AI-Generated Phishing: Why Your Staff Can No Longer Spot the “Dodgy Email”

Why AI Has Changed Phishing for UK SMEs — and Why Size Is No Protection

For years, phishing defence rested on a simple idea: train people to recognise the tells. Poor spelling. Strange grammar. “Dear Customer.” A sender address that didn’t quite match. It worked, up to a point, because producing a convincing, personalised scam email took time and skill that most criminals didn’t have.

Generative AI has removed that constraint. A fluent, correctly formatted, individually personalised email now takes seconds to produce — in perfect UK business English, referencing your company, your role, and your suppliers, all drawn from public sources such as LinkedIn and your own website. The same tools clone voices from short audio samples and, increasingly, generate convincing video.

The scale of the problem is documented, not speculative. The UK Government’s Cyber Security Breaches Survey 2025 found that 43% of UK businesses identified a cyber breach or attack in the previous 12 months — around 612,000 businesses — and that phishing was by far the most prevalent attack type, experienced by 85% of those affected, with respondents specifically citing the growing sophistication of AI-assisted impersonation. The NCSC’s own assessment of AI’s impact on the cyber threat concluded that AI will almost certainly increase the volume and impact of cyber attacks, and that generative AI already lowers the barrier for less-skilled criminals to produce convincing phishing — precisely because it removes the language errors that defenders relied on.

The assumption that a small business isn’t worth targeting no longer holds either. The same government survey found phishing hitting businesses of every size — micro and small firms included — because personalised attacks used to be reserved for large organisations only while crafting them was expensive. AI has collapsed that cost to nearly zero, so a 20-person financial services or property firm can now receive the same quality of targeted attack that was once aimed only at enterprises. Functioning business email accounts, client records and payment processes carry real value regardless of headcount.

None of this means the situation is hopeless. It means the defence has to change shape — from detection by individuals to layers of technical controls, verification processes and monitoring that don’t depend on anyone spotting anything.

This article walks through what has changed and what a proportionate response looks like for a growing UK business without an internal IT team:

Awareness training still has a role — but detection-based training, built around spotting typos and odd formatting, no longer reflects the threat. The National Cyber Security Centre (NCSC) is explicit that no training package can teach users to spot every phishing attempt.

If your team completed phishing training in the last few years, there’s a good chance it taught them to look for signals that AI-generated emails simply don’t produce. That training isn’t wrong so much as out of date — it prepared people for the previous generation of attacks.

There’s a second problem. Training that leans on blame — catching people out with simulations and penalising clicks — erodes trust and discourages reporting. The NCSC’s own guidance warns that punishing people for clicking makes employees less likely to report mistakes, which removes your most valuable early-warning system. People click for situational reasons: they’re busy, the email is relevant to their role, and responding to messages is literally their job.

Modern awareness training is process-based rather than detection-based: it teaches what to do when a request involves money, credentials or urgency — verify through a second channel — and makes reporting suspicious messages fast, expected and blame-free.

Not reliably by reading them. AI-generated phishing emails are fluent, correctly branded and often personalised. The dependable signals are now contextual — what the message asks for, not how it’s written.

This is the reframe that matters. Searching “how to spot phishing emails” still returns checklists of spelling errors and generic greetings — guidance written for a threat that has largely moved on. Judging an individual email on its craftsmanship is now close to a coin toss, even for careful readers.

What remains trustworthy is context:

• The request itself. Payments, bank-detail changes, credential prompts and gift-card purchases are high-risk regardless of how legitimate the message looks.
• Urgency and secrecy. “Do this now” and “keep this between us” are pressure tactics, not normal business behaviour.
• Channel switching. An email that pushes you to act by phone, or a call that references an email, is engineered to bypass your normal checks.
• Deviation from routine. A supplier changing bank details mid-contract, or a director requesting a transfer outside normal process, warrants verification every time.

The point isn’t that staff should stop paying attention — it’s that the organisation should stop depending on their attention as the primary control.

Filtering is essential — and imperfect. Some phishing will always reach inboxes, which is why the NCSC recommends layered defences rather than reliance on any single control.

The NCSC publishes a real-world case study from a UK financial sector company that illustrates the maths. Of 1,800 malicious emails sent to the organisation, email filtering stopped 1,750 — an excellent result. That still left 50 in user inboxes, 14 of which were clicked. Only the layers behind the filter — patched devices, reporting, monitoring and rapid response — kept the incident to a single cleaned machine instead of a business-wide compromise.

AI-generated phishing makes filter evasion easier, not harder. Because each message can be uniquely written, there is no repeated template for filters to fingerprint. Well-configured filtering and email authentication still remove the overwhelming majority of hostile mail — but the plan has to assume that some gets through.

A layered defence gives you multiple opportunities to stop an attack: before it arrives, when it’s read, if it’s clicked, and after. Missing any layer isn’t fatal — missing all but one is.

Four controls do most of the work for an SME: email authentication (SPF, DKIM, DMARC), multi-factor authentication, conditional access, and DNS/web filtering. All are available within a typical Microsoft 365 environment or alongside it.

Email authentication: SPF, DKIM and DMARC
These three DNS records work together to prove that email claiming to come from your domain genuinely does. SPF lists the servers allowed to send for you, DKIM cryptographically signs your messages, and DMARC tells receiving servers what to do when a message fails those checks. Configured correctly, they make it far harder for criminals to impersonate your business to your clients and suppliers — and improve deliverability of your genuine email as a side effect.

Multi-factor authentication (MFA)
If a phishing email does capture a password, MFA is the control that stops it becoming a breach. App-based MFA — Microsoft Authenticator, for example — adds seconds to a login; recovering from a compromised account takes days. Prefer app-based approval over SMS codes, which are vulnerable to SIM-swap attacks.

Conditional access
Conditional access policies add context to every sign-in: block logins from unexpected countries, require compliant devices, and challenge unusual patterns. A stolen credential used from an unfamiliar location simply doesn’t work.

DNS and web filtering
Most phishing emails need the recipient to reach a malicious site. DNS filtering blocks the connection at the point of lookup — so even a convincing email that gets clicked leads nowhere. It protects staff wherever they work, in the office or remotely.

With a mandatory second-channel rule: any request to move money, change bank details or share credentials is confirmed through a different channel — a known phone number, in person, or your finance system — before it is actioned. The rule applies to everyone, including directors.

This single procedure defeats the most damaging AI-enabled attacks — CEO impersonation, invoice fraud and voice cloning — because it doesn’t depend on judging whether the request is genuine. However convincing the email, the voice or the video, the fraud fails at the verification step.

The common objection is that verification will slow the team down and irritate clients. In practice, a call-back to a known number takes under a minute. Recovering from a fraudulent transfer takes days at best, and a compromise involving personal data can trigger an obligation to notify the Information Commissioner’s Office (ICO) within 72 hours under UK GDPR. Clients and suppliers in compliance-sensitive sectors increasingly read verification procedures as a mark of professionalism, not friction.

Make it work in practice:

• Verify using contact details you already hold — never those supplied in the request itself.
• Make the rule absolute. Urgency, seniority and confidentiality are precisely the pressures attackers apply.
• Pair it with a no-blame reporting route: a suspicious message reported in one click is intelligence; the same message deleted in embarrassment is a missed warning.

Less like the clumsy scams of a decade ago, and more like ordinary business correspondence with one altered detail. These patterns are appearing across exactly the sectors growing UK SMEs operate in.

CEO/CFO impersonation and payment fraud. A finance administrator receives an email from the managing director — correct tone, correct signature, plausibly timed — asking for a same-day payment to a new supplier. This is business email compromise (BEC), and AI has made the impersonation fluent enough that tone is no longer a tell.

Invoice and supplier fraud. A genuine-looking invoice arrives from a real supplier, matching your actual billing cycle, with one change: the bank details. Sometimes the attacker has compromised the supplier’s mailbox, so the message truly does come from their domain — another reason verification beats inspection.

Voice cloning. A director’s voice can be cloned from a short public sample — a webinar, a podcast, a company video. The finance team then receives a call in that voice authorising an urgent transfer. A call-back to the director’s known number collapses the fraud instantly.

Personalised spear-phishing at scale. AI tooling assembles a target list from LinkedIn and your website, then writes an individual email to each person referencing their role, colleagues and current projects. What once took an attacker a day per target now takes seconds per hundred.

Deepfake video calls. Rarer, but documented: synthetic video used to impersonate senior staff in remote meetings authorising payments. It merits awareness rather than alarm — and the same second-channel verification defeats it.

Layers reduce risk; they don’t eliminate it. Proactive monitoring catches the attack that gets through early, and a rehearsed response plan turns a potential crisis into a contained incident.

Monitoring means someone — or something — is watching for the signals a successful phish produces: sign-ins from unexpected locations, new mailbox forwarding rules, unusual permission changes, anomalous email volume. In a compromised-account scenario, these appear within minutes. Caught then, the incident is an inconvenience; caught weeks later, it may be a regulatory notification and a client-trust problem.

Incident readiness is the unglamorous half: knowing in advance who isolates the account, who informs affected clients, how backups are restored, and when the ICO threshold is met. Businesses with a documented, tested response recover in a fraction of the time — the NCSC’s layered-defence case study above ended with one device cleaned in hours precisely because detection and response were already in place.

Finally, this is not a one-off project. The threat landscape that produced AI phishing will keep evolving, which is why Astec IT builds continuous improvement into the service model: continuous monitoring and email threat detection as standard, with strategic quarterly reviews that reassess your security posture as your business grows and the threats change.

The AI-Era Phishing Resilience Checklist

Why Growing UK Businesses Work With a Dedicated IT Partner on AI-Era Security

Very little in this article requires enterprise budgets. It requires something growing businesses rarely have internally: a structured, proactive IT function that configures the controls, watches the monitoring, runs the reviews and keeps the whole posture current as threats evolve.

Astec IT works with UK SMEs in financial services, insurance, real estate, hospitality and professional services to put exactly that in place. Every client is assigned a dedicated primary and secondary engineer — named individuals who understand your infrastructure, your industry and your risk profile, not a shared helpdesk pool. The methodology follows the same sequence this article recommends: a structured risk assessment; an access and vulnerability review covering MFA, conditional access and email authentication; security awareness improvement updated for AI-era phishing; proactive monitoring and email threat detection; incident prevention and response readiness; and continuous improvement through strategic quarterly reviews.

If your current IT provider is reactive — resolving tickets when problems arise, but not reviewing access controls, watching for compromise signals or reassessing your defences each quarter — then email security is being handled in the narrow sense only. The gap between reactive support and proactive security is precisely where AI-enabled phishing operates.

Delivered in practice: for a UK business requiring unified cybersecurity across multiple sites, Astec IT consolidated layered security controls, centralised monitoring and a formal incident response process into a single engagement — with quarterly reviews maintaining the standard since. Read the Unified Cyber Security Solutions case study.

Criminals are not the only source of AI risk. The AI tools your own staff use create a separate exposure your IT partner may not be able to see.

Frequently Asked Questions About AI Phishing