Redactor

🔒 Remove client names, SSNs, and account numbers from your notes while keeping everything the tax law needs.

Live demo, fake data

Want to use this on your own files? Download your own copy.

Below is a walk-through on fake sample data. The "Ask Claude" step sends only the redacted sample to the Lab's key, so it's demo-only, don't paste real client data here. To use it on real files, download the offline copy: a single file that runs fully offline, nothing leaves your machine.

Download the offline tool Power users: the Python tool (PDF/Word, batch)

⚠️Redaction is a safety floor, not a guarantee: read every result before you share it. A clean result is still confidential client data, so whether it can go into a given AI tool is your call, it depends on that tool's data terms and on §7216 if it feeds a return.

How this works, in plain English

You paste text. The tool reads it on your own computer and looks for things that identify a person or business: names, Social Security and tax ID numbers, account numbers, emails, phone numbers, and addresses. It swaps each one for a neutral placeholder like [Person_1] or [SSN_1], and keeps a small private list (the "local key") matching each placeholder back to the real value. That list never leaves your machine.

You take the placeholder version to your AI tool, get your answer, then paste the answer back here. The tool uses your key to put the real names back in, right in this browser.

What it catches reliably: structured items with a clear pattern, like SSNs, EINs, emails, phone numbers, account numbers, and labeled ID / license / passport / policy numbers.
What it can miss: names, business names, and addresses are educated guesses, so it will occasionally miss one or redact too much. That is why you read the result yourself before sharing.
What it does not do: it does not make the amounts and dates anonymous, and it does not decide whether your AI tool is allowed to receive the data. Those are still your calls.

Start with this sample

This is sample data, not a real client. One click strips the identifying details and asks AI the four questions below: it works the numbers without ever knowing who the client is.

Here's what's safe to share

Review it, then copy it into your firm-approved AI tool. The client's identity stays with you.

✓ De-identified, review before sharing

See exactly what was removed

🔑 Your local key (stays on this machine)

This key is stored only on your local machine and never goes to the cloud. It is what's used to bring the AI's answers back and re-apply them to the original client, swapping tokens like [Person_1] back to the real names, right here in your browser.

Token	Type	Original value

↓

☁ Cloud AI step · only the redacted text is sent

Ask Claude

Each question runs against the built-in sample's safe text, never your own. Hit Send: the analysis works the numbers, but it cannot name the client or the SSN, because that never left your machine.

🔒These answers are pre-vetted examples of the analysis AI returns on the de-identified sample, shown here so the demo is always correct and reproducible. Your own text is never sent anywhere, this step is locked to the fixed sample. Treat AI output as a draft, verify every figure and citation against primary source before relying on it.

These four questions are preset, you can't edit them. Just hit Send.

⚖️Do not rely on these results for the final determination. The AI drafts and organizes; the judgment is yours, not the model's. Verify every figure, rule, and citation against primary source before it touches a return, a workpaper, or a client.

↩

Bring the answer back

When the AI replies with tokens like [Person_1], paste its answer here. Using your local key, we'll swap the real names back in, in your browser. Nothing is uploaded.

Run the demo above first, then the local key is ready here.

Keeps tax facts (amounts, %, dates, states, entity type). A safety floor, not a guarantee, eyeball the result before sharing.