‹ Back to Home

How to Prepare Sample Data Before Sending Anything to AI

Test an AI tool without touching real customer records. A seven-step exercise for building made-up test data, with OAIC guidance and what India's DPDP Act means for it.

Keerthika 8 min read
Follow on Google
Privacy Tips How to Prepare Sample Data Before Sending Anything to AI 8 min left Follow on Google
How to Prepare Sample Data Before Sending Anything to AI
0:00
0:00
🔒 Listen is for subscribers. Subscribe

Say you want to try sorting a small business's enquiry emails with the help of AI. Grabbing the last ten emails from real customers might look like the quickest way to start. But those messages may contain more than names, addresses and phone numbers: health conditions, money trouble or family details can be in there too. The first question to ask is whether you need to send any of that out just for a demo.

There is another way to run an early test: leave real people's records alone and write made-up records in the format you need. What follows is a small exercise for doing exactly that. It is not a customer study, a performance result, or a privacy certificate for any AI service.

1. Write down the question your test should answer

"Does the AI work well?" is far too broad. Instead write: "Can it sort enquiries into four groups: new booking, change, complaint, unclear?" Decide the output you expect as well: the group name, one sentence of reasoning, and a flag for whether a human should look at it.

You don't need anyone's name or phone number for this. Deciding up front what information the test actually needs stops you from adding fields you don't need. Test only the sorting at first. Keep actions such as replying to the customer or changing a booking out of this exercise.

2. Start from a blank sheet

Instead of copying your customer list and changing the names, open a new file. Three fields are enough: record ID, message, and expected group. On the very first line, write "Entirely made-up test data; not real customer records."

If the test needs a name, use an obvious placeholder such as "Test Person A". If you don't need a phone field, delete it. If you do, use a non-working placeholder that your isolated test setup accepts, rather than inventing a number someone could actually be reached on. Separately, check that integrations such as email sending are switched off.

3. Don't write only easy messages

Ten made-up records make a reasonable starter set. That is only a suggested size for a small exercise, not a defined size for proper quality testing. Let each record test a different situation.

For example:

  • Record TEST-01: "I need to book a new slot for next week." Expected: new booking.
  • Record TEST-02: "Can I move the time I already booked?" Expected: change.
  • Record TEST-03: "I've asked three times with no reply; I want to speak to someone in charge." Expected: complaint; human review needed.
  • Record TEST-04: "Change that." Expected: unclear; ask for more detail.

In the rest, add spelling mistakes, Tamil-English mixing, a message with two requests, an empty message and so on. But don't paste in a real complaint with small changes. Write fresh only the problem you want to test.

Plan for different formats too. Write the same request as one line and as a long message. Leave the needed detail out of one record. In another, put a complaint alongside a booking question. Decide in advance whether to force both into one group or allow several. That is part of the test design too; don't accept whatever the AI picks as the correct definition.

4. Removing the name is not enough

Concluding "no name, so no personal information" is a mistake. Someone's job, a small town, a specific date and an unusual incident, taken together, can help identify them. So don't blank out the name and number in a real message and then treat it as fully safe.

Australia's privacy regulator, the OAIC, explains that removing direct identifiers alone is not enough for de-identification. The context in which the information is used and the chance of re-identification both matter [1]. That is why, in this early exercise, we write fresh on a blank sheet rather than editing old records.

5. Read the free text separately

Even when the table fields are clean, details can hide inside the message. Read sections such as "notes", "complaint details" and "attached message" in full. Don't forget signatures, chains of old replies, file names, attachments and screenshots.

Ask someone else to check whether any real identity slipped into the practice file by mistake. A search can spot email addresses or long numbers, but don't assume an automated search will catch everything. Make a human read of the whole file part of this exercise.

6. Check the service before you send the file

Having made-up data ready doesn't mean you can send anything to any service. Check the service's current terms: whether your account is allowed for business use, how long records are kept, who can access them, and whether they are used for training. Don't fill gaps you don't understand with guesses.

As a matter of best practice, the OAIC recommends not entering personal information, and especially sensitive information, into publicly available generative AI tools [2]. This is Australian guidance; don't treat it as the same legal answer for every country. At the later stage where you need real data, you will need a privacy and security assessment that fits your organisation.

What about India?

In India, the Digital Personal Data Protection Act, 2023 (DPDP Act) is the core law here. An organisation handling personal data (a Data Fiduciary) has to use it only for the purpose it was collected for, keep only what it needs, and protect it with security safeguards. Pasting an email a customer sent for one service into an AI tool for testing may well fall outside that purpose.

The DPDP Rules, 2025 were notified on 13 November 2025 and are being phased in over 18 months, with the main obligations on organisations due by May 2027 [3]. There's no need to wait for that: starting with made-up data is a safe first step for Indian businesses too. This is a general explainer only; check your organisation's legal position with an expert.

7. Record the differences, not just the answers

For each made-up record, write the group the AI gave next to the group you expected. If it got it wrong, note "What was confusing about this record?" After you change your instructions, run the same set again. Keep the date, the instructions used and the file version.

Did it treat a complaint as an ordinary enquiry? Did it decide on an unclear request instead of asking for more detail? Notes like these help design the next test. But being right on ten made-up records doesn't mean you'll get the same result with real customers.

Write a stop rule before you test, too. For example, you might decide that if a reply auto-closes a complaint that needed a human, you won't move to the next stage until that error is fixed. That's your decision for this exercise, not a safety standard for every organisation. Likewise, keeping the made-up data and the expected answers in separate files helps you avoid accidentally feeding the answers in with the input during testing.

Think about what someone else would need to repeat this exercise. A file name alone isn't enough. They need to know which version, which instructions, which group names and which results you checked. You can write that record without any personal information. Give a new version number every time you change something. The point is to compare the two results, not to pass off an old test as a new one.

A clear line before the next stage

Don't treat this fully hand-written made-up set as equivalent to every kind of synthetic data generated from real data. Information generated from a real dataset can carry different privacy risks and needs its own assessment. The word "synthetic" is not a safety guarantee in itself.

The aim of the early exercise is modest: to sharpen your question, your groups and your evaluation method without sharing real records. Mark the file "for testing only". Fix the expected answers in advance. Log the mistakes. Only then decide, separately, whether a next stage with real data is needed.

Sources

[1] OAIC, De-identification and the Privacy Act. Guidance on why removing direct identifiers alone falls short, and on context-based risk assessment.

https://www.oaic.gov.au/privacy/privacy-guidance-for-organisations-and-government-agencies/handling-personal-information/de-identification-and-the-privacy-act

[2] OAIC, Guidance on privacy and the use of commercially available AI products. Best-practice recommendation on entering personal information into publicly available generative AI tools.

https://www.oaic.gov.au/privacy/privacy-guidance-for-organisations-and-government-agencies/guidance-on-privacy-and-the-use-of-commercially-available-ai-products

[3] PIB (Government of India), Government notifies DPDP Rules to empower citizens and protect privacy, 14 November 2025. (India section added by the TamilTech editorial team.)

https://www.pib.gov.in/PressReleasePage.aspx?PRID=2190014

Sources checked on 18 November 2026. The steps here are a practice exercise suggested by the author, not a testing protocol set by the OAIC. This is not legal advice.

About the author

An educational article prepared on behalf of IT Impress. Written with AI assistance and checked against the cited sources. More about IT Impress: itimpress.au. Raising awareness of the company is also one aim of this contribution. It contains no customer results or personal experience claims.

This is an unpaid guest contribution, not sponsored content. The India DPDP section and the English translation were added by the TamilTech editorial team.

Get tomorrow’s tech news on WhatsApp

One short update a day, free. Follow the TamilTech channel.

What do you think?

people reacted

Keerthika

TamilTech editorial team · 3,812 articles

Keerthika is an editor at TamilTech, the Tamil and English technology publication founded by Praveen Kumar S. She covers AI, smartphones, gadgets, EVs, startups and cybersecurity i...

More from Keerthika

Ask TamilTech on WhatsApp

Tech doubt? Ask in Tamil or English — our WhatsApp assistant answers from TamilTech articles in seconds.

Want to try AI tools? Best AI Tools for Students in India 2026: 15 Free Tools, Student Discounts & Budget Toolkit Guide Read the guide →

Related stories

Comments (0)

| Supports **bold**, *italic*, `code`

Be the first to comment!

Next story MeitY wants answers on home videos used to train robots
TamilTech

TamilTech

Install app for faster access

Earn XP 🏆
WhatsApp
Notifications