Key takeaway: There is no universal list that covers every client, contract and AI service. But some information should trigger an immediate stop—or at least a much stricter review—before it is pasted into an AI tool. Credentials, payment details, highly sensitive personal information, confidential legal material, security secrets and client data that you are not authorized to disclose are good examples.
AI assistants make it very easy to move information.
Copy.
Paste.
Ask.
That convenience is exactly why client-data mistakes can happen so quickly.
You may start by asking AI to improve a paragraph and accidentally include an entire email thread below it.
You may upload a spreadsheet to analyze sales totals without noticing that another sheet contains customer names.
You may ask AI to troubleshoot software and paste a configuration file containing an API key.
The practical question is therefore not only:
“Which AI service am I using?”
It is also:
“What information should never have entered this prompt in the first place?”
Here are the categories that deserve the strongest caution.
1. Passwords and login credentials
Do not paste:
- passwords;
- one-time passwords;
- recovery codes;
- authentication tokens;
- private keys;
- session cookies;
- secret access links.
An AI model does not need your real password to explain why a login is failing.
Instead of:
Why doesn’t this login work? Username: john@example.com, Password: ClientPassword123
describe the problem without the credential:
A user can authenticate successfully on the website but receives an unauthorized error through the mobile app. What should I check?
If AI needs an example, use fictional credentials.
2. API keys and software secrets
Developers and technical consultants need to be particularly careful here.
Configuration files can contain:
- API keys;
- database passwords;
- OAuth credentials;
- private certificates;
- cloud access tokens;
- webhook secrets;
- encryption keys.
A developer may paste 100 lines of code because of one error on line 83 without realizing the secret is sitting on line 12.
Before asking AI to debug configuration or logs, search for strings such as:
`API_KEY`
`SECRET`
`TOKEN`
`PASSWORD`
`PRIVATE_KEY`
`AUTH`
Replace real values with placeholders:
`API_KEY=[REDACTED]`
If a real secret has accidentally been exposed somewhere it should not have been, simply deleting the chat should not be your only response. Follow the relevant security procedure, which may include revoking or rotating the credential.
3. Banking and payment information
Avoid sending unnecessary financial credentials or payment data such as:
- complete card details;
- card security codes;
- online banking credentials;
- payment authentication information;
- private banking documents;
- sensitive account-access information.
If you need AI to help categorize transactions, it may not need the account holder’s identity or complete account details.
For example:
Instead of uploading a raw bank statement, prepare:
| Date | Description | Amount |
|---|---|---|
| `[DATE]` | Software subscription | €49 |
| `[DATE]` | Client payment | €2,400 |
| `[DATE]` | Office expense | €125 |
Remove information that is irrelevant to the analysis.
4. Identity documents
Be extremely cautious with:
- passport scans;
- national ID cards;
- driver’s licences;
- residence permits;
- tax identifiers;
- social-security identifiers;
- identity-verification documents.
These documents are difficult to meaningfully sanitize while preserving the reason they exist.
If your task is:
“What information normally appears on an EU passport?”
you do not need a client’s actual passport.
Use a description or synthetic example instead.
5. Highly sensitive personal information
Some client projects involve information that deserves particularly careful handling.
Examples can include information concerning:
- health;
- medical treatment;
- disabilities;
- sexuality;
- religious beliefs;
- political beliefs;
- biometrics;
- racial or ethnic origin;
- criminal allegations or records;
- children;
- employee disciplinary matters.
Exactly how these categories are regulated depends on jurisdiction and context.
But from a practical client-work perspective, they should immediately trigger the question:
Does this information really need to leave the approved system at all?
Replacing:
“Maria”
with:
`[PERSON_1]`
does not solve the problem if the remaining description still says:
`[PERSON_1]`, the only employee with [specific medical condition] at the company’s [small location] office…
The content may remain identifiable.
6. Attorney-client or confidential legal material
Do not casually use public or personal AI tools to process:
- confidential legal advice;
- litigation strategy;
- privileged communications;
- unreleased settlement discussions;
- internal legal investigations.
The consequences here are not simply about AI training.
Disclosure to a third party can raise legal and professional questions that depend on the jurisdiction, relationship, terms and circumstances.
If legal confidentiality matters, use only an explicitly approved workflow.
AI can still help with abstract questions.
For example:
Create a checklist for organizing documents for a contractual dispute.
That does not require the confidential dispute file.
7. Security vulnerabilities and infrastructure secrets
Suppose you manage a client’s website or cloud environment.
You discover a security problem and want AI to help.
Avoid supplying unnecessary details such as:
- production credentials;
- private IP addresses where sensitive;
- secret endpoints;
- access tokens;
- unpatched vulnerability details tied to an identifiable client;
- firewall credentials;
- database connection strings;
- complete internal architecture diagrams.
Transform the problem into a technical pattern.
Instead of:
Here is Client X’s entire production configuration. Find how an attacker can get in.
you can often ask:
An application uses this fictionalized authentication architecture. What classes of access-control errors should I test for?
Provide the minimum technical detail required to solve the problem.
8. Confidential deals and unreleased business information
Client documents may contain information such as:
- acquisitions;
- restructuring;
- unreleased financial results;
- future pricing;
- layoffs;
- upcoming product launches;
- confidential bids;
- investment decisions;
- unreleased partnerships.
Removing a company name may not be enough.
Consider:
`[COMPANY]`, Finland’s only manufacturer of [rare product], plans to acquire its Swedish competitor in October.
Most of the identifying information is still there.
Unique commercial events can identify an organization even when explicit names have been removed.
9. Entire customer or employee databases
AI makes bulk analysis tempting.
You have a spreadsheet with 8,000 customers.
Why not upload it and ask:
“Find patterns?”
Because the AI may not need the identifiable records.
First decide what analysis you actually want.
Maybe you need:
- customers by country;
- average order value;
- monthly churn;
- support tickets by category.
You may be able to aggregate the data locally first:
| Customer group | Customers | Average order |
|---|---|---|
| Group A | 1,420 | €84 |
| Group B | 970 | €113 |
| Group C | 530 | €61 |
Now the AI can help identify patterns without receiving thousands of individual records.
This is an important principle:
Aggregate before uploading when individual-level information is unnecessary.
10. Information your client explicitly said not to share
This category overrides many others.
Perhaps the information looks harmless.
But the contract says:
Information may only be processed using systems approved by the client.
Or:
Project documentation may not be transferred to third-party services.
Then your personal judgment that the information is “not very sensitive” is beside the point.
Your obligation comes from the agreement.
Check:
- NDA;
- contract;
- project security requirements;
- client AI policy;
- data-processing terms;
- employer policy.
“AI didn’t train on it” is not a response to a contractual prohibition against sending it there.
11. Information you do not have authority to share
Client documents frequently contain information about third parties.
For example:
- a customer’s complaint;
- an employee CV;
- a subcontractor’s proposal;
- another company’s pricing;
- emails from external partners;
- interview transcripts.
Your client gave the document to you for a particular purpose.
That does not necessarily mean you have unlimited authority to redistribute every part of it to another service.
When reviewing a document, ask:
Whose information is actually in here?
This is particularly important with email chains.
The person who hired you may not be the only person whose data appears in the thread.
12. Data the AI simply does not need
This is the largest category of all.
Information does not have to be extremely sensitive to be unnecessary.
Suppose you want AI to improve this sentence:
We can deliver the revised report on Friday.
The AI does not need:
- the client’s full name;
- their address;
- €80,000 project budget;
- names of five employees;
- the entire contract;
- six months of email history.
You could ask:
Rewrite this politely: “We can deliver the revised report on Friday.”
The safest unnecessary information is information that never enters the system.
A useful traffic-light system
Before sending client information to AI, classify it.
| Level | Examples | Practical approach |
|---|---|---|
| 🟢 Low concern | Public website copy, public product descriptions, generic wording | Normal review may be sufficient |
| 🟡 Review first | Internal project notes, ordinary business emails, unpublished drafts | Minimize, redact and check permissions |
| 🟠 High sensitivity | Personal data, contracts, employee information, detailed financial data | Use only with an approved and justified workflow |
| 🔴 Keep out unless specifically authorized | Credentials, authentication secrets, restricted legal material, highly sensitive records, prohibited client data | Do not paste into an ordinary AI workflow |
This is not a legal classification system.
It is a practical pause mechanism.
The point is to stop treating all information as equally suitable for copy-and-paste.
“But my AI provider doesn’t train on my data”
That can be important.
It still does not answer all the relevant questions.
“No model training” does not automatically mean:
- no storage;
- no logging;
- no administrator access;
- no connected systems;
- no legal obligations;
- no contractual restrictions;
- no security risk;
- no unnecessary disclosure.
This is why account-level protection and data minimization should work together.
See:
Free vs Business AI Accounts: What Changes When You Use Client Data?
“What if I anonymize it?”
Often, that is useful.
But anonymization is not magic.
Removing obvious identifiers can greatly reduce exposure, especially when the AI does not need them.
However, a document may still identify someone through:
- role;
- location;
- dates;
- unusual events;
- financial amounts;
- project details;
- combinations of otherwise ordinary facts.
For a practical workflow, see:
How to Anonymize Client Documents Before Using AI
Three ways to use AI without sending the original data
When the source material feels too sensitive, you can often redesign the task.
1. Describe the problem
Instead of uploading the document:
A client rejected a proposal because the scope is unclear. Give me five questions I should use to clarify the scope.
2. Create a synthetic example
Replace the real situation with invented data that preserves the structure of the problem.
Instead of a real employee evaluation:
Create an example performance-review summary for a fictional employee who met three goals but missed a major deadline.
3. Extract only the necessary facts yourself
Read the sensitive source locally.
Then provide:
The contract contains three deliverables, two deadlines and one ambiguous acceptance criterion. Help me create questions for clarifying the acceptance criterion.
AI can help with the reasoning task without ever seeing the contract.
Before you paste: the 10-second test
Ask:
1. Is it secret?
Passwords, keys or security credentials should stop you immediately.
2. Is it personal?
Determine whether someone can be identified directly or indirectly.
3. Is it confidential?
Check commercial, legal and contractual restrictions.
4. Is it necessary?
Could the AI complete the task without it?
5. Is this the right environment?
Check the actual account and workspace—not simply the AI brand.
If the data fails one of those checks, change the workflow before pressing Enter.
The principle behind Client-Safe AI
Using AI safely with client work is not about memorizing a list of forbidden words.
It is about developing a habit:
Give the AI the minimum information needed to perform the task.
Sometimes that means removing a name.
Sometimes it means replacing an entire document with three sentences.
And sometimes it means deciding that the information should not go into an AI system at all.
Those are all valid AI workflows.
The most sophisticated use of AI is not always giving the model more context.
Sometimes it is knowing what context to keep out.
Continue the Client-Safe AI series
Start here: Can You Paste a Client Email Into ChatGPT? What to Check First
Hands-on guide: How to Anonymize Client Documents Before Using AI
Account guide: Free vs Business AI Accounts: What Changes When You Use Client Data?
Full review: AI Privacy Checklist: 15 Questions Before Uploading Client Data