Artificial intelligence is not the reason cybercrime exists. Criminals were already exploiting weak passwords, unpatched systems, rushed payments and distracted staff long before generative AI arrived.
What AI changes is the speed, scale and credibility of those attacks.
A criminal no longer needs exceptional writing skills to produce a convincing phishing email, or a large team to tailor messages to dozens of employees, suppliers and customers. Advanced AI models can help create believable copy, translate it, imitate a company’s tone and process publicly available information at pace. Voice and video tools can also make impersonation attempts more persuasive.
For business owners, the important question is not whether AI will create entirely new forms of cybercrime. It is whether the business has controls designed for a world in which old scams are cheaper to run, harder to spot and easier to personalise.
That is a business-risk issue, not simply an IT issue.
AI makes familiar attacks more convincing
The most immediate risk is not usually a sophisticated attack on a company’s AI model. It is social engineering: persuading someone to hand over access, approve a payment or share sensitive data.
Business email compromise remains particularly costly. The FBI’s Internet Crime Complaint Center reported more than $2.77 billion in reported business email compromise losses in 2024. Its wider data, covering October 2013 to December 2023, recorded $55.5 billion in global exposed losses linked to this type of fraud. FBI IC3 data makes the point clearly: these scams affect small local businesses as well as large corporations. (ic3.gov)
AI can improve a fraudster’s chances of success by helping them to:
- write emails without the poor grammar and generic language that once gave many scams away;
- research executives, finance staff, suppliers and current projects from public sources;
- create variants of the same scam for different teams, regions or languages;
- mimic a senior colleague’s tone in email or messaging platforms;
- generate convincing scripts for phone calls, including urgent payment requests;
- produce manipulated audio or video intended to support an impersonation attempt.
The result is that “does this email look suspicious?” is no longer an adequate test. A well-written message can still be fraudulent. A familiar voice on a call may still not be the person it claims to be.
Businesses need processes that assume a request to change bank details, release a payment, reset access or disclose sensitive information could be manipulated, however plausible it appears.
The risk is also inside the business
AI can create security exposure through legitimate use as well as criminal misuse.
Employees may paste customer records, source code, commercial proposals, legal documents or internal strategy into public AI tools to save time. That can breach confidentiality obligations, create data-protection issues or expose information that should never leave the organisation’s approved systems.
The same applies to AI assistants connected to email, cloud storage, customer relationship management platforms or finance tools. These integrations can be useful, but they expand the number of systems, permissions and suppliers that need to be understood and managed.
This is particularly relevant for businesses using customer information to improve marketing. Data can make marketing more useful and efficient, but it also creates an obligation to handle that data carefully. Our guide to marketing and data for small business owners covers the importance of transparency, consent and security when using customer information.
A practical AI policy should answer straightforward questions:
- Which AI tools are approved for work?
- What information must never be entered into them?
- Who can connect AI tools to company systems?
- What access can an AI assistant have, and for how long?
- How will the business review suppliers’ security and data-handling terms?
- Who is accountable when an AI-generated output leads to a customer, financial or security decision?
Policies should not be an excuse to ban useful tools without thought. The better approach is to make safe use easier than unsanctioned use.
Security needs to focus on identity and verification
Traditional basics matter more, not less, when AI makes deception more credible.
The US Cybersecurity and Infrastructure Security Agency recommends multifactor authentication across business systems, with a preference for phishing-resistant methods. It specifically highlights email, file storage, remote access and administrator accounts as priorities. CISA’s guidance on multifactor authentication also makes an important distinction: a text-message code is better than no MFA, but it is weaker than options such as security keys or app-based authentication with number matching. (cisa.gov)
For many businesses, the most valuable change is to remove trust from high-risk requests and put verification into the workflow.
| Risky request | Better control | Owner |
|---|---|---|
| Supplier changes bank details | Call a known number already held on file | Finance team |
| Executive requests an urgent payment | Require a second approver outside email or chat | Finance lead |
| Employee receives a password-reset prompt | Verify through the official sign-in page or IT channel | All staff |
| New AI tool requests access to company data | Review permissions, supplier terms and business purpose first | IT and data owner |
This is not bureaucracy for its own sake. It is a deliberate pause at the point where an attacker wants someone to act quickly.
Businesses should also maintain tested backups, apply security updates promptly, limit administrator privileges, keep logging switched on and practise their incident-response process. CISA’s StopRansomware Guide recommends phishing-resistant MFA, patching, staff awareness training and rehearsed response procedures, among other measures. (cisa.gov)
Cyber insurance can help, but it cannot replace controls
Cyber insurance has become a more established part of business risk planning, but it should be treated as a financial recovery tool rather than proof that a company is secure.
A policy may help with costs such as incident-response specialists, legal advice, customer notification, data recovery, business interruption, cyber extortion and certain third-party claims. The exact cover, exclusions, limits, waiting periods and sub-limits vary considerably. Businesses should not assume that a standard property, liability or professional indemnity policy will respond to a cyber incident.
The insurance market is adapting to a more complicated threat environment. The National Association of Insurance Commissioners reported that US cyber-insurance claims rose by almost 40% in 2024, to nearly 50,000 reported claims. It also noted that insurers look favourably on stronger cybersecurity controls when underwriting cover, even as incidents involving ransomware, business interruption, litigation and regulatory investigations make claims more complex. NAIC’s 2025 Cybersecurity Insurance Market Report provides a useful overview of these trends. (content.naic.org)
In practice, this means insurers increasingly want a clearer picture of a company’s security posture. Expect questions about MFA, backups, endpoint protection, patching, incident response, staff training and controls around payment fraud. AI-related exposure is likely to become part of that conversation too, especially where a business relies on AI tools, uses sensitive data in them or faces risks from impersonation and automated fraud.
Before buying or renewing cover, ask:
- Does the policy cover business email compromise and fraudulent funds transfers, or are they subject to a separate limit?
- Is business interruption covered if a key cloud or technology supplier suffers an outage or breach?
- Does the insurer provide an incident-response panel, and must it be contacted before appointing external advisers?
- What security controls are required by the policy, and could a gap affect a claim?
- Are AI-related events addressed explicitly, or left to general wording and exclusions?
A broker, lawyer or specialist adviser can help interpret the wording. The cost of that review is modest compared with discovering a gap during an incident.
AI should improve defence as well as attack
There is a more positive side to this shift. AI can help security teams sort through alerts, identify unusual account activity, prioritise vulnerabilities and speed up routine analysis. Smaller businesses may benefit through managed security providers and cloud platforms that use automation behind the scenes.
But AI-assisted defence needs oversight. It can produce false positives, miss context and expose sensitive information if used carelessly. NIST’s developing Cyber AI Profile frames the challenge around three connected areas: securing AI systems, using AI to support cyber defence and resisting AI-enabled attacks. NIST’s Cyber AI Profile is a useful reminder that AI risk management cannot sit separately from the wider security programme. (nccoe.nist.gov)
The sensible response is not panic, and it is not blind faith in AI security tools. It is stronger identity controls, clearer approval processes, careful handling of data, tested recovery plans and insurance that reflects the risks the business actually carries.
Cybercrime is becoming more efficient. Business security now needs to become more deliberate.
