What Will AI Do With My Tiny Data? The Wrong Question

AI saves us hours. But are we paying for it with our client's trust?

Be honest. During this audit season, how many client PDFs did you upload somewhere to remove a password, merge files or convert a bank statement to Excel?

Five? Fifty? Did you even notice?

I raised this with a few fellow CAs recently. Almost every one of them gave me the same reply:

"Arre, what will Google or OpenAI do with one client's Bank Statement? They have bigger things to worry about."

Fair point. Nobody in California is waiting to read the bank statement of a kirana shop owner in Alibag.

But that is the wrong question.

The right question is: Did I have the right to send it there at all? And do I know where it went after I clicked Upload?

First, Let's Give AI Its Due

I use AI every day. It drafts notice replies in minutes. It explains a new circular in plain words. It fixes my Excel formulas and writes client emails in Marathi.

But look closely at that list. Not one of these jobs needs the client's actual data. AI's value is in what it knows, not in what your client's PAN or bank balance is.

Keep that in mind. It is the key to everything below.

Where Your "Tiny Data" Actually Goes

Nobody needs to misuse your client's data for things to go wrong. Here's what can happen to a file after you upload it. All of it has already happened.

It can stay for five years. You close the chat and assume it's gone. It isn't. Since 2025, Anthropic keeps consumer chats for up to five years if the user hasn't switched off model training. OpenAI and Google have similar settings. Default: usually ON.

"Deleted" may not mean deleted. In 2025, a US court ordered OpenAI to preserve ChatGPT chats, even deleted ones, for a copyright case filed by the New York Times. Later, it ordered 20 million chat logs handed over to the newspaper's lawyers. Users in India had nothing to do with that case. Their chats got pulled in anyway.

It can land on Google. In 2025, thousands of shared ChatGPT chats showed up in Google search, with names, resumes and office work. One tick-box. That's all it took.

Even Samsung got caught. Its engineers pasted secret code into ChatGPT to fix a bug. No hacker involved. Samsung banned AI tools on office devices soon after. If a company with Samsung's security budget can slip, so can a five-person CA office.

The real villain may be the "free PDF website". We worry about ChatGPT. Then we upload a client's password-protected bank statement to some free site to unlock it. Who owns that site? Which country? How long do they keep the file? Nobody knows.

Small pieces become a full profile. One Form 16 looks harmless. But PAN + Aadhaar + bank account + salary + address, collected over a year? That is exactly what loan fraudsters dream of.

Forget the AI Company. Look at Our Own Act.

Here's the part that should worry us more than any data leak.

As far as I know, Under the Chartered Accountants Act, disclosing information gained during a professional engagement to anyone other than the client, without the client's consent, is professional misconduct (Clause 1, Part I, Second Schedule). (Correct me if I am wrong)

Read it again. The clause does not ask whether the third party misused the data. It doesn't ask whether there was a leak. The disclosure itself is the misconduct.

And an AI company or a free PDF website is very much "anyone other than the client".

Two more things are coming our way:

• The DPDP Act, 2023 is now notified, and its main duties kick in from May 2027. A CA firm holding client data will have to protect it with reasonable safeguards. Penalties can run into Millions.
• Professional bodies abroad have already issued AI guidance. In the UK, members are warned that uploading identifiable client data into uncontrolled AI tools is a serious risk. ICAI is unlikely to be far behind.

And here's the simplest test of all: would your client be comfortable watching you upload their ITR, bank statement and Aadhaar to a website? They gave those documents to you. Not to your tools.

AI Is a Brilliant Junior Who Lies With Confidence

Confidentiality isn't the only trap.

AI can quote a section that doesn't exist. It can mix up the old and new regimes. It can cite a judgment that was never delivered. And it says all of this in the same confident tone as a correct answer.

Indian tax law changes with every Finance Act and every circular. AI's knowledge does not.

When something goes wrong, the department won't issue notice to ChatGPT. It will issue it to you. "AI ne bola tha" is not a defence.

So treat every AI answer like a draft from a bright but careless article assistant. Check every section, rate and date before it reaches the client.

So Should We Stop Using AI? No. Use It Smarter.

The trick is one line:

Let AI bring the brain. Keep the client's data on your own computer.

Here's how.

1. Don't Ask AI to Do the Work. Ask It to Build the Tool.

This one changed my office.

Instead of uploading client PDFs, I asked Claude to write small programs (Python scripts) that run on my own computer. The AI never sees a single client file. It only writes the instructions. My computer does the work, offline.

What I now do in seconds, without uploading anything:

• Remove passwords from bank statements, 26AS and AIS, a whole folder at once
• Merge 20 PDFs into one loan file
• Convert PDF bank statements into Excel
• Split a big PDF into separate files

Do you need to know coding? No. I just typed something like: "Write a Python script that removes the password from all PDFs in a folder." The AI gave me the script and told me how to run it. Python is free. The tool is mine forever.

(Tip: test it on a dummy file first and keep a backup of originals.)

2. Ask About the Case, Not the Client

Don't upload the notice. Describe it:

"An individual got a notice for AY 2024-25 for a ₹4.2 lakh mismatch between AIS and the return. How should the reply be structured?"

No name. No PAN. No address. Same quality of answer.

3. Use Offline Software You Already Have

Acrobat and many free desktop PDF tools unlock, merge and split files without going online. Excel's Get Data → From PDF pulls tables out of most software-generated PDFs. Rule of thumb: installed software over upload websites.

4. Switch Off Training. Today.

If you use ChatGPT, Claude or Gemini, open Settings → Privacy and turn off the option that lets your chats train the model. Two minutes. It doesn't make uploading client files okay, but it limits the exposure on everything else.

5. Bigger Firm? Get a Proper Business Plan.

Business and enterprise AI plans come with contracts that generally rule out training on your data. That's much closer to what the DPDP Act expects. It costs more, and you still need client consent.

6. Tech-Savvy? Run AI on Your Own Machine.

Free tools like Ollama or LM Studio let you run an AI model entirely on your office computer. Nothing leaves the machine. Less powerful than ChatGPT, but your data stays home.

The 30-Second Test Before Every Upload

• Does this task need the client's real data, or just the facts?
• Have I removed name, PAN, Aadhaar, account numbers?
• Do I know who runs this website?
• Could I do this offline instead?
• Is chat training switched off?
• Have I checked every section and date the AI gave me?
• Would I do this with the client sitting next to me?

If the last answer is "no", don't click.

The Bottom Line

AI is not the enemy. CAs who ignore it will fall behind.

But our clients don't hand us their bank statements because we are fast. They hand them over because they trust us.

So the next time a colleague asks, "What will they do with my tiny data?", ask back:

"Did your client say you could give it to them?"

Use AI for its brain. Keep the data at home.

For general awareness only; not legal advice. Views are personal.