What data should you keep out of AI tools at work?

An export file about to be pasted into a chat box

At the end of the month, your manager asks you to find out what customers of Nhà Mầm, a fictional online shop selling baby goods, complain about most. You export 200 complaint tickets from the support system. Each row has the customer's full name, phone number, email, order ID, delivery address, amount and complaint text.

Reading all 200 rows would take a whole morning. An AI tool could group the themes in a few minutes. You have already selected the whole file and are about to paste it in. While you are at it, you also plan to paste a clause from the contract with the delivery partner, to ask who is responsible for late deliveries.

This is an illustrative situation. There is nothing wrong with the intent: you want to work fast and well. The problem is what comes along with the file. This article helps you separate what should stay out of AI tools, how to clean a piece of data before using it, and when to stop and ask the person who is responsible.

• Check AI-drafted documents before sending them to the team

Why is sensitive data hard to spot?

Sensitive data rarely sits in one obvious place. It is mixed into what a PO uses every day: export files, interview notes, screenshots, chats with developers, meeting notes. A few things make it easy to miss.

• Pasting a whole block is faster than filtering line by line. Under pressure, cleaning is the first step to go.
• The data looks harmless. An order ID or a phone number alone says little, but combined with a name and an address it points to exactly one person.
• Removing the name may not be enough. NIST, the US standards institute, lists data privacy among generative AI risks, including leaking data and de-anonymising people whose data was meant to be anonymous.
• The company policy is unclear, or you have not read it. Many teams have only just started using AI and have not written guidance yet.
• Company information is sensitive too. New POs tend to think only of customer data and forget revenue figures, unreleased plans or partner terms.

OWASP, a software security community, includes “Sensitive Information Disclosure” in its top risks for applications built on large language models. The kinds of information it names include personal identifiable information, financial details, health records, confidential business data, security credentials, and legal documents. OWASP also says users should know how to interact safely and avoid entering sensitive data.

In Vietnam, the Law on Personal Data Protection No. 91/2025/QH15 has been in effect since 1 January 2026. Put briefly, it sets out the rights of people whose data is processed and the duties of organisations that process personal data, and it bans acts such as buying and selling personal data. This is not legal advice. Whether customer data can go into an outside tool is a question for your company's legal team or data owner, not one a PO should settle alone.

What data should you not put into AI tools?

The table groups six kinds of data a PO often touches. “Should not” here means not without clear permission from the company. If your company has an approved tool and a specific policy for that kind of data, follow the policy.

Look at the “Safer alternative” column. In almost every case you can still do the job, just with less data. To know what customers complain about, you need the complaint text, not the phone number. To ask what a log error means, you need the error message, not the token.

Four checks before you paste

One approach is to go through the same four steps every time, whether the data is long or short. For a short piece, they take a minute or two.

Step 1 is easy to skip. Ask your manager or IT three things: which AI tools the company allows, with which accounts, and for which kinds of data. Do not infer it from what colleagues use, and do not rely on what you have heard about how a vendor stores or uses data. That changes over time and between plans. The right source is your company's policy and the current terms the company has reviewed.

Step 2 starts with a question: what do I need the tool to help with? Grouping complaint themes needs the complaint text. Rewriting an announcement needs the draft announcement. Every column and row that does not serve that task stays in the company's systems.

Step 3 cleans what is left. Replace names with generic labels, remove phone numbers, emails, order IDs and exact addresses, round amounts, drop staff names. If you do not need real data at all, a fictional example is often the safest choice, especially when you only want help with layout or structure.

How do you know if a piece of data can go into AI?

Use the filter below for each specific piece. A “yes” moves you on. A “no” means doing the side action first, then coming back.

The last two questions split company information into two treatments. Unreleased figures and plans can often be generalised and still carry the point, for example “revenue rose slightly compared with last month”. Contracts and HR information should not go in in any form without permission. If your question is “what does this clause mean?”, the answer should come from legal or whoever negotiated the contract, not from an AI tool.

How do you anonymise customer data?

Back to Nhà Mầm's complaint file. Below is one illustrative ticket; every name, number and address is made up. The task is grouping complaint themes, so only the problem itself is really needed.

After cleaning, the ticket still shows three problems: late delivery against the promise, damaged goods and a hard-to-reach hotline. That is everything theme grouping needs. The cleaned version is hard to use to work out who Customer A is.

Watch out for rare details. If the text says “I'm the only customer in Đà Lạt who bought three cots last week”, someone inside the company could guess who it is even with the name removed. When you see a detail like that, generalise it to “a customer who bought several items at once”.

If you later need to reach the right customer, for example to apologise, keep the mapping between the Customer A label and the order ID inside the company's systems. That mapping never goes into the AI tool. The same approach applies to user interview notes and to summarising customer feedback.

• Using AI to summarise customer feedback
• What to ask in a user interview

A checklist before pasting into an AI tool

Use this checklist each time you are about to put work data into an AI tool. Copy it into your notes or pin it in your team's chat channel.

Before putting this piece into an AI tool:
[ ] The tool and account I use are on the company's approved list.
[ ] This kind of data is allowed with that tool under the current policy.
[ ] I can say in one sentence what I need the tool to help with.
[ ] Only the part that serves that task is left.
[ ] Names are replaced with generic labels; phone numbers, emails, order IDs and exact addresses are removed.
[ ] Passwords, keys, tokens, internal links, source code and security details are removed.
[ ] Unreleased figures are rounded, turned into trends, or used only where permitted.
[ ] No contract terms, partner information or HR information.
[ ] No rare detail that would let someone guess who it is.
[ ] Screenshots, if any, are cropped or masked to what is needed.
[ ] Where I was unsure, I asked my manager, IT or legal.

The first two lines need answering once per tool, then again when the policy changes. The rest apply to each piece. The screenshot line is easy to forget: a screenshot can include a sidebar with other customers' names, a search box still holding a phone number, or a browser tab showing an internal link.

Common mistakes when using AI with company data

• Pasting the whole export. The phone, email and address columns go along, while the task only needs the complaint text.
• Thinking that removing the name is enough. An order ID, an address or a rare detail can still point to the person.
• Using a personal account for company work. A company-approved tool and one you use privately may be under different terms. Ask IT first.
• Guessing the vendor's policy. Do not rely on articles or rumours about how a tool stores data. Follow your company's policy and the current terms it has reviewed.
• Pasting a key or password by accident and staying quiet. Tell the system owner straight away so it can be rotated. The sooner you report it, the shorter the exposure.
• Asking AI to interpret a contract instead of legal. An answer that sounds certain is not necessarily right for your contract.

Summary

• Six kinds of data to hold back without clear permission: customer personal data, unreleased figures and plans, contracts, passwords and keys, source code and security details, HR information.
• Check the policy and approved tools first, then each piece of data.
• Take only what the task needs, replace names, remove identifiers, round numbers, drop rare details.
• If you do not need real data, use a fictional example.
• When unsure, ask your manager, IT or legal. If you pasted a key or password, report it straight away.

I am writing more PO situations in this same format, with templates you can use straight away. If you would like the next one by email, subscribe to the newsletter below.

• Get new articles by email

Frequently asked questions

Can I put customer data into an AI tool?

This article recommends doing so only when your company has a policy allowing it and the tool is approved for that kind of data. If unsure, ask your manager, IT or legal first; even when it is allowed, remove names, phone numbers, emails, order IDs and addresses the task does not need.

What does anonymising data before using AI mean?

Replacing people's names with generic labels, removing identifiers such as phone numbers or order IDs, rounding amounts, and dropping rare details that could reveal who it is. Keep only what the task actually needs.

What should I do if I pasted a password or API key into an AI tool?

Tell the system owner or IT straight away so they can rotate that key or password. Do not stay quiet or decide on your own that it is fine.

Sources to check

• NIST AI 600-1: Generative AI Profile
• OWASP: LLM02:2025 Sensitive Information Disclosure
• Cổng thông tin Hà Tĩnh: Luật Bảo vệ dữ liệu cá nhân có hiệu lực từ 01/01/2026

Related articles

• Using Claude as a PO: where it helps, what you still do: AI drafts fast. This article shows four mistakes to check and a checklist to use before a document reaches the team.
• Can AI replace a Product Owner? What still needs a PO: A balanced answer for new POs: AI helps with drafting, summarising and suggesting, while accountability for product value and the decisions behind it stays with a person.
• Using AI to summarise customer feedback: what to check: AI tools group feedback and name themes fast, but they can invent quotes, lose frequency and drop small groups of customers. This article shows each failure mode and how a PO checks a summary before using it for a decision.
• Numbers from AI analysis: what to check before reporting: How a PO checks numbers an AI tool calculated: reconcile totals, check the denominator and the period, recompute a few key figures, and label what is still uncertain.