The tension
Better context usually means better answers. An assistant that knows the audience, the goal and the background can give a far more relevant response than one working from a one-line request. But the easiest way to supply context is to paste everything: the entire email thread, the full spreadsheet, the customer’s details. That is where problems begin.
The aim is not to share nothing. It is to share what the answer depends on and leave out what it does not.
Ask what the answer actually needs
Before including a detail, ask whether the response would change if you left it out. Most of the time, identity does not matter; structure and situation do.
- A request to help word a difficult message to an employee needs the situation (a pattern of missed deadlines, a supportive tone) but not the employee’s name, ID or health circumstances.
- A request to tidy a customer list needs the column layout and a few fake rows, not real customers.
- A request for help with a contract clause needs the clause’s wording, perhaps without the party names.
Often a small shift is enough: describe the type of thing instead of the specific thing.
Common kinds of information to keep out
Categories that usually call for caution:
- Personal data: names, addresses, phone numbers, account numbers, government identifiers and anything that can identify a person, especially customers, patients, students and employees.
- Sensitive personal matters: health, finances, legal trouble, disciplinary issues and similar subjects. Our article on health privacy covers why these deserve more care.
- Confidential business information: unreleased plans, pricing, source code, credentials, security details and trade secrets.
- Third-party material: documents covered by a confidentiality agreement or that belong to a client.
- Passwords and access keys: never include these.
Rules differ between organizations and countries. The topic of data protection explains the principles, and your workplace policy will list what is allowed with which tool.
Techniques that keep the usefulness
Use placeholders. Replace real details with labels: “[CLIENT],” “[PRODUCT],” “[AMOUNT].” Fill them in on your own side afterwards.
Generalize. “A mid-sized client in the logistics sector” can often stand in for a named company.
Use invented examples. If the assistant needs to see the shape of the data, create sample rows with made-up values. State clearly that they are made up.
Summarize instead of pasting. Write two or three sentences describing the situation rather than uploading the full thread. This also forces you to decide what matters.
Trim documents. Share the relevant paragraph, not the whole file.
Add real details later. Get the structure or wording from the assistant using placeholders, then insert specifics offline.
An example: instead of “Write a reply to Dana Whitfield at Brightline Ltd about invoice 20418 for $4,300 that is 45 days overdue,” try “Write a firm but courteous reply to a client whose invoice is 45 days overdue. Mention the invoice number [NUMBER] and amount [AMOUNT] as placeholders. Offer a call to resolve any issue.” The result is nearly as useful and carries no private details. (All the names and numbers in this paragraph are invented for illustration.)
Know your tool
Not every tool treats your input the same way. Some may keep conversations, use them to improve the product, or let administrators see them; others offer stricter business settings. Details change and vary by provider and plan, so read the terms, check privacy settings, and see what data retention means for the tools you use. If your company has approved a specific tool, use it for work content rather than a personal account; skipping approved routes is how shadow AI starts.
When uncertain, ask your IT or privacy contact before pasting. A short question is cheaper than a data incident.
Think about what comes back, too
Context affects output. If you paste a document with errors or bias, the response may repeat them. If you include opinions about a person, the assistant may build on them. Keep descriptions factual and neutral, especially when the topic involves people, as in hiring or performance. Concerns about algorithmic bias do not disappear because you supplied the context yourself.
A quick pre-send check
Before pressing enter, scan your request for:
- names, contact details or identifiers that are not needed;
- figures or plans that are confidential;
- anything covered by an agreement or regulation you are unsure about;
- credentials or links that give access to systems;
- whether this tool is approved for this kind of content.
If any answer gives you pause, replace the detail with a placeholder or leave it out.
Careful context is a skill like any other. With practice you learn how little real detail most requests need, and you build habits that protect other people’s information as well as your own.