ChatGPT and client data: what actually happens when you paste a document
For accountants and corporate advisors, a useful AI prompt can contain more client information than the task requires. A review before pasting changes what you share.
A paste is a disclosure
Suppose you want ChatGPT to summarize a corporate registry extract. You copy the text, paste it into a chat, and ask for the ownership structure. The service receives the text you pasted to process your request. The same basic point applies if you paste it into Claude: the text leaves your browser and goes to the AI provider.
That may be an acceptable decision for a particular task. But it should be a decision. A registry extract can contain directors’ names, tax codes, addresses, company VAT numbers and other details. An engagement letter may add a client’s email address, telephone number and bank account. Those details are often beside the question you actually want answered.
The relevant question is not simply “Can this AI read the document?” It is “Which parts of this document does the AI need to read to do this job?”
The identifiers are part of real work
A professional document rarely arrives as a tidy, anonymous example. The names in an engagement letter identify the parties. An IBAN may sit beside payment terms. A tax code can appear repeatedly in headers or signatures. When you paste the entire text, those details travel with the clauses you wanted to discuss.
Consider a request such as: “List the deadlines and responsibilities in this engagement letter.” The AI needs the obligations and dates. It may need to distinguish the firm from the client, but it usually does not need their real names, tax codes or bank details. Placeholders can preserve that distinction: [CLIENT], [FIRM] and [IBAN] still let the model follow the document.
Masking has limits. A distinctive transaction, address or combination of facts might identify someone even after obvious identifiers are replaced. A professional still needs to read the prepared text and decide whether its remaining detail is appropriate to share.
Put a review step before the prompt
Riducta helps you choose what data to share with AI. You bring in the document, and Riducta detects tax codes, IBANs, VAT numbers, email addresses and telephone numbers in supported formats. You indicate or confirm the names to mask. It prepares text with placeholders in place of the selected values, and you check that text before copying it into ChatGPT or Claude.
That order matters. Detection is useful, but a name can have several meanings in a document, and an identifier can appear in an unexpected format. Your review is where you catch what needs attention. Read the prepared text as if it were about to leave your desk, because pasting it is exactly that next step.
The document preparation and masking run in your browser. The AI provider receives the text you choose to paste. Riducta does not make the AI conversation private by itself, and it cannot decide whether a particular client matter belongs in a third-party service. It gives you a practical point at which to make that choice.
After the AI replies, you can paste the response back into Riducta to restore masked values from the placeholders. Check the restored answer too, especially before putting it into client work. The AI may reword, omit or alter parts of its response; a restored draft still needs professional review.
Use the AI for the part you need
Take the registry extract again. Your aim might be a short note identifying shareholders, directors and changes worth checking. First, decide whether the extract is suitable for the service you plan to use. Then prepare it in Riducta, review the names and detected identifiers, and inspect the resulting text. If a particular address is unnecessary, remove it before pasting. If a name is necessary to answer the question, decide whether to leave it visible or use a consistent placeholder.
Now give the AI a narrow instruction: “Summarize the ownership and management details in this text. Flag anything that needs checking against the original extract.” Compare its answer with your source. Paste the response into Riducta if you need the masked values restored, then verify the finished note.
This workflow is also useful when preparing text from PDFs. It is separate from the decision about personal data, but it can make prompts smaller. With Claude, Riducta-prepared text used about half the tokens of uploading the same PDF as a document in a September 2026 test: a median 52% reduction across 20 public PDFs, measured with Anthropic’s official token count. Results vary by document. For ChatGPT, Riducta’s Fedele mode keeps every word of the extracted text and used about 4% fewer tokens in the stated comparison; Snella also removes repeated lines and indexes and used about 6% fewer. Layout and images are outside those text comparisons.
The useful habit is simple: inspect the source, choose what the AI needs, and inspect what you are about to send. Riducta makes that step part of the workflow. Try it at riducta.com. Free for a single document, Pro for working every day.
ChatGPT is a trademark of OpenAI; Claude is a trademark of Anthropic. Riducta is not affiliated with or endorsed by either.