Using AI Tools at Work Without Leaking Personal Data

How personal data and secrets leak into AI tools through staff prompts and AI features in your apps, and the practical controls that reduce the risk.

AI assistants have become part of everyday work. People use them to draft emails, summarise documents, write code and analyse spreadsheets. The productivity gains are real, and so is the risk: every prompt is data leaving your organisation.

This guide covers where the leaks happen and practical controls that reduce them without stopping people from using AI.

Where the leaks happen

Staff pasting into chat tools. Someone pastes a customer complaint with the customer’s phone number. A recruiter uploads a stack of résumés to compare candidates. An engineer pastes a log file containing access tokens. None of them meant any harm.

AI features inside your own applications. When your product sends user content to a model provider, whatever is in that content goes too, including personal data your users typed in.

Files and attachments. Uploading a spreadsheet or document sends everything in it, not just the part someone wanted help with.

Agents and integrations. AI agents connected to your email, files or databases can pull far more data into a prompt than a person would paste by hand.

Why “just be careful” does not work

Most leaks are not deliberate. People are busy, the tool is helpful and the sensitive detail is buried in the middle of a long paste. Training matters, but it cannot catch everything. You need controls that work even when someone is in a hurry.

Practical controls

1. Decide which tools are approved

Pick the AI tools your organisation supports, prefer plans with clear terms on data retention and model training, and tell staff which ones to use. Unapproved tools on personal accounts are the hardest to see or control.

2. Write a short, clear rule

A one-page rule beats a long policy nobody reads. Cover what can never go into an AI tool, what is fine, and who to ask when unsure. Use real examples from your own work.

3. Check prompts before they leave

Technical controls can inspect prompts and files on their way to an AI model and act on what they find:

  • Redact a value, replacing it with a placeholder.
  • Tokenise a value, so the model sees a stand-in and the real value is restored in the answer.
  • Warn the person and let them decide.
  • Block the request entirely for the most sensitive data.

For staff, this usually means a managed browser extension. For applications, it means routing model calls through a gateway that checks each request.

4. Roll out in warning mode first

Start by warning rather than blocking. You learn what people actually send, tune detection to cut false alarms and avoid breaking work on day one. Tighten rules once you understand the patterns.

5. Keep records without keeping the data

Record what was detected and what action was taken, but avoid storing the prompt text itself. Otherwise your safety system becomes another place sensitive data piles up.

6. Know which AI systems you run

Keep a register of the AI tools and features in use across the organisation, what data each one touches and who owns it. That register is the starting point for any wider AI governance work.

A simple starting plan

  1. List the AI tools staff use today. Ask them; the honest answers are useful.
  2. Approve two or three, and publish a one-page rule.
  3. Add prompt checking in warning mode for the most common tools.
  4. Review what was detected after a month, and adjust.
  5. Bring your legal and security teams into the review, especially for uses involving customer or employee data.

AI tools are not going away. The goal is to let people use them while keeping personal data where it belongs. For the wider picture of where personal data lives in your organisation, see how to find the personal data hiding across your systems.

Where Arborworld fits

Arbor Guard stops personal data and secrets reaching AI models, from your applications and your people.

Explore Arbor Guard

Questions people ask

Should we just ban AI tools at work?

Bans tend to push usage to personal accounts and devices, where you have no visibility at all. Most organisations get better results by approving specific tools, setting clear rules and adding technical controls.

Are enterprise AI plans safe for personal data?

Enterprise plans often offer better terms on data retention and training, which helps. They do not decide what your staff paste in, so you still need rules and controls on what data reaches the model.

What data should never go into an AI tool?

Start with identity numbers, financial details, health information, passwords, API keys and customer records. Your legal and security teams can extend the list for your organisation.

Vishwanath CK
Written by

Vishwanath CK

Vishwanath CK is a founder of Arborworld, which builds software for hiring, HR, and personal data protection. Arborworld's two founders bring over 45 years of shipping products at scale across travel, e-commerce, fintech, gaming, and mobile.

More from Arborworld resources