Your Boss Didn't Email You. An AI Trained on His Last Three Replies Did.

The old tell for a fake-boss email was bad grammar and a weird greeting. That tell is gone. Generative tools now ingest a scraped email thread and mimic a colleague's exact cadence, punctuation, and sign-off, which is why business email compromise still drains companies — the FBI's Internet Crime Complaint Center logged $2.9 billion in reported BEC losses in 2023 alone. The only reliable defense left is moving the conversation to a different channel before you act.

Why can't I trust an email that sounds exactly like my manager?

Because "sounds like him" is now the easiest part to fake. BEC used to rely on a hijacked or lookalike address paired with clumsy text; the awkward phrasing was the giveaway. Feed a large language model three real replies from someone and it will match their rhythm — the way they open with "Hey," skip punctuation when rushed, sign off with just a first initial. Attackers get those replies from a compromised mailbox, a forwarded thread, or a reply-all chain that leaked outside the company.

Verizon's 2024 Data Breach Investigations Report found the human element involved in roughly 68% of breaches, with social engineering and pretexting a dominant pattern. Style is no longer evidence. The person who writes like your CFO might be a model trained on fifteen minutes of your CFO's outbox.

What does one of these upgraded fakes actually look like?

It looks like a short, plausible, slightly urgent request that fits a real conversation already in progress. The classic move: a reply dropped into an existing thread about a vendor payment, with new wire instructions and a line like "bank flagged the old account, use this one." No typos, correct project names, your boss's real speech tics.

The technical catch is that the display name can read perfectly while the sending domain is off by a character — @company-invoices.com instead of @company.com — or the Reply-To quietly points somewhere else. Authentication helps here: a message that fails SPF, DKIM, or DMARC alignment is suspect no matter how human it reads. AI mail tools, including the scam and invoice-fraud flagging built into AI email with scam flagging, check those headers automatically, because a person skimming on a phone never will.

So how do I actually verify a request I can't trust on its face?

Switch channels and confirm through one you already know, not one the email gave you. If the email asks for money or credentials, call the person on a number from your contacts or company directory — never the number in the signature. Walk to their desk. Message them on the internal chat tool. The whole scam depends on you staying inside the email, where the attacker controls both sides.

Make it a rule, not a judgment call. Any payment change, any wire, any new banking detail gets a second-channel confirmation, full stop — the IC3 consistently recommends exactly this verbal verification for BEC. Rules survive a busy Tuesday afternoon; instincts don't. And stop treating a convincing writing style as proof of anything. The one thing a model trained on scraped replies can't fake is your boss picking up the phone and saying he never sent it.