Classify Your Work Data Before You Prompt

The first question before any AI task at work is: what data am I about to share? Not all work data is equal. A public press release and a client's payroll file need different care. Once you can sort data into a few simple levels, most AI decisions become easy.

This lesson gives you a five-level model you can use in seconds. Your company may use its own labels, and if it does, its labels win. The ideas carry over either way.

What You'll Learn

• Five common levels of work data, with examples
• What each level means for AI use
• How to reduce data before you share it, so a task moves to a safer level
• How to use AI itself to help you remove names and details

The five levels

Many companies sort data into levels. The names differ, but most look something like this.

Public. Already published or meant for anyone. Your website, a public price list, a published job post, a press release.

Internal. For staff only, but would cause little harm if it leaked. Meeting room rules, an internal newsletter, a general process guide, an org chart without personal details.

Confidential. Would hurt the company if it leaked. Unreleased product plans, financial forecasts, source code, pricing strategy, contract terms, security details, legal matters.

Personal data. Anything that identifies a living person. Names with contact details, ID numbers, health or salary details, performance reviews, customer records. This applies to staff, customers, and anyone else. Privacy laws like GDPR set rules here.

Client data. Anything a client gave you or that you made for them. This often comes with its own contract terms, which may limit where it can be stored or processed. Client data can also be confidential and personal at the same time.

What each level means for AI use

Two points stand out.

First, the tool matters as much as the data. Internal data in your company's approved AI tool may be fine. The same data in a personal account is not. The lesson on personal vs company accounts explains why.

Second, "allowed" is not the same as "needed." Even when a tool is approved, share only what the task needs. If you want help with the tone of a reply, the AI does not need the customer's account number.

When a file has more than one level

Real files mix levels. A project update might include public facts, an internal timeline, a confidential budget, and a list of client contacts. Apply the strictest rule of any data in it to the whole file. If one row is personal data, the whole file is personal data for AI purposes. If it holds both client data and personal data, follow both sets of rules.

Reduce the data, lower the risk

Often you can make a task safer by changing what you share, not by giving up on AI. A few ways:

• Remove names. Use "Client A," "Employee 1," or a role like "the account manager."
• Remove numbers that identify. Account numbers, ID numbers, phone numbers, addresses.
• Round or change figures. If you need help with the structure of a budget memo, the real totals may not matter. Use made-up numbers with the same shape.
• Describe instead of paste. "Write a polite reply to a customer who was charged twice and is upset" needs no real email at all.
• Share the part, not the whole. Paste the one paragraph you need help with, not the full contract.

This is called minimizing data. It is one of the most useful habits in this course. A task that starts as "confidential" can often become "internal" or even "public" with two minutes of editing.

Let AI help you remove details

You can ask an approved AI tool to help remove names and details from text that is already safe to share with it, for example a sample you made up, or internal text in a company tool. Then you check the result yourself before using the cleaned text elsewhere. Try it with the sample below.

Two warnings:

• The AI can miss things. Always read the cleaned text yourself. Indirect details, like a job title plus a small town, can still identify someone.
• Only clean text in a tool that is allowed to see the original. Pasting a real customer file into an unapproved chatbot "to anonymize it" is still sharing the real file.

A quick test when you are unsure

If you cannot decide the level, ask yourself: "Would I be comfortable if this text appeared in a news story with our company name on it?" If the answer is no, treat it as confidential at least. If it names people, it is also personal data, and if it came from a client, it is also client data, so follow those rules too. And if you still are not sure, ask your manager or IT before you paste.

Key Takeaways

• Sort work data into public, internal, confidential, personal, and client levels. Your company's own labels take priority.
• The safe choice depends on both the data level and the tool.
• A mixed file takes the strictest rule of any data in it.
• Minimize data: remove names and identifiers, change figures, describe instead of paste, share only the part you need.
• AI can help remove details, but check its work, and only use a tool that is allowed to see the original.