dynamic behind AI writing most phishing emails now reaching inboxes. Detection built on typos has no signal left to read.

What This Means for Your Inbox

If you have ever approved, forwarded, or been copied on a vendor invoice, you sit inside this campaign's addressable market. The impersonation only works because most of us treat a familiar name in the From field as identity. It is not. The display name is a free text field any sender can fill with anything.

There is a reconnaissance dimension too. Campaigns at this scale need to know which mailboxes are live before investing in a tailored follow up, and remote images report back the moment a message renders, a technique we covered in how attackers use tracking pixels to find live inboxes.

Nothing here depended on a Microsoft 365 specific weakness, so the identical playbook works against Google Workspace, Zoho, and any provider where a bulk sending platform can relay mail with a spoofed display name.

How Do You Stop a $50,000 ACH Request?

You stop it with process, because the message itself is designed to survive inspection. Four controls do most of the work:

• Verify out of band, always. Confirm any new payee or changed bank detail by calling a number already on file for that vendor, never a number printed in the email.
• Require two approvers above a threshold. Pick a number finance can live with and make single approver release impossible above it. Executive impersonation fails the moment authority alone cannot move money.
• Tag external mail visibly. A message signed by your own CEO but banner flagged as inbound from outside is a contradiction anyone can see.
• Harden authentication at the edges. Publish DMARC at p=reject, follow NIST SP 800-177 Trustworthy Email for SPF and DKIM alignment, and move finance staff to phishing resistant authenticators per NIST SP 800-63B.

Compliance teams should note the clock. Because most BEC losses move by wire or ACH, recovery windows run in hours, and a detection that lands the next morning lands too late.

What to Watch Next

Watch whether bulk email delivery providers tighten display name controls on accounts sending to corporate recipients, because that is the pressure point this campaign exposed. Microsoft's threat intelligence reporting is the best public place to track whether the technique spreads.

ServiceNow was chosen because enterprises genuinely pay ServiceNow invoices. Salesforce, Workday, and every other large recurring vendor fit the template equally well, and swapping branding costs an attacker nothing. Assume the next million messages carry a different logo.