How to Give an AI Context Without Sharing Too Much Data

The whole document is not the best prompt

When an AI asks for context, the natural reaction is to paste the full email, contract, or conversation. It is fast and can produce a more fluent answer. But it confuses two things: supplying enough information to do a task and transferring everything you hold to a tool.

The useful rule is not “never share data,” which would block legitimate uses. It is this: share the minimum context that allows the result you want. The habit resembles a familiar data-protection idea: Article 5 of the GDPR requires personal data to be adequate, relevant, and limited to what is necessary for the purpose. This is practical explanation, not legal advice; the applicable rules depend on the case and provider. As a working habit, though, it applies widely.

The capability to take away: turn a real document into useful minimum context by separating the goal, necessary facts, identifiers, and details that are not needed.

Start with the output, not the file

Before opening a chatbot, write what you want back. “Put this contract’s payment obligations and dates into a table” is a task. “Read this contract and tell me what you think” invites more material and produces an answer that is hard to review.

A defined output decides what information is needed. To extract dates, selected clauses and a calendar may be enough; to draft a commercial reply, a product sheet, tone, and the customer’s objections may be enough, without names, phone numbers, or full history. The goal turns minimisation into an operation rather than a slogan.

Make four columns before you paste

Read the material and sort each element into one of four columns:

1. Goal and constraints. What the AI must do, for whom, and in which format.
2. Necessary facts. Information without which the task changes: dates, conditions, aggregated amounts, technical requirements, or applicable rules.
3. Identifiers. Names, addresses, email, phone numbers, case numbers, customer IDs, signatures, credentials, and private links.
4. Surplus details. Comments, histories, attachments, metadata, and conversations that do not affect the requested output.

The first two columns form the context. The third needs a specific choice: remove it, replace it with a label, or use a tool approved for that kind of data. The fourth normally stays out. Do not confuse a label with a guarantee: changing “Marta López” to “[CUSTOMER A]” reduces exposure in the text, but does not by itself remove re-identification risk or an organisation’s obligations.

Replace without distorting the task

Useful anonymisation preserves the relationships the task needs. If you are comparing discounts, you can use “customer A” and “customer B” while keeping relevant amounts and dates. If you are drafting a reply, you can say “returning customer, annual contract, open incident” without supplying an address or ID number.

There is an important exception: do not remove a fact that changes the meaning. A medical condition, country, exact date, or confidentiality clause can be essential to the answer. The solution is not blind deletion; ask, “If I replace this, can the AI still answer correctly?” If not, raise the level of care: use the authorised environment, review settings and permissions, and consult the appropriate person in your organisation.

Review the route, not only the text you see

Before sending, consider where the context will live. Is this a personal or work account? Are there connectors to email, storage, or third-party tools? What do settings say about retention, training, export, and access? Who else can open the result?

Privacy management is not reduced to crossing out a name. The NIST Privacy Framework treats privacy through organisational risk management. For an individual, the practical equivalent is to know the information’s route before choosing a tool. For a company, the decision needs policy, contracts, controls, and accountable people; it is not solved in a prompt.

A template that reduces exposure

Instead of pasting a full file, try this structure:

• Task: “Extract obligations and dates into a table.”
• Necessary context: only relevant clauses, with labels for people and organisations.
• Constraints: “Do not infer facts absent from the text; mark gaps.”
• Output: defined columns and a reference to the source fragment.

The last instruction has two benefits: it reduces pressure to invent and makes human review easier. If the AI needs information you did not include, it should ask; that is not a preparation failure, but a signal that the minimum context is not yet sufficient. Add it deliberately rather than opening the gate to the whole document.

What should stay out by default

Treat passwords, API keys, recovery codes, full bank details, identity documents, medical histories, children’s data, trade secrets without agreed protection, and access-granting links as pause signals. A model being able to read a datum does not mean you should provide it. If work depends on it, do not improvise a text solution: use the authorised tool and process.

NIST’s Generative AI Profile identifies governance, content provenance, pre-deployment testing, and incident disclosure as central considerations. The everyday lesson is simple: before placing AI inside a flow with sensitive information, define who may use it, with which data, how output is checked, and what happens if something goes wrong.

Less context, a better conversation

Reducing data is not only a cautious measure. It often improves the question’s quality. Clean context makes it clearer which fact supports each conclusion, lets you repeat the query, and makes another person’s review easier.

The next time an AI asks for “more context,” do not automatically reply with the full file. First ask: Which exact fact does it need to complete this task? Answering that protects you from oversharing and teaches a more precise way to work with AI.