What do small businesses get wrong about AI risk assessment?

Why This Matters for Security Teams

Small businesses often underestimate AI risk because the technology is introduced as a productivity tool, then quietly becomes part of customer support, content generation, code assistance, or decision support. That shifts risk from a single model choice to a broader control problem across data handling, access, reliability, and compliance. The NIST AI Risk Management Framework is useful here because it frames AI risk as lifecycle governance, not a one-time approval.

NHIMG’s research on why NHI security matters now shows that security debt around machine access is common long before a breach becomes visible. For small firms, the same pattern appears when AI tools inherit API keys, file access, or customer data permissions without a separate risk review. The mistake is not just missing a control. It is assuming one AI policy can cover every use case, when the exposure changes dramatically between a drafting assistant and an automated workflow that touches records or payments. In practice, many small businesses discover AI risk only after an external tool has already been granted broad access to sensitive systems.

How It Works in Practice

A practical AI risk assessment starts by separating the use case into discrete parts: what data the system consumes, what it outputs, what actions it can take, and who is accountable for those actions. That structure aligns with the NIST Cybersecurity Framework 2.0 and makes the assessment usable for small teams that do not have a dedicated AI office.

For each use case, security teams should ask four questions:

• Does the system process personal, regulated, or confidential data?
• Can the output be acted on without human review?
• Does the tool connect to SaaS apps, internal files, or credentials?
• What happens if the model is wrong, manipulated, or unavailable?

That last question matters because AI risk is not limited to accuracy. A system can be technically “working” while still exposing secrets, amplifying biased decisions, or creating records that trigger legal or contractual problems. NHIMG’s Top 10 NHI Issues research is relevant when an AI tool is effectively acting as a machine identity with delegated access. If it can call tools, retrieve data, or initiate workflows, it needs its own review of permissions, logging, and revocation paths.

For control design, current guidance suggests using the NIST Cyber AI Profile (IR 8596) alongside basic security controls such as least privilege, change tracking, and vendor due diligence. Small businesses do not need a heavyweight model governance program to start. They need a documented inventory of AI use cases, a data classification check, and a decision on whether a human must approve any external or high-impact action. These controls tend to break down when one admin account and one chatbot token are shared across multiple departments because accountability and containment disappear.

Common Variations and Edge Cases

Tighter AI oversight often increases operational overhead, so small businesses have to balance speed against review discipline. That tradeoff is real when a team is trying to automate marketing, support, and document drafting with limited staff.

There is no universal standard for this yet, so best practice is evolving. Low-risk uses, such as internal brainstorming with no sensitive data, may only need a light review. Higher-risk uses, such as summarising customer records, generating employment decisions, or triggering transactions, need stronger controls and more explicit approval. The same tool can fall into different risk categories depending on the data and the action it can take.

One common edge case is vendor-hosted AI inside a larger SaaS product. Businesses may assume the vendor has handled risk, but the organisation still owns the decision to upload data, grant access, or rely on the output. Another edge case is open-source or locally hosted AI, where privacy may improve but security and maintenance responsibilities increase. For both cases, the Oasis Security & ESG research found that 72% of organisations have experienced or suspect they have experienced a breach of non-human identities, which is a reminder that machine access often fails before the AI model itself does. Small businesses should treat AI risk as a portfolio of use-case decisions, not a one-time checkbox.

Related resources from NHI Mgmt Group

• What do security teams get wrong about breach risk in small businesses?
• What do security teams get wrong about maintaining assessment readiness for federal frameworks?
• What do security teams get wrong about preparing entry level staff for IAM and AI security work?
• What do security teams get wrong about passwordless authentication and AI risk?