The data map

The approved list says which tools people can use. The data map says what can go into them. Most of the real risk with AI isn't the tool. It's what gets pasted in, and most people pasting have never been told where the line is.

What people are pasting

When companies measure what actually goes into AI tools at work, a large share of it is the material they'd least like to see leave the building: customer records, staff details, financial data, contracts. Ask people directly and nearly half say they've done it.

None of this means people are careless. It means nobody gave them a map. “Be careful with sensitive data” asks every person to draw the line themselves, every time, under deadline. A data map draws it once.

Four classes, and where each may go

The pilot kit uses four classes. Every piece of information your organisation holds fits one of them.

• Public: website copy, published prices, press releases that have gone out. It can go into any approved tool.
• Internal: plans, internal emails, numbers that don't identify anyone. Approved tools on business plans only.
• Confidential: client work, contracts, personal data about staff. Only tools the approved list marks “confidential OK”.
• Regulated: patient data, card data, anything a law or industry standard governs. Only tools with the required agreement in place, such as a business associate agreement (BAA) for health data under HIPAA.

Two rules make the map usable. When something could fit two classes, treat it as the higher one. And your organisation may move an example up a class, but never down: if you treat budgets as confidential, say so, and the map follows.

Try it. Each of these is the kind of thing someone pastes into a tool on an ordinary Tuesday.

Where people get it wrong

Most mistakes with the map aren't careless. They come from four places, and your examples should cover each one.

• Drafts. A press release is public once it's out. Until then it's internal, and if it mentions a client deal, it's confidential.
• Personal data that feels routine. A CV, a staff rota with phone numbers, an email thread with a customer's address in the signature. None of it feels sensitive in the moment. All of it is personal data.
• Pieces that add up. Sales by region is internal. Sales by region with each customer's name is confidential. Combining harmless files can move the result up a class.
• Attachments and recordings. People think about what they type. They forget the spreadsheet they attached, or the meeting recording a notes tool picked up.

One more rule closes the loop: what comes out takes the class of what went in. A summary of a confidential contract is confidential, however short it is. So is a draft reply built from a patient's notes.

Regulated data needs the agreement first

Regulated data is the one class where good intentions aren't enough. The law decides, and it usually asks for a signed agreement with the vendor before any data goes in. In US healthcare, that's a business associate agreement under HIPAA. For personal data under the GDPR, it's a data processing agreement. Card data falls under the payment card industry's own security standard, and it belongs nowhere near a general AI tool.

Which vendor plans include these agreements changes often, so treat it as a fact to check, not a fact to remember. Question 3 on the vendor checklist asks it every time, and a tool only goes up to “regulated” on your list once the agreement is signed. Even then, use the least data the task needs. Healthcare calls this the minimum necessary, and it's a good rule everywhere.

Put the map into the tools

A map on a page helps people decide. The tools are what make the decision stick. The column you filled in last lesson, the highest class each tool may take, is where the two meet. Your IT lead turns it into settings: single sign-on so only business accounts work, admin switches for the AI features, and, where you have them, data-loss rules that warn before a card number or a patient ID goes somewhere it shouldn't.

Set it below. Each row is a tool from your list; each column is a class. Then write the examples in your own words, because “client work” means something different in a law firm, a clinic and a construction company.