Protecting customer data when using AI

From the brief stage, Sinh Vũ is already exposed to sensitive client data: names, contact numbers, contracts, unpublished strategies. When AI begins to participate in the workflow, the question is not "Is AI useful?" but rather "What data is fed into AI, and what data is absolutely off-limits?" The answer to that determines your credibility with clients, just as much as the quality of the product.

Data classification: the first step that cannot be overlooked.

Not every piece of data requires the same level of protection. It is necessary to clearly distinguish four layers:

• Public data: information that has been published externally, with low risk when inputting into AI.
• Internal data: processes, templates, internal documents that have not been shared externally. More caution is needed.
• Confidential data: contracts, strategies, financial reports. Do not input into public AI.
• Personal information (PII, meaning personally identifiable information): name, phone number, email, address, customer payment data. Highest level of protection, requiring clear legal basis each time it is processed.

The most common mistake is treating all data as a single block, then making decisions based on convenience rather than risk levels.

Choose the right tool for the type of work.

What needs to be checked before using any tool: does this tool have a DPA, will the data I input be used for training, and who has access to that data?

Minimization: core principle

Data minimization means only including the necessary parts for specific tasks in AI, not the entire record "just in case." This principle comes from international data protection standards and has clear practical reasons: data not included cannot be exposed.

• Before inputting data into AI, ask: what data fields does this task really need?
• Blur or replace names, phone numbers, and addresses before submission if the task does not require identifiable information.
• Schedule deletion: delete data that is no longer needed, do not keep it "just in case."

Don't process personal data just because it might be useful later. Each processing needs a legal basis.

ICO, Data minimisation (AI and data protection toolkit)

Common mistakes and how to recognize them

• Directly inputting client lists into free AI: a familiar action but carries the risk of data exposure. It’s not just because AI "repeats" immediately, but because there is no control over where that data goes afterward.
• Using personal AI accounts for company work: when the account is beyond the organization's control, no one can check or revoke access when an employee leaves.
• Thinking anonymity is enough: Many pieces of data combined can still identify individuals even if each piece has had names removed. Anonymization needs to be accompanied by minimization.
• Unaware of the data policy of the tools being used: this is the most common gap. Using without reading the terms means accepting conditions without knowing what you are agreeing to.

The viewpoint of Sinh Vũ

Data discipline is a condition of credibility, not an additional burden. Customers trust you with sensitive information because they believe in how you operate, not because they are sure you won't be careless.

Sinh Vũ operates along three clear lines: confidential data and personal information of clients do not enter public AI tools. When AI support is needed, priority is given to environments committed to data protection, and only minimal, anonymized data is used. Any decision involving personal data requires approval. When issues exceed our scope, we refer them to security or legal experts instead of handling them ourselves.

You don’t need to be a technical expert to practice this. Just have a clear internal policy: which data is allowed into which AI, who is allowed to use it, and when in doubt, ask before proceeding.