TL;DR:

  • AI-generated phishing emails are now grammatically perfect, tonally appropriate, and personalised using publicly available information from LinkedIn, Companies House, and social media — the “spot the typos” advice no longer works
  • The tells have shifted: focus on the sender domain, the nature of the request, urgency framing, and payment/credential asks rather than writing quality
  • Update your staff training to reflect 2026 phishing tactics — awareness programmes built before 2024 will be teaching people to look for the wrong signals

Phishing emails have been a recognisable threat for two decades partly because they were recognisable. Bad grammar, obvious translations, implausible scenarios, sender addresses that didn’t match the claimed organisation. These signals trained people to be sceptical, and they worked to filter out low-effort attacks.

The calculus changed when large language models became widely accessible. The criminal ecosystem adapted quickly — LLMs are now integrated into phishing toolkits available on underground markets, and the barrier to producing a polished, convincing phishing email has effectively collapsed. The consequence for UK small businesses is that a significant category of protective instinct — reading emails for quality signals — is no longer reliable.

What’s Different About AI-Generated Phishing

The most obvious change is writing quality. A modern AI-generated phishing email targeting a UK business owner will be written in fluent British English, with correct VAT terminology, plausible company names, and accurate reference to UK-specific processes (HMRC submissions, Companies House filings, Bacs payments, CHAPS transfers). The era when “Dear Valued Customer” and three grammatical errors in the first sentence were reliable warning signs is over.

The more concerning development is personalisation at scale. Before LLMs, spear phishing (personalised, targeted phishing) required manual research per target — time-consuming and expensive enough to limit it to high-value targets. Now attackers can scrape publicly available information (LinkedIn profiles, Companies House director details, Twitter/X posts, website “about us” pages, press releases) and use it to generate personalised phishing emails in bulk.

A small business owner might receive an email that:

  • Addresses them by their first name
  • References their company by correct legal name
  • Mentions a specific supplier or partner by name (scraped from the website or LinkedIn)
  • Uses accurate UK business language for the sector they’re in
  • Has a plausible request that fits normal business operations

This isn’t a theoretical risk. UK fraud reporting to Action Fraud has shown a shift in reported phishing characteristics: victims increasingly describe emails that “seemed completely legitimate” and “knew things about our business” — patterns consistent with AI-enhanced targeting.

What to Look For Instead

Since writing quality is no longer a reliable signal, shift your scepticism to these indicators:

Sender domain scrutiny: The actual sending domain (visible in email headers or by hovering over the sender name) remains the most reliable signal. An email from “HMRC” sent from hmrc-compliance-2026.co.uk is phishing regardless of how well it’s written. Check: does the domain exactly match the organisation it claims to represent? Not a lookalike — exactly match. HMRC sends from @hmrc.gov.uk. Companies House sends from @companieshouse.gov.uk.

Unsolicited payment or credential requests: Legitimate suppliers don’t change bank details by email without a phone confirmation. HMRC doesn’t demand immediate payment by email. Your bank doesn’t ask you to re-enter credentials. If an email is requesting you transfer money, provide banking information, or enter credentials via a link, treat it as phishing regardless of quality.

Urgency and threat framing: “Your account will be closed in 24 hours” and “immediate action required or legal proceedings will commence” are manipulation tactics. Urgency is designed to prevent the pause-and-verify response that would expose the fraud. Real organisations allow reasonable time for responses.

Unexpected attachment types: Even polished phishing emails often rely on attachments with embedded malicious content — particularly HTML files that simulate browser-based credential harvesting. Be sceptical of unexpected attachments from any sender, even known contacts (whose accounts may be compromised).

Technical Defences

DMARC, DKIM, and SPF email authentication are your first layer. These don’t directly protect against AI-generated content but they do prevent attackers from spoofing your own domain to attack your suppliers, and they help filter emails that spoof legitimate organisations. If you haven’t configured these for your domain, do it — a guide is in the related articles section.

AI-enhanced email security platforms (Microsoft Defender for Office 365, Abnormal Security, Proofpoint Essentials) are now standard recommendations for SMBs with Microsoft 365 or Google Workspace. These platforms analyse email behaviour patterns rather than just content — they look at whether a sending address has been seen before, whether the request pattern is unusual for that sender relationship, and whether the email is soliciting an action inconsistent with normal correspondence. They catch AI-generated phishing better than rule-based filters because they’re also using AI.

For Microsoft 365 users: Microsoft Defender’s Safe Links and Safe Attachments features rewrite URLs and scan attachments in a detonation sandbox before delivery. These are available in Microsoft 365 Business Premium and provide meaningful protection against the attachment and link vectors common in phishing.

Updating Staff Awareness Training

If your last staff training round was built before 2024, the content is likely teaching people to look for signals that are no longer reliable. A 2026 update should:

  • Explicitly tell staff that AI-generated emails are now indistinguishable from legitimate email by writing quality alone
  • Shift the verification instinct to domains and phone confirmation: if in doubt, call the supplier on a number you already have (not one in the email)
  • Cover the specific UK scenarios: HMRC payment demands, supplier invoice fraud, Companies House compliance emails, and banking messages — these are the most commonly impersonated organisations in current UK phishing campaigns
  • Run simulated phishing exercises with AI-quality templates — if your simulations are still using obviously poor-quality templates, staff aren’t being prepared for what they’ll actually encounter

The underlying vulnerability hasn’t changed: phishing works because people are busy, trust email, and are conditioned to respond to authority and urgency. What’s changed is that the signals we trained people to spot are now unreliable. Adjust training accordingly.