How does the On Premises PII Cleanser work?

Why teams ask for it

Many firms already use Claude or ChatGPT connected to Microsoft 365, Outlook, and SharePoint. Those shared drives hold years of tax returns, bank statements, payroll files, and driver’s license scans. The assistant can often tell you a file contains a Social Security number, which also means it has read it. Business plans from the major AI companies promise not to train on your data, but the files are still processed on their servers.

A password or API key can be rotated after a leak. A client’s Social Security number can’t.

How it works, step by step

1. Your staff ask as usual. “Summarize this bank statement” or “What’s missing from this tax packet?” with the file attached.
2. A local model finds the personal details. It runs on the Cleanser, in your office, and marks names, addresses, Social Security and tax ID numbers, account and routing numbers, dates of birth, phone numbers, and emails.
3. Each one becomes a placeholder. “Jane Smith” becomes PERSON_1, her SSN becomes SSN_1. The same value always gets the same placeholder, so the cloud model can still follow who is who.
4. The cloud model answers. It sees the numbers and structure it needs to do the work, such as balances, line items, and dates of transactions, but not who they belong to.
5. The real details go back in. The Cleanser swaps the placeholders back before your staff see the answer. The table that links placeholders to real values never leaves the Cleanser.

Is it safe to upload bank statements or tax returns to ChatGPT?

It depends on the plan and on the rules you work under. Consumer AI plans can use chats to train future models unless the user turns that off. OpenAI says it doesn’t train on ChatGPT Business or Enterprise data by default, and Anthropic’s commercial plans say the same. In every case the file is processed and stored on the vendor’s servers for some period under the vendor’s terms.

For tax preparers there’s an extra layer: Section 7216 limits disclosing tax return information to anyone outside the firm. Our CPA page covers that in detail. The Cleanser cuts down what leaves your office, but whether a redacted return still counts as tax return information is not settled, so ask your counsel before relying on it.

Can I use Claude or ChatGPT without sharing PII?

Yes, with the Cleanser in front of it, or by running the AI entirely in your office. The two fit different needs:

What it doesn’t hide

The Cleanser removes identifiers, not facts. A cloud model still sees amounts, dates, business names you choose to keep, and the shape of the document. Someone who already knows a client could sometimes recognize them from the facts alone. Detection is also not perfect: an unusual name or a number in an odd format can slip through. We tune it on samples of your own documents, show you what it caught and missed, and log every question so you can check what was sent.

If a file can’t leave the office in any form, use the private server instead.

Can I build this myself?

You can. Open-source tools like Microsoft’s Presidio detect personal data, and LangChain documents a reversible version that restores the originals afterward. The hard parts come after that: catching PII in scanned PDFs and spreadsheets, keeping placeholders consistent across a long conversation, connecting to Microsoft 365 and SharePoint, per-user logins, and keeping the mapping table secure. That’s the part we install and maintain.

Private AI without the server price

It runs on a mini PC and costs well below our GPU servers; we’ll quote it once we know your file types and number of users.