How to Use ChatGPT for Work Without Sharing Sensitive Data

AI assistants are useful for work. They are also cloud services that process everything you send them. For most business use, this creates a real tension: the tasks where AI could save the most time — contract review, client communication drafting, sensitive analysis — are also the tasks where you cannot simply paste everything into ChatGPT.

This guide covers practical techniques for getting productive value from AI tools while keeping confidential information protected.

Understand what actually happens to your data

Before making data decisions, understand the actual policies:

ChatGPT (OpenAI):

Free tier: Conversations may be used to improve models unless you opt out in settings
Plus and Pro: Data controls available; you can disable model training on your conversations
Team: No training on business conversations by default
Enterprise: BAA available, no training on your data, 30-day retention limit

Claude (Anthropic):

Free and Pro: Conversations may be used for training unless you opt out; opt-out in privacy settings
Enterprise: BAA available, stricter data handling

The opt-out: On most paid plans, you can disable training on your conversations. Do this in settings if you have not already. It does not prevent Anthropic or OpenAI from storing your conversations, but it removes them from training pipelines.

The key distinction: "not trained on your data" is different from "not stored." Even with training disabled, most AI providers retain conversations for some period for abuse detection and safety monitoring.

Techniques for working with sensitive content

Technique 1: Anonymize before pasting

The most common mistake: pasting a contract or client email with all the real names, company names, and specific details.

Before pasting sensitive content:

Replace real company names with placeholders (Company A, Client B)
Replace personal names with roles (Client, Lawyer, CEO)
Remove contract values, account numbers, and specific identifiers
Replace specific product names with generic descriptions

"Review this contract and flag unusual indemnification language" with an anonymized version gets you the same legal analysis without exposing client identity.

This works for most analysis tasks. The AI does not need to know who Company A is to analyze their contract language.

Technique 2: Describe the structure, not the content

For many analysis tasks, you can get useful output by describing the content rather than sharing it.

Instead of: "Here is the P&L for Q3, what should I focus on?"

Try: "I have a P&L showing 23% gross margin (down from 31% last year), a 40% YoY increase in R&D spend, and flat SaaS revenue with declining professional services. What are the likely causes and the questions I should investigate?"

The AI works with the relevant numbers without seeing the actual file. This is often sufficient for analytical questions.

Technique 3: Use AI for the format, fill in the sensitive parts yourself

AI is useful for creating templates and structures that you then fill in with the real data.

"Give me a structure for a board presentation on Q3 financial performance, covering quarterly results, drivers, risks, and forward outlook."

You get the structure from AI, then add the actual numbers and specifics yourself in your own tool. The sensitive information never enters the AI conversation.

Technique 4: Use an enterprise plan or local AI for high-sensitivity work

For genuinely high-sensitivity work — client documents, M&A materials, healthcare data, legal privileged content — the safest approach is either:

Enterprise AI with appropriate contracts: ChatGPT Enterprise or Claude Enterprise with a signed Business Associate Agreement (for healthcare) or Data Processing Agreement. Understand exactly what data retention looks like.
Local AI: Run a local model via Ollama or LM Studio. Your data never leaves your machine. Quality for document analysis using Llama 4 Maverick or Qwen 2.5:32b is sufficient for most business tasks.
Dedicated enterprise tools: Harvey (legal), Nuance DAX (healthcare), and sector-specific AI tools built with the appropriate compliance architecture for that industry.

Technique 5: Check your organization's policy first

Many organizations are drafting or have drafted AI acceptable use policies. Before working around data concerns individually, check if your organization has guidance. Violating an employer's data handling policy — even with good intentions — creates compliance and employment issues.

If no policy exists and you work with genuinely sensitive data, raising the question with IT or legal is appropriate before experimenting.

Not all information is equally sensitive. These categories are generally safe for public AI assistants:

Publicly available information you are asking to summarize or analyze
Generic business writing that does not identify clients or contain proprietary data
Code that does not contain API keys, passwords, or proprietary algorithms
Career and professional development questions
General industry analysis and research

The principle: would you be comfortable if this conversation were seen by a regulator, opposing counsel, or your client? If yes, the risk is low.

The local AI option is more accessible than most people know

The default assumption is that local AI is worse and harder to set up. In 2026, both parts of this assumption are increasingly wrong.

Ollama installs in one command. Llama 4 Maverick or Qwen 2.5:32b on a machine with 32GB RAM produces output that handles most business document analysis tasks adequately. For a 45-minute setup investment, you have an AI assistant that never sends any data to any server.

For professionals in law, healthcare, finance, and any regulated industry, this setup deserves serious consideration alongside enterprise cloud options.