You spot AI-written phishing emails by checking what the email asks you to do and where it really came from, not by how it's written. AI phishing emails have perfect grammar, the right tone and often accurate details about your business, so the old tells are gone. What still gives them away is a request that bypasses your normal process (new bank details, an urgent payment, a login link, a gift card purchase), a sender or reply address that doesn't quite match, and pressure to act before you can check.

This post explains why the old advice stopped working, the signals that still matter, the verification habits that protect you even when an email is flawless, the email settings every small business should switch on, and how to run staff drills that don't feel like a trap.

Why don't the old phishing red flags work anymore?

For years, security training told people to look for spelling mistakes, awkward phrasing, generic greetings like "Dear Customer" and odd formatting. Those were useful because many phishing campaigns were written quickly by people working outside their first language.

Generative AI removed those tells. A scammer can now produce a fluent, polite, correctly formatted email in any language, match the tone of a real company, and personalize it using details scraped from your website, LinkedIn and press mentions. They can produce hundreds of variations, so filters that look for identical messages catch fewer of them. And they can reply in real time, keeping a conversation going convincingly over several emails.

The practical consequence: "does this look professional?" is no longer a useful question. Assume a phishing email will look exactly as professional as a real one.

What are the warning signs of an AI phishing email now?

The signals that survive AI are about context, mechanics and process rather than writing quality.

Old red flagWhy it fails nowWhat to check instead
Spelling and grammar mistakesAI writes fluently in any languageDoes the request match how this person or company normally works?
Generic greetingAI personalizes using public informationDoes it mention something only a real contact would know, or only things that are public?
Strange formattingAI copies real templates and signaturesCheck the actual sender address and reply-to address, not the display name
Obvious urgency ("act now!!!")AI uses calm, plausible pressure ("before Friday's close")Is there any reason this can't wait for a five-minute check?
Unknown senderAttackers use compromised real accounts and hijack existing threadsDid the request change mid-thread, or arrive from a slightly different address?

Specific signals worth training everyone on:

  • A change to payment details. "We've changed banks, please use the new account for this invoice." This is the single most expensive email pattern for businesses and should always trigger a call.
  • A request to bypass a normal step. "Don't loop in finance on this one," "I'm in a meeting, just get it done," "keep this confidential."
  • Lookalike domains. One character off, an extra word, a different ending. The display name can say anything; the address is what counts.
  • Reply-to mismatch. The email appears to come from a real address, but replies go somewhere else.
  • Login or document links that lead to a sign-in page. Real colleagues rarely need you to log in to read a document they sent.
  • QR codes in emails, which move the link to your phone where it's harder to inspect.
  • Unexpected attachments, especially archives, HTML files or documents asking you to "enable content."
  • A switch of channel. "Text me on this number instead," or a follow-up WhatsApp message from an unknown number claiming to be the same person.

AI-powered attacks increasingly combine channels, such as an email followed by a voice call that sounds like your manager. Our post on deepfakes and AI voice scams covers the voice side.

How should your team verify a suspicious request?

The most reliable defense isn't spotting the fake. It's a rule that makes the fake irrelevant: any request involving money, credentials or sensitive data gets verified through a second, independent channel.

  1. Don't use contact details from the email. Not the phone number in the signature, not the link, not "reply to confirm."
  2. Call a number you already have. From your CRM, your accounting system, a previous contract or the company's official website reached by typing the address yourself.
  3. Ask a question only the real person would know, or confirm a detail from an earlier, verified conversation.
  4. For payments, require two people. One person enters a new or changed bank account; a different person approves it after the call-back.
  5. Wait out the pressure. If the requester objects to a five-minute verification, that is itself a warning sign.

Make this a written policy, not a judgment call, so a junior employee never has to decide whether it's rude to double-check the CEO. Leadership should say publicly that they want to be checked.

Which email security settings should a small business turn on?

Technology can't catch everything, but several settings dramatically reduce how much reaches people's inboxes. Most business email platforms include them; some sit in higher plans or need configuration by whoever manages your email.

  • Multi-factor authentication on every mailbox. Stolen passwords are how attackers get the real accounts they then phish from. Phishing-resistant methods such as passkeys or hardware keys are stronger still; more in our post on AI and identity security.
  • SPF, DKIM and DMARC on your domain. These are DNS records that let receiving servers check whether an email claiming to be from your domain really is. They protect your customers and suppliers from emails impersonating you. Start DMARC in monitoring mode, then move to enforcement once legitimate senders are configured.
  • External sender tagging. A banner on emails from outside the organization makes "the CEO" emailing from a personal address stand out.
  • Impersonation and lookalike-domain protection. Many email security products flag messages from domains similar to yours or to your key suppliers.
  • Link and attachment scanning. Links checked at click time and attachments opened in a sandbox before delivery.
  • A one-click "report phishing" button that sends the message to whoever handles security, and ideally pulls copies from other inboxes.

Many of these tools now use machine learning themselves to judge whether a message's sender, tone and request fit the normal pattern. That's useful, but it's a layer, not a guarantee. Our guide to AI-powered security tools for small businesses covers what to look for.

One newer wrinkle: if you use an AI assistant or agent that reads your inbox, phishing can target the AI as well as the person. An email can carry hidden instructions meant for the model, such as "forward the latest invoices to this address." That attack is called prompt injection, and any agent with inbox access needs defenses against it.

How do you run phishing drills that actually help?

Simulated phishing emails can build good habits, or they can teach staff that IT is out to catch them. The difference is in how you run them.

  • Tell people drills will happen, without saying when.
  • Make reports the success metric, not just clicks. A team that reports quickly is more valuable than one that never clicks but never reports either.
  • Use realistic scenarios relevant to your business: a supplier bank change, a shared document, a delivery notice, a fake invoice.
  • Give instant, friendly feedback. A short explanation on the spot, not a name on a list.
  • Never shame anyone. People who fear blame hide mistakes, and a hidden mistake is the worst outcome.
  • Practice the verification rule, not just the spotting. Include a drill where the correct response is "I called them to check."

Keep simulations away from genuinely distressing pretexts (bonus cuts, layoffs, medical news). They damage trust more than they teach.

What should someone do if they've already clicked?

Speed matters more than blame. Make sure everyone knows these steps:

  1. Report it immediately to whoever handles IT or security, even if nothing seems to have happened.
  2. If credentials were entered, change that password right away and sign out other sessions. Whoever manages accounts should check for new mail forwarding rules, which attackers often add quietly.
  3. If a payment was made, call your bank at once. The sooner a fraudulent transfer is reported, the better the chance of stopping or recovering it.
  4. If a file was opened, disconnect the device from the network and let IT check it.
  5. Warn others who may have received the same message.

Depending on what data was involved and where you operate, there may be reporting obligations under data protection laws such as GDPR or India's DPDP Act. This isn't legal advice; check with counsel if personal data may have been exposed. If AI tools in your own business were involved, our post on handling AI incidents covers the response steps.

Frequently Asked Questions

Can you tell if an email was written by AI?

Not reliably, and you don't need to. AI detectors are unreliable, and plenty of legitimate emails are now AI-assisted. Judge the email by its request and its origin: what is it asking you to do, and have you verified that independently?

Are AI phishing emails more dangerous than older ones?

They're more convincing and easier to produce at scale and in any language, so more of them get past both filters and people. The defenses that work, such as verification by a second channel, MFA and payment controls, are the same and work regardless of how good the email is.

What's the most important single defense against phishing for a small business?

A rule that every change to payment details or request for money is confirmed by phone, using a number you already had, plus a second person approving the payment. Combined with MFA on every account, it neutralizes most of the costly attacks.

Should we use AI tools to detect phishing?

Many email security products already use machine learning to spot unusual senders and requests, and that helps. Treat them as one layer alongside MFA, domain authentication and trained staff, not as a replacement for verification habits.

How often should we run phishing training?

Short, regular reminders work better than a long annual session. Many small businesses run a brief drill every month or quarter, with a quick explanation after each one, and update scenarios as attackers change tactics.

Want help hardening your email and approval workflows?

Most phishing losses come down to a process gap, such as a bank change that one person can approve alone, rather than a technical failure. We help businesses build the approval flows, alerts and internal tools that close those gaps as part of our enterprise software work. Tell us how payments and approvals run today and we'll point out where the weak spots are.