Start with the workflow
- List the AI tools people actually use and the business data they submit or connect.
- Set permitted and prohibited data types, human review and account ownership for each use case.
- Test vendor retention, sharing, export and administrative controls before connecting a data source.
- Route suspicious AI-generated requests through existing payment and incident procedures.
Last updated: September 24, 2026
A blanket “AI is safe” or “AI is dangerous” rule is not useful to a 20-person business. Risk depends on the task, data, provider settings and whether a person verifies the output before it changes a business decision. NIST's AI Risk Management Framework and its Generative AI Profile provide voluntary risk-management structure, not a certification for a tool or business.
Inventory real use before writing policy
Ask each team which services they use for drafting, summarizing, coding, support or analysis. Include browser tools, app integrations, meeting bots and features embedded in products you already pay for. For each use case, record the owner, account type, data entered, connected systems, output destination and who approves a consequential result. Do not assume a tool is unused because the IT team did not purchase it.
| Use case | Main risk | Control to test |
|---|---|---|
| Drafting customer communications | Confidential input or inaccurate output | Approved data categories and human approval |
| Summarizing meetings | Unexpected recording, retention or broad sharing | Consent, storage, deletion and participant access |
| Connecting a help desk or drive | Excessive retrieval permissions | Scoped service account and access test |
| Writing code or scripts | Secrets in prompts; unsafe generated changes | Secret scanning and human review before deployment |
| Finance or HR decision support | Unverified facts or sensitive personal data | Restrict inputs and prohibit autonomous decisions |
Write a short, enforceable use policy
Name approved tools and business accounts. Specify which customer, employee, financial or credential data may never be submitted without review. State whether outputs must be checked for facts, security and intellectual-property concerns. Require an owner to approve integrations and maintain an exit path if a provider changes terms. Train staff on a few real examples rather than asking them to memorize a long policy.
Before purchasing an enterprise AI tier, check the exact contract and administrative controls. Look for retention and training settings, data location, audit logs, SSO, role management, support and export. A vendor's broad “private” claim does not establish that your chosen plan and configuration meet the need. The privacy-first guide helps map the data journey.
Handle AI-assisted fraud with process controls
Generative tools can make fake voices, images and written requests more convincing. The finance team should confirm bank changes with a known contact and a second approver, regardless of how authentic an email or call sounds. The deepfake defense guide and BEC verification guide give the decision procedure. If an AI integration leaks data or changes access unexpectedly, activate the incident-response plan and preserve the relevant logs.
A small pilot with evidence
Pick one low-sensitivity workflow. Limit access to named users, test a prohibited-data scenario, inspect what logs and deletion controls exist, and review sample outputs with the process owner. Record time saved only if measured against the prior workflow; do not infer productivity or risk reduction from vendor demos. Expand after the owner can monitor and reverse the integration.
First three actions
List actual AI services, stop unreviewed sensitive-data uploads, and require known-channel confirmation for payment or access changes. Those actions need an owner before they need a new product.
Sources checked September 24, 2026: NIST AI RMF, NIST Generative AI Profile, and NIST Privacy Framework. This page is source-based operational guidance, not a tested AI deployment.