Using AI at Work Without Leaking Company Data

Most companies now have employees pasting internal documents, code, and customer data into ChatGPT, Claude, or Copilot every day, often without anyone signing off on it. The problem isn't AI itself. It's that the default habits people bring to these tools treat them like a private notepad, when in reality you're sending data to a third-party server with terms of service most people never read.

Know what actually happens to your input

Read the data usage terms for whatever tool you're using before you assume anything. ChatGPT's consumer tier (the free and Plus plans) may use conversations to train future models unless you opt out in settings. ChatGPT Enterprise and Team plans, along with the API, generally don't train on your data by default and offer stronger contractual guarantees. Same split applies to Claude (consumer vs. Claude for Work/API) and Gemini. The gap between "free personal account" and "business account with a data processing agreement" is the single biggest factor in your actual risk, and it's the one most people skip checking.

If your company hasn't purchased an enterprise tier for any AI tool, assume every prompt you type into a free-tier chatbot could end up reviewed by a human trainer or baked into a future model. That's not paranoia, it's the plain reading of most consumer terms of service.

Strip the data before you paste it

Before pasting anything into a prompt, ask what would happen if this exact text showed up in a data breach headline tomorrow. If the answer is bad, redact it first. Concretely:

• Replace real customer names, emails, and account IDs with placeholders (CUSTOMER_A, user@example.com) before asking for help drafting a response.
• Strip API keys, tokens, and internal hostnames out of code snippets. A quick grep -riE "(api[_-]?key|secret|token|password)" over the file you're about to paste catches most obvious leaks.
• Summarize the shape of a problem instead of pasting a whole proprietary file. "I have a function that rate-limits requests per user using a sliding window, and it's dropping legitimate requests under bursty traffic" gets you useful help without handing over your actual algorithm.
• For contracts, financials, or anything under NDA, don't use general-purpose chat tools at all unless your legal team has explicitly cleared that tool for that data class.

Set up guardrails your team will actually follow

A policy nobody reads gets ignored. A policy that fits into existing workflow gets followed. A few things that work in practice:

• Get IT to provision a business-tier AI account (ChatGPT Team, Claude for Work, Microsoft Copilot with your M365 tenant) so the default option for employees is the safer one, not the free public one.
• Use browser extensions or DLP (data loss prevention) tools like Microsoft Purview or Nightfall to flag when someone's about to paste something that looks like a secret or PII into a web form. Many companies already have DLP for email; extending the same rules to browser input to AI sites is a natural next step.
• Write a one-page policy that says, in plain language, which tools are approved for which data classes. "Public documentation and open-source code: any tool. Customer PII, financials, source code for proprietary systems: enterprise-tier tools only, or not at all." Specific and short beats long and vague.
• For code specifically, decide whether GitHub Copilot or a self-hosted alternative (like a local Code Llama or StarCoder deployment) makes more sense for your most sensitive repos. Copilot Business lets you exclude specific repos from suggestions and has code-matching filters to reduce the chance of regurgitating licensed snippets.

Watch for the sneaky leak paths

Direct copy-paste isn't the only risk. Browser extensions that summarize pages can quietly send whatever's on your screen, including internal dashboards, to a remote API. AI-powered note-taking bots that join your video calls (Otter.ai, Fireflies, etc.) record and transcribe everything said, including anything confidential mentioned in passing. Before adding any of these to a meeting, ask whether everyone in the room is fine with a third-party service holding a transcript of that conversation, and whether the vendor's retention policy matches your company's data handling rules.

Check what plugins or connectors are enabled in your AI chat client too. Some let the model browse the web or access connected apps like Google Drive, which can pull in more context than you intended to share.

Build the habit, not just the policy

The goal isn't zero AI use, it's AI use where everyone treats the input box like an external email to a vendor, because that's functionally what it is. Once that mental model sticks, most of the risky behavior stops on its own.

For more on this, check out Korra Studio's segments on data security fundamentals and on practical AI tooling for engineering teams.