AI-Enabled Business Email Compromise: The Old Warning Signs Are Not Enough
Updated: Oct 5
Business Email Compromise, or BEC, has been financially devastating for years. It does not need malware, a malicious attachment, or a sophisticated technical exploit to work. A convincing message, sent at the right moment, can be enough to trigger a payment, change banking information, release sensitive data, or alter another business process.
Artificial intelligence adds another layer to an already effective form of fraud.
Generative AI can help attackers produce polished, context-specific messages quickly. Public company websites, social media, press releases, presentations, job postings, compromised mailboxes, and other sources can provide the names, roles, terminology, relationships, and business context needed to make those messages believable.
In one 2024 sample, 40% of BEC emails appeared to be AI-generated. VIPRE Security Group reported the estimate, but it should be interpreted cautiously. AI-content detection is imperfect, and the analysis does not establish that 40% of all BEC worldwide is generated by AI.
Still, the finding points to an important shift.
The old warning signs have not disappeared.
They were just never strong enough to carry the defense by themselves.
BEC Does Not Have to Look Like Phishing
Business Email Compromise is fraud built around trusted business communication.
Sometimes an attacker creates an external account that resembles a real executive, vendor, employee, or business partner. In other cases, the attacker compromises a legitimate mailbox and sends fraudulent requests from the actual account.
That second scenario is especially important.
A fraudulent message may have the correct sender address. It may appear inside an existing conversation. It may refer to a real invoice, transaction, customer, or project because the attacker has been monitoring the account.
There may be no suspicious attachment. No broken English. No obviously strange link. Nothing that screams phishing.
The attack succeeds when the message influences a real business process.
AI Makes Some Old Advice Less Reliable
Employees have long been taught to look for spelling errors, awkward grammar, unusual phrasing, generic greetings, and other signs that something about an email feels wrong.
Those indicators can still be useful. They just cannot be treated as requirements for fraud.
Sophisticated BEC existed well before modern generative AI. Attackers were already researching organizations, studying executive communication styles, compromising accounts, monitoring legitimate email threads, and timing requests around real transactions.
AI did not invent that playbook.
What it can do is reduce the effort required to execute parts of it.
An attacker can generate professional language, revise tone, translate messages, imitate business terminology, and create multiple versions of a communication quickly. The FBI has warned that generative AI can help criminals make fraudulent communications more believable while reducing some of the spelling and grammar problems that once served as warning signs.
A fraudulent message can now be grammatically perfect, specific to the organization, and consistent with the way a legitimate business request might actually be written.
If the mailbox itself has been compromised, even the sender address may be real.
At that point, "Does this email look legitimate?" becomes a weak control.
BEC Targets Business Processes, Not Just Employees
An email alone does not create the financial loss. Something has to happen after the email arrives.
A payment gets approved. Vendor banking information gets changed. Payroll is redirected. Credentials are reset. Sensitive information is released. Access is granted. A normal procedure gets bypassed.
The employee's decision matters, but that decision sits inside a larger process. Consider two organizations receiving the same convincing fraudulent request to change a vendor's bank account.
At the first company, one employee can make the change based on an email.
At the second, banking changes must be independently confirmed through established vendor contact information and approved through a separate workflow.
The employee at either organization may believe the email. The consequences are very different. That changes the security question.
It matters whether employees can recognize a fraudulent email. But so does what happens if they do not. If one convincing message can trigger a high-value transaction or sensitive disclosure with no other control intervening, the organization has made individual judgment carry too much weight.
AI may make deception easier.
The process should make the consequences harder.
Verification Should Work Even When the Email Looks Perfect
Strong BEC defenses do not depend entirely on proving that a message is fake. They verify the action being requested. The FBI recommends independently confirming payment and purchase requests and verifying changes to account numbers or payment procedures through trusted contact information.
That distinction matters.
If a vendor sends an email saying future payments should go to a new bank account, an employee should not have to determine with certainty whether the email was written by the vendor, an attacker, or an AI system.
The banking change should have its own verification process.
That may involve calling a previously verified number, using an established vendor-management workflow, requiring a second approval, or confirming the change through another trusted system. The exact control will vary by organization.
The principle does not:
The objective is not simply to decide whether the message is authentic. The objective is to establish that the requested action is authorized.
That approach still works if the email contains spelling mistakes. It works if the email looks flawless. It can even work if the attacker is sending from the vendor’s legitimate compromised account.
That is what makes verification a business control instead of a phishing-detection trick.
The Psychology Has Not Changed Much
AI can change the quality of the message, but the pressure behind BEC is familiar.
Attackers still lean on authority, urgency, trust, routine, secrecy, and the desire to be helpful.
The CEO needs the payment processed immediately.
The vendor changed banks and needs today's invoice sent to the new account.
The executive is traveling and cannot talk.
The transaction is confidential.
The request cannot wait until tomorrow.
The employee is not necessarily ignoring obvious danger. The request may fit closely enough with normal work that stopping feels less natural than acting.
That is one reason awareness training has limits. Someone can know exactly how BEC works and still encounter an unusually convincing request during a busy day.
Knowledge helps. The surrounding environment matters too.
Security Culture Determines Whether Verification Is Usable
A company can write a good verification procedure and still make it difficult for employees to use. Suppose a finance employee receives what appears to be an urgent payment request from the CEO. Policy requires independent verification.
Now suppose that same organization rewards speed, treats executive requests as untouchable, or makes employees feel that questioning senior leaders is disrespectful.
The control exists on paper. In practice, the employee has been taught that using it carries a social cost.
That is a security culture problem.
Employees should know that delaying a high-risk action long enough to verify it is expected behavior, even when the request appears to come from someone powerful or impatient.
Leaders reinforce that expectation through their own behavior.
If executives routinely bypass normal controls when something is urgent, employees learn very quickly which rule actually matters.
Saying “I need to verify this before I proceed” should not feel like an accusation.
It should feel like the process working properly.
Do Not Turn Employees Into AI Detectors
Employees should know that attackers can use generative AI to create convincing social-engineering content. They do not need to become forensic analysts.
Asking an employee to determine whether an email was written by AI is usually the wrong problem. A fraudulent email written entirely by a person is still fraudulent. A legitimate email drafted with AI is still legitimate.
What matters is the request.
Does it involve money? Access? Sensitive information? A change to an established account? A request to bypass a normal procedure? If so, the employee needs to know when additional verification is required, how to perform it, and where to escalate uncertainty.
Training can support those decisions. It should not create the impression that employees are expected to identify every synthetic communication correctly.
Perfect detection is not realistic. A resilient process is.
Email May Be Only the First Step
BEC is also becoming harder to think about as an email-only problem.
An attacker can reinforce an email with a phone call, text message, cloned voice, manipulated video, or fraudulent messaging account.
Moving to another channel may feel like verification even when it is not independent verification.
The 2024 fraud involving engineering firm Arup illustrates the problem. A finance employee in Hong Kong reportedly questioned an initial phishing message. The employee was later placed into what appeared to be a video conference with the company’s CFO and other colleagues.
Hong Kong authorities said the conference had been fabricated using deepfake technology. The employee ultimately authorized transfers totaling about HK$200 million, roughly US$25.6 million.
The employee had doubts. Additional evidence appeared to resolve them. The evidence itself was fake.
HumanSecIQ examines that problem separately in AI Voice and Video Impersonation: When You Can No Longer Trust What You See or Hear.
The case also makes a broader point for BEC defense. Changing channels is not enough if the attacker controls the new channel too.
Verification needs to be independent of the communication being verified.
What Leaders Should Be Looking At
AI-enabled BEC should push organizations to look beyond phishing click rates and annual awareness training.
Those questions reveal far more about BEC resilience than whether everyone can identify a perfectly simulated phishing message.
One question matters especially:
What controls remain if the employee believes the communication is legitimate?
If the answer is "none," the organization is relying on detection to work every time. Eventually, it will not.
Comments