Updated November 7, 2026. Data settings were checked against each vendor’s own help and privacy pages this month. Menus and defaults change often, so confirm them in your own account. This guide is general information, not legal advice.
Most freelancers don’t lose clients over bad work. They lose them over a broken promise, and pasting a client’s unreleased product plan into a chatbot that trains on your chats is exactly that kind of promise. The good news is that the fix takes about an hour: a few settings, the right plan for client work, a short rule for what never gets pasted, and a clear note to clients about how you use AI.
This guide walks through AI client data privacy step by step. You’ll get a table of where each major tool’s training setting lives, a traffic-light system for sorting client data, a redaction template, an NDA rider you can adapt, and a client-facing checklist you can send with your next proposal.
Quick Answer: Sort client material into what can, can’t, and can only go into AI after redaction. Then turn off model training in every AI tool you use (ChatGPT, Gemini, Copilot, Grammarly, and others train on personal-plan data by default or ask you to choose), use temporary or incognito chats for sensitive one-offs, and move regular client work onto a business plan, where the major vendors say they don’t train on your data by default. Read each client’s NDA for AI or “third-party processor” language before uploading anything, and tell clients in writing which tools you use and how. Review your settings every quarter, because defaults change.
Why Client Data Needs Its Own AI Rules
When you paste text into an AI tool, four things can happen to it that your client may not expect:
- Training. On many personal plans, your chats can be used to improve the vendor’s models unless you switch that off.
- Human review. Google’s Gemini Apps Privacy Hub says “a subset of chats are reviewed by human reviewers” and asks users not to enter “confidential information that you wouldn’t want a reviewer to see.”
- Retention. Even with training off, most tools keep chats for a period (often 30 days or 72 hours) for safety and abuse checks.
- Exposure through you. Shared links, a skill or project file uploaded to the wrong account, a meeting bot that joins calls automatically, or a browser extension that reads every page you open.
None of this means you should stop using AI on client work. You just need the habits you already use for email and cloud storage: know where data goes, keep sensitive parts out, and be able to explain your process in two sentences.
AI Data Settings at a Glance
| Tool | Personal plans: trains on your data? | Where to turn it off | Business plans |
|---|---|---|---|
| ChatGPT | Yes, unless you opt out | Settings > Data Controls > “Improve the model for everyone” | Business, Enterprise, Edu, and API: not trained on by default |
| Claude | Only if you allow it (choice set at sign-up) | Settings > Privacy > model improvement toggle | Team, Enterprise, and API: not trained on by default |
| Gemini | Yes, while Keep Activity is on; some chats are human-reviewed | Gemini Apps Activity > turn off Keep Activity | Workspace accounts: not used for training, no human review |
| Microsoft Copilot | Yes, unless you opt out or are in an exempt group | Profile > Privacy > “Training on conversation activity” and “Training on voice conversations” | Work (Entra ID) accounts: not used to train foundation models |
| Grammarly | Yes, on by default for individual accounts | Account > Privacy > “Product Improvement and Training” | Enterprise and sales-bought plans: off by default; website-bought team plans: admin can turn it off |
| Otter.ai | Trains on de-identified recordings and transcripts | No training-specific toggle found; check the vendor’s site | Not stated; ask Otter before recording client calls |
Table: Checked November 7, 2026, on each vendor’s help or privacy pages. Labels and defaults vary by app and region, so check your own settings.
Step 1: Sort Client Data Into Red, Amber, and Green
Before you change a single setting, decide what kind of material you handle. A simple traffic-light system works for most freelancers and is easy to explain to clients.
- Red: never goes into an AI tool. Passwords, API keys, bank and card details, government ID numbers, health records, signed contracts, unreleased financials, anything the client marked “confidential,” and anything an NDA says can’t leave their systems.
- Amber: only after redaction or on a business plan. Draft strategy documents, customer lists, interview transcripts, internal emails, and unpublished campaign plans. Strip names and numbers first, or use a plan where the vendor contractually doesn’t train on your data.
- Green: fine on any plan with training off. Published web copy, public product pages, your own drafts, generic briefs, and style guides the client has approved for sharing.
Write the list down in one short document. It becomes the backbone of your client-facing policy in Step 7.
Step 2: Turn Off Training in Every Tool You Use
This is the biggest single change you can make, and it’s free. Do it on every account you use for client work, including old free accounts.
- ChatGPT. OpenAI’s data controls FAQ gives the path: profile icon > Settings > Data Controls, then turn off “Improve the model for everyone.” OpenAI says that once it’s off, new conversations still appear in your history but aren’t used for training. Temporary Chats are never used for training and are deleted after 30 days.
- Claude. On Free, Pro, and Max, Anthropic asks you to choose whether your chats can be used to improve its models. Its privacy help page puts the toggle under Settings > Privacy. With it on, de-identified data can be kept for up to 5 years; with it off, Anthropic’s consumer terms update says the 30-day retention period applies. Anthropic says incognito chats aren’t used for training even when the setting is on.
- Gemini. Open Gemini Apps Activity and turn off Keep Activity. Google’s privacy hub says chats are then kept with your account for 72 hours and aren’t used to train its models unless you send feedback. Temporary chats get the same 72-hour treatment. With Keep Activity on, the default auto-delete period is 18 months.
- Microsoft Copilot. Microsoft’s Copilot privacy controls page lists two toggles, “Training on conversation activity” and “Training on voice conversations.” On copilot.com they’re under your profile > Privacy; in the desktop app, profile > Settings > Privacy. Microsoft’s privacy FAQ also lists some countries and account types excluded from training, but switching both toggles off removes any doubt.
- Grammarly. Grammarly’s help page says “Product Improvement and Training” is on by default for individual users. Turn it off in your account’s privacy settings. Grammarly reads almost everything you type in a browser, so this one matters more than people think.
- Everything else. For any other AI tool, search its help center for “training,” “model improvement,” or “data controls.” If you can’t find a clear answer, treat the tool as Green-only.
One catch to know: thumbs-up and thumbs-down feedback can override your settings. Anthropic says feedback stores the whole related conversation and may be used for training, and Google says the same for Gemini feedback. Don’t rate responses in client chats.
Step 3: Put Regular Client Work on a Business Plan
Personal plans with training off are fine for Green material. If you regularly handle Amber material, a business plan is worth the extra few dollars a month, because the data promises are contractual rather than just a setting you might forget.
| Plan | Training default | Price (checked November 2026) | Good fit for |
|---|---|---|---|
| ChatGPT Business | Not trained on by default; DPA available | $25 per user per month, or $20 billed annually (standard seat) | Freelancers who live in ChatGPT and want admin controls |
| Claude Team | Not trained on by default; DPA included in terms | $25 per user per month, or $20 billed annually (standard seat) | Long documents and client projects in Claude |
| Google Workspace with Gemini | Prompts treated as customer data; no training, no human review | From $7 per user per month on an annual plan (Business Starter) | Freelancers already on Gmail and Drive for client work |
| Microsoft 365 Copilot Chat (work account) | Enterprise data protection; not used to train foundation models | Check the vendor’s site | Freelancers whose clients work in Microsoft 365 |
Prices from each vendor’s pricing page or announcement, checked November 7, 2026. Taxes, promos, and regional pricing vary.
Some details are worth knowing before you upgrade:
- Most business plans need a minimum number of seats or a work domain. Check the sign-up page; some let a solo freelancer buy one or two seats, others don’t.
- A DPA is the document clients’ legal teams ask for. OpenAI’s business data page says it supports customers signing a Data Processing Addendum, and Anthropic’s DPA is incorporated into its commercial terms.
- Data residency is usually Enterprise-only. OpenAI offers regional storage for eligible Enterprise, Edu, Healthcare, and API customers, not Business. If a client requires data to stay in the EU or UK, confirm it in writing before you start.
- Business plans still retain data. Google says Workspace admins choose 3, 18, or 36 months for Gemini app history. Set the shortest period that works for you.
If you’re weighing which assistant to standardize on, our ChatGPT alternatives and Claude alternatives guides compare the main options by workflow, and Best AI Tools for Business covers team plans more broadly.
Step 4: Check Your NDAs and Contracts
Many NDAs written before 2023 never mention AI, but they often ban sharing confidential information with “third parties” or require it to stay on “approved systems.” An AI vendor can count as a third party. Before uploading Amber material for a client, read their NDA and contract for these phrases:
- “Third party,” “subprocessor,” or “service provider” disclosure limits.
- Requirements to store data in a specific country or on client-approved systems.
- Any explicit ban on AI, machine learning, or “automated tools.”
- Ownership and confidentiality of “work product,” which may include your prompts and drafts.
- Breach notification duties and timelines.
If the contract is silent, don’t assume silence means yes. Ask, and get the answer in writing. Here’s a short rider you can adapt and propose. Have a lawyer review it for your jurisdiction.
AI Tools Rider
1. Contractor may use the AI tools listed in Schedule A to assist with drafting, editing, research, and analysis.
2. Contractor will only use accounts on which the provider does not use Client data to train its models, either by plan terms or by account setting.
3. Contractor will not enter Client Confidential Information marked "Restricted" into any AI tool without Client's prior written approval.
4. Contractor will review all AI-assisted output for accuracy before delivery and remains responsible for the work.
5. On request, Contractor will update Schedule A and confirm the data settings in use.
Schedule A: [Tool name, plan, training setting, retention setting]Step 5: Redact Before You Paste
For Amber material, replacing names and numbers with placeholders lets you get most of the AI’s help without exposing the sensitive part. You keep the key on your own machine and swap the real values back in after.
- Make a copy of the document. Never redact the original.
- Replace each identifying item with a labeled placeholder and record it in a private key file.
- Round or band numbers if exact figures aren’t needed (“revenue between $1M and $2M”).
- Remove file metadata, tracked changes, and comments before uploading. They often contain names.
- Paste the redacted version, then restore the real values in your final draft offline.
Copy this key template into a local note:
REDACTION KEY: [Client code] [Date] (keep offline, never upload)
[CLIENT] =
[PRODUCT-1] =
[PERSON-1] = (role: )
[PERSON-2] = (role: )
[CITY] =
[FIGURE-1] =
[URL-1] = A prompt line also helps: “Placeholders in square brackets are intentional. Keep them exactly as written in your answer.” The AI will then return text you can restore in seconds.
Step 6: Lock Down Meeting Bots, Extensions, and Connectors
Chat windows get the attention, but the bigger leaks often come from tools that run in the background.
- Meeting note takers. Otter.ai’s privacy policy says it trains its own AI on de-identified audio recordings and transcripts, and its security page says recordings aren’t manually reviewed by a human without explicit consent. Whatever note taker you use, turn off auto-join, ask the client before recording, and delete recordings once the summary is approved. Our AI meeting note takers roundup covers each tool’s options.
- Browser extensions. Writing assistants and AI sidebars can read every page you open, including client dashboards. Disable them on client portals or use a separate browser profile for client work.
- Connectors. When you connect an AI tool to Gmail, Drive, or Slack, it can read whatever that account can. Connect a work account scoped to client folders, not your whole personal drive.
- Shared links and projects. Review who can see shared chats, projects, and saved skills. Our guide to building reusable AI skills explains why client secrets should never live inside a skill file.
Step 7: Tell Clients How You Use AI
Clients rarely object to AI. They object to finding out late. A one-page summary, sent with your proposal or onboarding pack, turns a risk into a selling point. Here’s a client-facing checklist you can copy:
How I protect your data when using AI
[ ] I use these AI tools: [list tools and plans]
[ ] Model training is turned off on every account I use for your work, or the plan's terms exclude training.
[ ] I never enter passwords, payment details, personal ID numbers, or files you mark "Restricted."
[ ] Customer names and figures are replaced with placeholders before any AI use, unless you approve otherwise.
[ ] Meeting recordings happen only with your permission and are deleted after you approve the notes.
[ ] Every AI-assisted draft is checked and edited by me before you see it.
[ ] I'll update this list if my tools change and tell you within [X] days of any data incident.
Questions or preferences? Reply and I'll adjust my process for your account.And a short email to send to existing clients:
Subject: Quick note on how I use AI tools
Hi [Name],
I use a few AI tools to speed up drafting and research. I've attached a one-page summary of how I keep your information safe, including which tools I use and what I never share with them.
If your company has its own AI policy or anything you'd like me to avoid, just let me know and I'll follow it.
Thanks,
[Your name]If you’re also reworking your proposal process, our guide to AI proposal software for freelancers shows where a document like this fits.
Step 8: Review Your Setup Every Quarter
Defaults change. Anthropic changed its consumer training rules in 2025, and Grammarly’s help pages now sit under a new parent brand. Put a 20-minute reminder in your calendar every three months:
- Recheck the training setting in every tool from Step 2.
- Delete old chats, uploads, and recordings for finished projects.
- Remove connectors and extensions you no longer use.
- Update the tool list you’ve sent to clients.
- Skim each vendor’s privacy page for changes to retention or training.
Common Mistakes to Avoid
Mistake 1: Assuming paid means private. A paid personal plan such as ChatGPT Plus or Claude Pro still follows consumer data rules. Business plans are where “no training by default” is standard.
Mistake 2: Forgetting about feedback buttons. Rating a response can send the whole conversation for training even with the main setting off.
Mistake 3: Treating temporary chats as invisible. They’re excluded from training, but vendors still keep them for a short period, typically 30 days or 72 hours.
Mistake 4: Letting meeting bots auto-join. A note taker that joins every calendar event will eventually record a call a client expected to be private.
Mistake 5: Never telling the client. A clear AI policy up front costs nothing. Explaining after a client discovers it costs trust.
Frequently Asked Questions
Is it safe to use ChatGPT with client data?
It can be, with the right setup. Turn off “Improve the model for everyone,” keep Red material out entirely, redact Amber material, and use a Business or Enterprise plan for regular client work, which OpenAI says isn’t trained on by default.
Does turning off training delete my chats?
No. It stops future chats from being used for training. Your history stays until you delete it, and vendors keep deleted or temporary chats for a short safety period.
Which AI tools don’t train on my data by default?
On business plans, ChatGPT Business and Enterprise, Claude Team and Enterprise, Gemini in Google Workspace, and Microsoft 365 Copilot with a work account all state they don’t use your data to train their models by default. On personal plans, Claude asks you to choose; most others are on by default.
Do I need my client’s permission to use AI?
Check your contract and NDA first. If they’re silent, ask anyway and get the answer in writing. This isn’t legal advice, so talk to a lawyer if the work involves regulated data such as health or financial records.
What should I do if I pasted something sensitive by mistake?
Delete the chat, check that training is off, and note what was shared and when. Then follow your contract’s notification terms, and tell the client promptly if confidential information was involved.
You May Also Like
- Best AI Meeting Note Takers for Freelancers (2026): 7 Picks
- How to Build Reusable AI Skills for Client Work (Gemini, ChatGPT, Claude)
- How to Use AI for Freelancing: Automate Client Work and Reclaim Your Time
- ChatGPT vs Claude for Marketing Copy: Which Is Better in 2026?
- The Freelancer’s Complete AI Toolkit (2026 Edition)
- Best AI Tools for Business (2026): 10 Picks for Every Team
