---
title: Sensitive Data in AI, and Legal Says No? A Data Protection Guide for Businesses
lang: en
source: https://mindsprt.dev/en/knowledge/ai-data-compliance/
---

# Sensitive Data in AI, and Legal Says No? A Data Protection Guide for Businesses

*Print Knowledge · 9 min read · 2026-07-26*

> Customer personal data, quotations, brand color values, can these really not be put into ChatGPT? Legal's "no" has a point, but it does not mean AI tools are completely off-limits. This article explains, from a practical angle, which data is risky, which is safe, and how companies can build AI usage rules that legal teams can actually approve

**Quick answer:** Customer personal data, quotations, brand color values, can these really not be put into ChatGPT?

## When Legal Says "No," What Are They Actually Worried About?

I have been asked this question many times recently, mainly because design teams have started using AI tools for proofreading, color matching, and layout. Then one day, someone from legal, or the boss, walks over and says, "Customer data cannot be put into that."

Legal's concern is not unnecessary. The public version of ChatGPT, meaning the version you open directly in a browser and log into, has one key behavior: what you enter may, by default, be used by OpenAI to improve its models. This is not a secret. OpenAI's privacy policy states it clearly. Most people just have not read it carefully.

Where exactly is the problem? If you paste a quotation confirmation sheet containing a customer's company name, contact person's name, and purchased items into ChatGPT and ask it to polish the wording, that data has been sent to OpenAI's servers. From a legal data-control perspective, it is no longer only in your hands.

Taiwan's Personal Data Protection Act requires companies to have a specific purpose when collecting, processing, or using personal data, and they may not go beyond the necessary scope. Sending customer personal data to a third-party AI platform, strictly speaking, crosses that line unless you informed the customer and obtained consent when collecting the data. In B2B printing sales, almost nobody does this.

That is the line legal is drawing when they say "no." They are not trying to make your life harder.

## What Is the Difference Between Public AI and Enterprise AI?

Banning AI across the board is the other extreme, and it is not necessary. The real point is knowing what kind of AI you are using, because the level of data separation can be completely different.

Public version, Consumer tier

・ChatGPT Free and Plus personal accounts: input data enters the training pool by default, unless you manually go into settings and turn off the "improve the model" option.

・Suitable for personal use and creative ideation that does not contain customer-identifying information.

Enterprise and API versions, Enterprise / API tier

・ChatGPT Enterprise and Microsoft 365 Copilot through Azure OpenAI: OpenAI clearly states that it does not train models on Enterprise customer data, and that data is encrypted both in transit and at rest.

・API calls, when the user parameter is set and the account is confirmed as an API plan, are also covered by the no-training data policy.

Local deployment

・Running open-source models, such as the Llama series, in your own server room or private cloud: data never leaves your environment, so compliance risk is lowest.

・The downside is IT setup cost. This is more suitable for printing plants or design companies of a certain scale.

Put plainly: even when both are called "AI," the fate of your data is very different between public and enterprise versions. Many companies say "AI cannot be used." What they should really say is, "Public AI cannot be used to process customer data." Those two sentences mean very different things.

## Variable Data Printing Is Far More Sensitive Than Most People Think

Variable Data Printing, VDP, is the scenario in the printing industry that needs the most caution.

In simple terms, VDP means that within the same batch of printed materials, each piece prints a different name, address, barcode, or personalized content. Direct-mail envelopes, membership cards, insurance policies, and medical report covers are all typical VDP applications. The nature of this work is clear: you will definitely have a large list of personal data in your hands.

I have seen a very common situation: a designer wants to check whether the layout of a VDP template is correct, so they paste an entire CSV containing 500 customer records into an AI tool and ask it to "check the field format." Just like that, the names, addresses, and phone numbers of those 500 records are sent out.

For VDP jobs, my recommendation is to use de-identified dummy data for template testing as a rule. From the moment the real customer list enters the workflow, it should run only inside an environment with information-security controls and must not touch public AI tools.

If you need AI to help improve the design logic or layout rules for VDP, turn the rules into text, such as "the name field allows up to 8 Chinese characters, and if it exceeds that length, reduce the font size to 8pt." Send the rules into AI, not the real data along with them.

## Which Data Can Be Used, and Which Must Be De-identified First?

This is the question people ask most often in real work. A simple classification looks like this:

Safe to put directly into AI, low risk

・Public information, such as product specifications, material descriptions, and printing process explanations.

・Completely fictional design-copy text, such as Lorem Ipsum for testing or self-made dummy data.

・Internal process documents that contain no identifying information about customers or employees.

・Technical questions, such as "What is the standard CMYK value for Pantone 485C?"

Must be de-identified first

・Customer requirement documents: replace company names with "Customer A" and personal names with "procurement contact."

・Past case references: blur all specific identifying details before describing the case.

・Quotation drafts: items and specifications may stay, but the customer name, tax ID, and purchaser name must be removed.

Never put into AI, high risk, regardless of AI type

・Any list containing a real person's name, phone number, address, or national ID number.

・Original design files or brand guideline documents covered by a confidentiality agreement with a customer.

・Financial data, contract terms, and quotation details, especially for competitive bids.

・Employee salary, performance, or HR data.

There is only one core principle for de-identification: after removal, no one should be able to work backward from the remaining data to a real person or a specific company. Replacing only the name while keeping a distinctive business description does not count as true de-identification.

If you need AI help preparing a print proposal, consider discussing with the [MINDS Knowledge Academy consulting team](https://mindsprt.dev) how to build an AI workflow that meets compliance needs, especially data classification rules for VDP and commercial printing scenarios.

## What Clauses Should You Check in an NDA With an AI Vendor?

Many companies look only at features and pricing when adopting AI tools, then skim through the confidentiality agreement. This is a very common mistake.

A trustworthy AI vendor NDA or terms of service should, at minimum, confirm the following points:

Data use

・Does the vendor clearly promise that input data will not be used to train models?

・Is there a clear data retention period, for example deletion within 30 days after processing is complete?

・Will the data be used only for the purposes specified in the contract, with no disclosure to third parties?

Data storage and separation

・Confirm server location: the EU's GDPR and Taiwan's Personal Data Protection Act both have rules on cross-border transfer. Data stored on U.S. servers is not always compliant.

・Does the Enterprise account provide logical separation to ensure your data is fully separated from other tenants?

Liability for breach

・If the vendor has a data leak, it must notify you within a set number of hours. 72 hours is a common industry standard.

・Are the compensation terms for breach specific, rather than only saying the vendor will "make best efforts" to prevent issues?

Audit rights

・Are you allowed to request data processing records from the vendor?

・ISO 27001 or SOC 2 certification is a basic threshold. It shows the vendor has passed a third-party information-security audit.

If the vendor's terms of service are a public online document and cannot be customized for you, then the only thing you can do is confirm whether the platform's usage policy meets your needs. If it does not, you are handling customer data in an untrustworthy way.

Major printing-industry SaaS tools, such as Adobe Creative Cloud and Canva Enterprise, have issued clearer GDPR and personal-data compliance statements in recent years. But those statements are not the same as an NDA. You only have real protection after signing an enterprise contract.

## A Compliance Checklist Before Bringing AI Into the Company

This is the checklist I use when helping printing plants and design companies introduce AI workflows. It is not a theoretical framework. It comes from actual implementation work:

Before adoption

・Confirm the version of the AI platform's terms of use and its explanation of data use.

・Assess whether you need an enterprise version, API version, or local deployment, based on data sensitivity.

・Confirm with legal whether existing customer contracts contain clauses such as "customer data may not be given to a third party for processing."

Build internal rules

・Create an "AI-allowed data classification table," based on the classification logic above.

・Specify which workflows may use AI and which are prohibited, such as banning public AI throughout the entire VDP process.

・Set up employee training on AI use, at least once a year.

Vendor management

・Require every AI tool vendor to provide a DPA, Data Processing Agreement.

・Confirm that the vendor has ISO 27001 or SOC 2 certification.

・After signing, review the vendor's privacy policy updates every year.

Daily operations

・Before pasting content into AI, run a "de-identification self-check": does this content include information that can identify a specific person or company?

・If there is any doubt, replace the sensitive parts before sending it. Do not gamble.

・After a sensitive task is complete, clear the conversation history on the AI platform. Most platforms provide a manual delete option.

If your company handles print jobs that require high personal-data protection, such as medical, financial, or government bid projects, consider working directly with a commercial printing provider with a full compliance process, such as [MINDS Printing](https://www.mindscmyk.com/). That keeps everything from design to final layout inside a controlled environment and reduces the chance of information-security gaps.

## Key Takeaways

・Public ChatGPT may, by default, use your input data to train models. Enterprise and API versions clearly promise not to do so. Their compliance risks are not in the same league.

・In Variable Data Printing, VDP, real customer lists should be fully separated from AI tools from the moment they enter the workflow. Use only dummy data for template testing.

・The bottom line of de-identification is this: after removal, the remaining data cannot be traced back to any real person or company. Replacing names while keeping distinctive business descriptions does not count.

・When signing with an AI vendor, check four things: data use, server location, notification obligations, and audit rights. Without a DPA, you do not have real protection.

・Internal AI usage rules cannot stay as talk. A rule such as "public AI is banned for VDP" has to be written into the workflow SOP before it counts as implemented.

## Further Thinking

When it comes to compliance, the printing industry actually has an advantage over most sectors, because we are already used to handling confidential customer materials: keeping design drafts confidential, preventing proof copies from leaking, and keeping quotations private. These are basic professional habits in the industry. Extending the same logic to AI tools is not hard.

The most practical next step is to do one thing: list every AI tool currently used in the company, then check one by one whether it is an enterprise version and whether a DPA has been signed. If any tool is being used through an employee's personal account, switch it to a company account now, or upgrade directly to Enterprise. This move is not complicated, but it can cut off 80% of the risk.

Design companies and printing plants do not need to wait until something goes wrong before acting on AI compliance. Building a data classification table now and teaching employees what can and cannot be pasted into AI is far cheaper than fixing the damage afterward.

## FAQ

### What are the specific data-security differences between the free version of ChatGPT and the enterprise version?

Free and Plus personal accounts allow OpenAI, by default, to use conversations to improve models, meaning the data may enter the training pool. ChatGPT Enterprise clearly promises not to train models on customer input data, and data is encrypted during transmission and storage. Simply put, in the personal version, your input may be seen and learned from by OpenAI. In the enterprise version, it will not.

### Can customer lists for Variable Data Printing, VDP, be processed by AI tools?

No. VDP customer lists usually contain personal data such as names, addresses, and phone numbers. Putting them directly into public AI tools violates the "specific purpose" principle under personal data protection rules. The right approach is to use dummy data for template testing, while real lists are processed only locally or inside a sealed, fully controlled workflow environment with information-security controls.

### Is replacing names enough for de-identification?

No. The standard for de-identification is this: after removal, no one can use the remaining information to trace it back to a real person or specific company. If you replace the name but keep a distinctive business description, such as "the largest manufacturer of a certain type in northern Taiwan," people in the industry can still identify it. That is not true de-identification.

### Which document matters most in a service contract with an AI vendor?

The most important document is the DPA, Data Processing Agreement. It clearly states how the vendor processes, stores, and protects your data, as well as notification obligations and compensation terms if a leak occurs. If you only have a service contract or online terms, but no DPA, you effectively have no legal protection.

### Is a locally deployed AI model always safer than a cloud service?

From a data privacy perspective, local deployment is indeed the safest option, because the data never leaves your environment and there are no third-party server compliance issues. But security still depends on your own IT management. If your internal servers lack proper access control and encryption, local deployment still carries risk. A cloud enterprise version, such as Azure OpenAI Service, may be more controllable in practice for small and midsize companies if it comes with a proper DPA and security certifications.


---

> HTML version: https://mindsprt.dev/en/knowledge/ai-data-compliance/
> MINDS — 麥思印刷整合有限公司 · https://mindsprt.dev