According to AI phishing threats for small businesses 2026 data from Sagiss’s 2026 report on how AI is changing cyberattacks on SOBs. Phishing is one of the oldest cyberattacks, and in the last few years AI has only made it more dangerous. Unlike the larger organizations, this is especially bad news for small and mid-sized businesses- those that can’t protect themselves without cybersecurity operations centers, dedicated incident response teams, etc.
In this blog, you can find what has changed in language models and what small businesses should do for their security.
Phishing used to be just a problem of quality. many numbers of generic emails were sent by attackers; they were easy to spot as they were full of grammatical errors and awkward phrasing. But that quality ceiling gets removed when large language models are introduced. It no longer takes hours but just seconds for a phishing email to be generated today that can match the tone, structure, and even the writing quirks of a real colleague, vendor, or executive.
Three shifts stand out the most.
Personalization now happens at scale. It used to be a choice between phishing (low quality, high volume) and spear phishing (high quality, labor-intensive, reserved for high-value targets) that attackers had to decide. That tradeoff has been collapsed by AI. A company’s LinkedIn presence, press mentions, or a breached vendor list-, all these details AI models can ppull even can generate individually tailored emails for hundreds of targets at once, with none of the manual research that spear phishing used to require.
Targeting moves faster. AI tools can research, draft, and launch a campaign within hours of a data breach or a public company announcement, whereas a human has to do days’ worth of research. targeted phishing attempt can do a layoff notice, an executive change, or a new vendor contract all this before the news cycle has moved on. That window has shrunk to days, sometimes hours, because generating a plausible pretext for a small business that announces a new CFO on LinkedIn on a Monday no longer requires a human researcher to piece the details together manually.
Voice and video have entered the picture. IBM’s 2025 Cost of a Data Breach Report found that roughly 1 in 6 breaches that year involved attackers using AI, most commonly for social media phishing and for deepfake impersonation. CrowdStrike separately reported a 442 percent increase in vishing, or voice phishing, between the first and second half of 2024. The classic “CEO fraud” scam, where an attacker impersonates an executive by email to request an urgent wire transfer, now sometimes arrives as a phone call or a video message that sounds and looks like the executive in question. For a small business where the owner or a handful of senior staff have a public speaking history, a podcast appearance, or even a few minutes of video on the company website, generating a convincing voice clone no longer requires much more than that.
The financial stakes back this up. The FBI’s Internet Crime Complaint Center recorded 21,442 business email compromise complaints in 2024, with reported losses of 2.77 billion dollars, making it the second-costliest cybercrime category the agency tracks. Business email compromise does not require malware or a network breach. It requires one convincing message and one employee with the authority to move money or share credentials. Verizon’s 2025 Data Breach Investigations Report puts a number on how often that human element is the deciding factor: roughly 60 percent of confirmed breaches in the dataset involved some combination of human error, manipulation, or misuse rather than a purely technical exploit. AI phishing is aimed squarely at that 60 percent, not at firewalls or unpatched servers.
It was a reasonable baseline three years ago to build security awareness training on phishing emails and an annual slideshow, but it is not enough now, and what the employees are being trained to notice is exactly what the gap is.
Bad grammar, mismatched sender addresses, generic greetings. People can be trained with click-test programs. Those tells are disappearing, but Unusual urgency, a request that bypasses normal approval steps, and pressure to act before checking with someone else are all underlying patterns of a social engineering attempt. Behavioral pattern recognition should be included in training rather than surface-level red flags. A quarterly simulated phishing test still has value for measuring whether training is sticking even if a request skips the normal process, regardless of how polished the message looks or sounds, and is protected against a threat that keeps changing its surface details. An employee should be able to find that out, but the content of that training has ended up shifting from “look for typos” to “notice when a request is trying to make you move faster than usual.”
An out-of-band verification step that does not depend on the same channel the requests, like wire transfers and payment changes, arrive through is needed. Verification means calling a phone number on file from a previous, trusted interaction, not one provided in the message itself. If a request to change a vendor’s bank account details arrives by email, verifying it by replying to that same email confirms nothing, since a compromised or spoofed account will simply reply. The majority of business email compromise and CEO fraud attempts can be closed off by this simple habit, and it costs nothing to implement. The businesses that get caught by these scams are rarely the ones with strong technical defenses in place; they are the ones without a written, non-negotiable rule that payment changes get a phone call before they get processed, regardless of who appears to be asking or how urgent the request sounds.
Multi-factor authentication fatigue is its own category worth addressing directly. Attackers who have already obtained a password will sometimes trigger repeated MFA push notifications, hoping an employee approves one out of frustration or habit rather than suspicion. The fix is not more training alone. It is switching to number-matching MFA, where the employee has to enter a code shown on the login screen rather than tap a single approve button, combined with a policy that any unexpected MFA prompt gets reported rather than dismissed. Businesses still relying on simple push-to-approve MFA are leaving a known, exploitable gap open.
None of these three changes require an enterprise security budget. Out-of-band verification is a policy decision. Number-matching MFA is usually a configuration change within an existing Microsoft 365 or Google Workspace subscription rather than a new purchase. Behavioral pattern training is a shift in what a training program covers, not a reason to buy a new platform. What they require is someone paying attention to phishing trends closely enough to know these are the specific gaps worth closing in 2026, which is exactly the piece that tends to fall through the cracks when security is one item on a long list of responsibilities for a single IT generalist.
Most small businesses with in-house IT have one or two generalists handling helpdesk tickets, network administration, and vendor management, often without dedicated time set aside for security monitoring. That is not a criticism of the people doing the job. It is a staffing math problem. AI-generated phishing does not follow a nine-to-five schedule, and catching it requires continuous monitoring, not a periodic review.
A managed security provider brings a few specific capabilities to bear that are hard to replicate with a generalist internal team. Continuous, around-the-clock monitoring means a suspicious login or an unusual email forwarding rule gets flagged and investigated within minutes rather than discovered days later during an unrelated support ticket. Threat intelligence feeds give the provider visibility into phishing campaigns and attacker infrastructure being used against similar businesses elsewhere, often before those tactics show up in a specific client’s inbox. Documented incident response playbooks mean that when a phishing attempt succeeds, and eventually one will, the first hour is spent executing a plan rather than improvising one: who isolates the affected account, who notifies leadership, who resets credentials and reviews forwarding rules, and who determines whether a wire transfer can still be recalled.
Sagiss, for example, rather than a subcontracted overflow desk it runs a 24/7 security operations center staffed by its own employees, paired with policy-driven MFA enforcement and email filtering tuned specifically against AI-generated phishing patterns. Phishing defense has moved from a training problem that any business can solve internally with an annual course to an operational problem that benefits from dedicated tooling and staffing most small businesses cannot justify building in-house. This is the broader shift in 2026 data points reflected by that combination. A firewall and a helpdesk can be managed by a generalist IT hire. Also, a security operations function can be run by very few people, and discovering the gap only after a phishing attempt has already succeeded usually happens in businesses that try to ask one person to do both.
The reason for alarm is not the 72 percent figure that opened this piece on its own. It is a signal that the people closest to these attacks, the employees actually receiving them, have noticed the shift before most security programs have caught up to it. The ones that turn a near-miss into a non-event instead of a wire transfer that cannot be recalled will be the Businesses that update their training, their verification habits, and their monitoring to match that shift.
Immediately change your passwords, enable MFA, contact your bank if financial data was compromised, and report the incident to your organization’s IT or cybersecurity department if the scam occurred through company channels.
Unfortunately, yes. As AI becomes more powerful and accessible, phishing scams will evolve. However, awareness, vigilance, and advanced security tools will continue to be strong countermeasures.
Absolutely. Many cybersecurity tools now use AI to detect suspicious behavior, analyze email content, and identify anomalies faster and more accurately than humans can.
