The redaction habit: what financial advisers should remove from client details before pasting anything into AI

The Redaction Habit: What to Remove Before You Paste Anything into AI

Table of Contents

To redact client data before using AI, remove anything that could point to a real person: names, NRIC numbers, contact details, account and policy numbers, exact dates and addresses, and any detail rare enough to identify someone on its own. Then paste only what the task needs, into a tool your firm allows.

It takes about a minute. It is also the habit that lets you use AI for client work without handing over your client.

Why this habit matters

AI tools make it very easy to paste first and think later. A client sends a long email, you want a clear reply, and the quickest route is to copy the whole thread into a chat window.

Many people do exactly that. In a TELUS Digital survey of 1,000 US employees at large companies who use AI at work, run in January 2025, 57% admitted entering sensitive information into public AI assistants. 31% said they had entered personal data, and 21% customer information. 68% said they used these tools through personal accounts rather than ones their employer provided.

The same survey found that only 24% said their company required AI training, and 44% said their company had no AI guidelines or that they didn’t know whether it did. The habit is easy to lose when nobody has spelled it out.

Advisers hold some of the most sensitive information people share with anyone: income, health, family, debts and plans. So the habit matters more for us, not less.

What counts as personal data, in plain words

Singapore’s Personal Data Protection Act (PDPA) is the law on how organisations collect, use, share and look after personal data. The Personal Data Protection Commission (PDPC), which oversees it, defines personal data as data about a person who can be identified from that data, or from that data together with other information the organisation has or is likely to have access to.

The second half of that definition is the one people miss. A name is obviously personal data. But so is “a 47-year-old dentist in Bukit Timah with three children and a recent cancer diagnosis”, because those details together can point to one person, even with no name attached.

Two more terms you’ll see:

  • Redaction: removing or covering details in a document before you share it. Here, it means taking identifying details out before text goes into an AI tool.
  • Anonymised data: data that can no longer be linked back to a person, even when combined with other information. Swapping a name for “Client A” is a good start, but it does not make text anonymised if the rest of the details still give the person away.

Your firm’s policy comes first

The PDPA places its duties on organisations. For most advisers, that means the firm you represent decides which AI tools you may use with client information, and how. If your firm says no client information goes into any AI tool, that is the answer, and nothing below changes it.

Three of the PDPA’s obligations, as the PDPC describes them, explain why firms take this seriously:

  • Purpose limitation: personal data may be used only for purposes a reasonable person would consider appropriate in the circumstances, and that the person has consented to. Your client shared their details for advice, not for testing a chatbot.
  • Protection: organisations must make reasonable security arrangements to prevent unauthorised access, use or disclosure of the personal data they hold.
  • Transfer limitation: personal data may be transferred out of Singapore only in line with the regulations, so that it keeps a comparable standard of protection. Many AI tools process data on servers overseas.

A firm-approved tool, set up under an agreement with the provider, is a very different thing from a free personal account. Redaction is still worth doing in both. It is the layer you control, whatever tool you use.

The redaction checklist

Before you paste, go through the text once for each line:

  • Names. The client, their spouse, children, parents, employer and anyone else mentioned. Check the email signature and the “Dear…” line too.
  • Identity numbers. NRIC, FIN and passport numbers. These should never go into an AI tool, redacted text or not.
  • Account and policy numbers. Bank accounts, policy and plan numbers, application references and claim numbers.
  • Contact details. Phone numbers, email addresses, home and office addresses, and social media handles.
  • Exact dates. A date of birth or a wedding date can identify someone. Use an age or a year instead: “early 40s” rather than “born 3 March 1984”.
  • Places. Swap a street or a condo name for a general area, or drop it if the task doesn’t need it.
  • Rare details. An unusual job, a well-known employer, a rare illness or a public role can identify a person on their own. Make them general: “a senior manager”, “a health condition”.
  • Health information. Leave it out unless the task truly needs it and your firm allows it. If it must stay, keep it general.
  • Exact amounts. Round them. “About $8,000 a month” does the job of “$8,137.42”.
  • The hidden extras. File names, document headers, earlier messages in a forwarded thread, and text inside screenshots.

Swap, don’t just delete

Deleting details can leave text that makes no sense. Swapping keeps it readable and keeps the AI’s answer useful.

Use clear placeholders in square brackets: [Client A], [Spouse], [Child 1], [Employer], [City]. The brackets also make it easy to find each one later, when you put the real details back in your own reply, outside the AI tool.

Keep the link between the placeholder and the real person in your own records, never in the chat.

A before-and-after example

Here is how it might look. The client and every detail are invented.

Before:

Hi, it’s Rachel Tan (S1234567D). Daniel and I just found out baby number two is due in May! I’m a senior physiotherapist at a hospital in Novena and take home $8,137 a month. Our joint account at the bank ends 4471. Can we meet before I go on maternity leave? My mobile is 9123 4567.

After:

Hi, it’s [Client A]. [Spouse] and I just found out our second child is due next year. I work in healthcare and take home about $8,000 a month. Can we meet before I go on maternity leave?

The task was “help me write a warm reply that suggests a review meeting”. The “after” version keeps everything that task needs: the news, the timing and the request to meet. It drops the name, the NRIC, the account digits, the phone number, the exact salary and the workplace, none of which the AI needed.

Notice what else changed. “A senior physiotherapist at a hospital in Novena” is specific enough to narrow things down a lot. “Works in healthcare” isn’t.

When redaction isn’t enough

Some material should not go into a general AI tool at all, however carefully you edit it:

  • Scanned documents and photos, such as an NRIC, a medical report or a policy schedule. Text inside an image is easy to miss.
  • A full fact-find or a complete client file. There is too much in it to redact reliably.
  • Recordings of client meetings. That is a separate question, with consent and storage issues of its own.
  • Anything your firm has told you to keep out of AI tools.

If a task needs the whole file to make sense, it probably belongs in your firm’s own systems, not a chat window.

Two prompts worth saving

To keep the placeholders in the answer. Paste this before your redacted text, so the AI doesn’t invent names or details:

The text below has been edited to remove personal details. Keep every placeholder in square brackets exactly as written, and do not make up names, dates or numbers. If you need a detail that has been removed, ask me instead of guessing.

A second pair of eyes. Once you’ve redacted the text yourself, add this to the end of your prompt:

Before you do the task, list anything in my text that could still help identify a real person, on its own or combined with other details.

Please note that the second prompt is a check on work you’ve already done, not a replacement for it. The text has reached the tool by the time it answers, so the first pass has to be yours.

Frequently asked questions

Can financial advisers put client information into ChatGPT?

Only if your firm allows it, and then only what the task needs, with identifying details removed. Your firm’s policy on AI tools comes first. Even with a firm-approved tool, redacting names, numbers and other identifying details is a good habit.

Is removing the client’s name enough?

Usually not. Under the PDPA, data counts as personal data if a person can be identified from it, alone or combined with other information. Age, job, area, family and health details together can point to one person even without a name.

Does the PDPA apply when I use AI tools?

The PDPA covers how organisations collect, use, disclose and protect personal data, whatever the tool. Its obligations, such as using data only for appropriate purposes, protecting it and limiting transfers overseas, apply to client data that goes into an AI tool just as they do anywhere else. Your firm’s policy explains how it applies them.

Is it safer to use a paid or business version of an AI tool?

Business versions often come with agreements on how your data is stored and used, which free personal accounts may not have. That can make a real difference, but it is your firm’s decision which tools are approved. Redaction is still worth doing in any tool.

What if I’ve already pasted client details into an AI tool?

Tell your firm, following its process for data incidents. Don’t try to fix it quietly. Under the PDPA, organisations must assess whether a data breach needs to be reported, and that assessment is your firm’s to make.

Related reading

Sources

Full Disclosure

The Content above is generated by AI. The objective of the #AiSeries is for me to test if I can generate website traffic using just AI-generated content. However, having that said, even I, have learned a lot from the content above.