KINSMC
In brief: Before you put work, customer or personal information into an AI tool, identify the account type, read the current data controls, minimise what you share and test the workflow with non-sensitive material first.
AI tools can save time with research, drafting, transcription and customer support. They can also receive information that should not leave your control. The useful question is not simply “is this AI tool safe?” A better question is: what information will this specific account, setting and workflow send to the provider or to other connected services?
This guide gives you a practical way to answer that question before you paste in a document, upload a spreadsheet or connect a business app. It is written for everyday users and small teams, not just IT departments.
Start with the data, not the tool
Make a short inventory of what you plan to use. A public product description is different from a customer export, a signed contract, a medical record or unreleased financial results. “Sensitive” also includes combinations of ordinary details that could identify a person, such as a full name plus booking date, phone number or address.
Separate your material into three groups:
- Public: already published information, such as a public web page or press release.
- Internal: work notes, procedures, pricing plans and drafts that are not public.
- Restricted: personal data, payment details, credentials, customer files, legal documents and anything covered by a confidentiality agreement.
For restricted material, pause. Ask whether the task can be completed with an anonymised extract or a short summary instead. Removing names, account numbers, precise locations and unnecessary attachments is often the fastest risk reduction.
Check the exact account and plan
Privacy terms can differ between a consumer account, a business workspace and an API. Do not assume that a company-wide policy applies to the account you happen to be using. OpenAI, for example, explains that personal ChatGPT workspaces have data controls, while Business, Enterprise, Edu and API customer content is not used to train its models by default. That does not remove your responsibility to check access, retention and sharing settings for your own organisation.
Look for answers to these questions in the provider’s current documentation:
- Can you opt out of using content to improve the service?
- Who in your organisation can view chats, files and connected data?
- How long is content retained, and can it be deleted?
- What happens when you give feedback, share a conversation or connect a third-party app?
- Does the tool clearly distinguish a personal account from a managed workspace?
A privacy label on a landing page is not enough. Use the support page or policy that names the feature you plan to use. Settings and product terms change, so record the date you checked them.
Use the minimum necessary input
Good data practice is also good prompt practice. Give the model only the facts it needs to complete the task. If you want help rewriting a customer email, replace the recipient’s name and reservation number with placeholders. If you need a summary of a report, remove staff contact details and appendices that are irrelevant to the summary.
Do not put passwords, one-time codes, card numbers, passport scans or private API keys in a prompt. An AI chat is not a password manager. For customer lists, use a small synthetic sample to test your workflow before you process any real record.
Review connections and sharing routes
The prompt box is not the only route for data. File uploads, browser extensions, cloud-drive connections and automation tools can have their own permissions. Before enabling one, check which folders, mailboxes or services it can access. Prefer the narrowest permission available. “Read selected folder” is safer than broad access to an entire drive when your task only needs one project folder.
Also check the output route. A generated summary may be copied into a public ticket, an email thread or a shared document. Review the output for private details before you send it. This is particularly important when a tool combines information from several sources.
Keep a simple approval rule
Small teams do not need a 40-page policy to make a difference. A short rule can work: public information may be used in approved tools; internal information needs a named workspace and a human review; restricted information needs written approval or must stay out of the tool. Assign one person to keep a list of approved services and their current settings.
NIST’s Generative AI Profile recommends managing generative-AI risks through governance, documentation, testing and human oversight. In practice, that means treating an AI workflow like any other business process: decide the purpose, identify what can go wrong, set a limit and review the result.
Test with a low-risk example
Before you build a process around a new assistant, run a short test using public or fictional data. Check the quality of its answer, whether it cites or exposes information you did not intend to include, and how easily you can remove the test material. This is also the right moment to test who can access the conversation in a shared workspace.
When the tool makes factual claims, verify them independently. Our guide to checking sources in an AI search answer explains a simple method: open the source, confirm the claim, then look for the date and the original context. For a wider selection process, see our checklist for choosing an AI tool.
A practical final check
Before you send data, stop for ten seconds and ask: Would I be comfortable if this exact text appeared in a shared work chat? Is there a smaller or anonymised version? Have I checked this account’s current controls? Who else can see the input and output? If any answer is unclear, use a safer sample or get approval first.
AI can be genuinely useful without becoming a shortcut around basic data care. The safest habit is simple: choose the right account, share less than you think you need, and verify both the tool’s current settings and its output.