AI-enabled phishing: why traditional training is no longer enough and what to do in an SMB

A recent KnowBe4 report indicates that 86% of phishing analyzed in the last quarter includes content generated or personalized with AI. Vishing and deepfakes, once anecdotal, are now part of the regular toolkit.

The classic awareness program, based on misspellings and suspicious links, increasingly fails to measure real risk. For an SMB, keeping that practice without updating it produces a false sense of security.

Short answer

Effective anti-phishing training in 2026 combines three changes: realistic simulations that match current attack quality, operational metrics beyond click rate, and verification processes that reduce damage when someone falls for it, because they will.

What changed in the last year

Before	Now with AI
Misspellings and reused templates	Idiomatic, recipient-personalized text
Generic addresses and clumsy domains	Coherent impersonation with real environment accounts
Vishing reserved for VIP accounts	Vishing aimed at help desk and admin staff
Rare and expensive deepfakes	Affordable short audio and video
Asynchronous email attacks	Multichannel conversations with follow-up

The common factor: the cost of crafting a credible lure dropped, so volume and quality went up.

Attack types worth covering

Vector	Typical case	Where it hits
AI-generated email phishing	Email from “the CEO” in correct tone and context	Finance, hiring, HR
Vishing to the help desk	Call to reset MFA or change phone	Identity and SaaS
Voice deepfake	Executive audio requesting urgent transfer	Payments and accounting
Vendor impersonation	Email changing IBAN after a legitimate prior thread	Procurement and operations
MFA fatigue + social engineering	Notification spam combined with a call	Any employee

An SMB does not need to cover every vector with the same intensity. It must cover those affecting processes with financial or identity impact.

Why traditional training fails

Common practice	Why it loses effectiveness
Generic simulation templates	Real attacks are now personalized
”Click rate” as the only metric	Does not measure severity or attack progression
Annual training without reinforcement	Habituation bias reduces attention
Punishing the individual who fell	Creates silence in the next case
”Do not open anything suspicious” message	Unworkable in roles that open things all day

What gets measured gets managed. If the metric is only clicks, the program optimizes that number and ignores the rest.

Modern training program

Component	Goal	How an SMB does it
Realistic simulations	Match current attack quality	Templates adapted to sector and role
Channel variation	Cover email, voice, and messaging	One quarterly simulation per channel
Microlearning	Reinforce concepts without saturation	3-5 minute pills at the moment of failure
Blame-free culture	Increase report speed	Recognition for reporting, not for guessing right
Verification processes	Reduce damage even when there is a click	Dual channel for payments, IBAN changes, resets
Post-incident review	Learn from each real case	Brief analysis shared with the team

The combination matters more than the tool. A realistic simulation without a verification process behind it only frightens.

Operational metrics

Metric	What it indicates	How to measure
Report rate	How fast someone alerts	Time from sending to first report
Failure depth	How far the attack reaches	Click only? Credentials? MFA approved?
Time to containment	Process effectiveness	Minutes until the account is locked
Per-person recurrence	Individual reinforcement need	Consecutive cycles with a miss
Process coverage	Critical areas with verification	% with documented dual channel

These metrics do not replace click rate, but they put it in context.

Processes that reduce damage

1. Mandatory dual channel for payments above an agreed threshold.
2. IBAN changes never by email; always with direct verbal verification.
3. MFA reset with out-of-band verification and manager approval.
4. Verbal confirmation of off-hours financial instructions.
5. Explicit policy: when audio or video from an executive demands urgency, verify before acting.
6. Clear channel to report suspicion without penalty.

These processes are cheap. The hard part is keeping them in daily life.

Common mistakes

1. Replacing the entire program with an expensive platform without changing processes.
2. Treating vishing as “out of scope” because it is not email.
3. Punishing the first employee who reports and silencing the rest.
4. Running seasonal simulations and forgetting them for months.
5. Excluding help desk and admin staff from the program.
6. Confusing training coverage with real compliance.

Progress indicators

Indicator	Good	Bad
Simulation realism	Sector- and role-adapted	Generic templates
Channel coverage	Email, voice, and messaging	Email only
Report time	Minutes	Days or never
Payment processes	Dual channel in >90% of cases	Email is enough to authorize
Error culture	Reports without penalty	Silence after a miss

Final criterion

Defense against AI-enabled phishing stops being only a training problem and becomes a process problem. A person who falls can still cause little damage if the processes behind them require verification. That is the lever that moves risk most in an SMB.

Working sources

• KnowBe4 research on the share of AI-enabled phishing in the last quarter.
• Awareness best practices with operational metrics beyond click rate.
• Official documentation of major identity providers for safe help desk processes.
• Technical and training decisions must be adapted to each company’s sector, size, and maturity.