Practical guide
7 Prompt Safety Tips Every Employee Should Know Before Using AI Tools
Practical prompt hygiene for employees — the seven most common ways sensitive information ends up in AI tools, and how to avoid them without slowing down your work.
Key takeaways
- The seven core prompt safety habits: never paste credentials, redact customer names and contact details, treat AI input like an email to a stranger, check which account and data settings you are using, paste only the section you need, review AI output before sending, and use only approved tools.
- Replace real credentials with placeholders (
API_KEY=your-api-key-here) when asking AI for configuration help. - You rarely need a customer's name, email, or account number for the AI task to work — remove them first.
- AI can state wrong things confidently; verify facts, figures, and dates against a real source before sending anything.
- Free ChatGPT accounts have historically allowed training on your inputs; business plans let you disable it — check the setting.
Most data security guidance focuses on hackers. But one of the most common ways small business data ends up where it should not be is employees sending it to an AI tool — with good intentions and no awareness that it is a problem.
This is not about blame. It is about building habits that protect your business and your customers, without making your team's work harder.
Here are seven prompt safety habits worth building.
1. Never paste credentials into an AI tool
This is the most critical rule. Passwords, API keys, access tokens, database connection strings, and private SSH keys must never be pasted into a chat interface.
The risk is not only that the AI provider might store or train on your input. It is also that if you paste a key into a shared AI account — your company ChatGPT Teams workspace, for example — other users with access to the conversation history might see it.
If you are asking an AI to help with a configuration problem that involves credentials, replace the real values with placeholders before pasting: API_KEY=your-api-key-here.
2. Redact names and contact details from customer data
When you are using AI to analyze, summarize, or respond to customer information, you usually do not need the customer's name, email address, phone number, or account ID for the task to work.
Before pasting a customer email or support ticket, replace identifying details:
- Customer name → "the customer"
- Email address → removed
- Account number → "[account ID]"
The AI can still help you draft a response. The customer's PII stays off the AI provider's servers.
3. Treat AI input like email
A useful mental test: would you be comfortable emailing this content to a stranger? If not, do not paste it into an AI tool. This covers most cases without requiring your team to memorize a data classification policy.
This rule is especially useful for:
- Internal financial summaries
- Draft contracts or pricing terms
- HR-related discussions
- Business strategy documents
4. Check what account you are using
Who owns the AI account you are using? What are the data-sharing settings on that account?
The default settings on free ChatGPT accounts historically allowed OpenAI to use conversation data for training. Business and enterprise plans typically offer different settings. Know which you are on, and check the settings — most providers allow you to disable training on your data.
If you are using a work AI account, your IT or operations team should own the configuration. If you are using a personal account for work tasks, that is worth flagging — your work data may not belong in a personal account.
5. Do not paste whole documents when you only need part of them
When using AI to help with a task that touches a sensitive document — a contract, a financial statement, an HR record — paste only the section relevant to the task.
If you need help improving the language in clause 4 of a supplier contract, paste clause 4, not the whole contract. This limits the exposure of unrelated sensitive information.
6. Review before you send
AI-generated text can confidently state things that are wrong. Before sending any AI-drafted communication to a customer, supplier, or colleague:
- Read the whole thing.
- Check any factual claims, figures, or dates against a real source.
- Verify the tone matches your brand and the relationship.
This step is especially important for anything legal, financial, or medical. AI tools are very good at producing plausible-sounding text that is subtly incorrect.
7. Know your business's approved tool list
Using an AI tool your business has not approved means your employer has no visibility into how your work data is being processed, and cannot protect the business or you if something goes wrong.
Before using a new AI tool for work tasks, check whether it is on your company's approved list. If it is not, ask before you start — the approval process is usually quick, and it protects both you and your customers.
If your business does not have that list yet, ChatGPT rules for employees sets out a four-tier model for deciding what needs approval, and how to write an AI usage policy covers the document that names the owner of it.
Check your prompts before you send
The Prompt Safety Checker on this site runs your draft prompt through 13 detection patterns — API keys, PII, credentials, internal financial data, and more — entirely in your browser, with nothing transmitted to any server.
Use it when you are not sure if a prompt is safe to send.
Prompt safety is one habit inside a wider set of AI usage rules. To give your whole team clear, written guidance — data input rules, an approved-tool process, and output review checklists — see the Starter Pack.