SOPHIA XT

AI for regulated firms

Your staff can use AI. Your clients' data stays yours.

Safe-Desk is an AI assistant built for law, accounting, insurance and other firms that hold client data. Before any message reaches the AI, it swaps Social Security numbers, account numbers and the rest for placeholders. When the answer comes back, the real values go back in.

Not sure your staff put client data into AI? Run the free one-week check and see what they send, before anything is hidden.

$19 per staff member a month · 14-day free trial · includes $5 of AI per person each month, and AI beyond that is billed at the model's cost plus 30% · optional setup: $149 for up to 10 staff, $10 each extra · cancel any time

Scroll
Example
Your staff type

Draft a note to Ana: her SSN 123-45-6789 is on file and her refund of $1,310 is on the way.

The AI sees

Draft a note to [NAME_1]: her SSN [SSN_1] is on file and her refund of [TAX_FIGURE_1] is on the way.

Your staff get back

Hi Ana, we have your Social Security number (123-45-6789) on file, and your refund of $1,310 is on its way.

The owner's log

2:14 PM · Front desk · Alerted: Social Security number, Tax return figure, Person's name (detected)

Watch

See one paste get caught, in 30 seconds.

Windows app

Protects what staff paste into the ChatGPT, Claude and Copilot desktop apps. Needs a Safe-Desk account to sign in.

Download for Windows

Windows 10 or 11. Installs for the signed-in user, no admin rights needed.

Chrome and Edge extension

Protects ChatGPT, Claude, Gemini, Copilot, Perplexity and DeepSeek in the browser.

How to install it

Coming to the Chrome Web Store. Until then, sign in to Safe-Desk and download it from Settings.

The problem

Your staff already paste client data into ChatGPT.

Surveys put the share of employees who paste work data into AI tools above three in four, most of it from personal accounts the firm cannot see. For a firm bound by client confidentiality, IRC 7216 or GLBA, each paste is a disclosure nobody approved and nobody logged.

Banning AI doesn't stop it. It moves it somewhere you can't see. Safe-Desk gives staff AI they can use for real work, with the protection built in and a record you can read.

How it works

Three steps, every message, every time.

A staff member's message goes to the Safe-Desk gate, which checks it, swaps client data such as a Social Security number for a placeholder like [SSN_1], and sends only the placeholder version to the AI provider over TLS. The answer comes back with placeholders, the gate puts the real values back, and the staff member reads the real values. The owner's log records rule names only. Passwords and keys are stopped and sent nowhere. The Chrome extension applies the same gate on ChatGPT, Claude, Gemini, Copilot, Perplexity and DeepSeek. STAFF MEMBER types: SSN 123-45-6789 reads: SSN 123-45-6789 SAFE-DESK GATE 01 Check it against your rules 02 Swap client data for placeholders 03 Encrypt and send 04 Put the real values back Swap table: sealed, encrypted blob Chat text is not stored AI PROVIDER sees only: SSN [SSN_1] never sees: 123-45-6789 SSN 123-45-6789 TLS SSN [SSN_1] TLS reply with [SSN_1] TLS SSN 123-45-6789 TLS CHROME EXTENSION same gate for text sent into ChatGPT, Claude, Gemini and more OWNER'S LOG + PHONE ALERT 2:14 PM · Alerted: SSN rule names only, never values STOPPED passwords, keys not sent anywhere A staff member's message goes to the Safe-Desk gate, which checks it and swaps client data such as a Social Security number for a placeholder like [SSN_1]. Only the placeholder version goes to the AI provider over TLS. The answer comes back with placeholders, the gate puts the real values back, and the staff member reads the real values. The owner's log records rule names only. Passwords and keys are stopped. The Chrome extension applies the same gate on ChatGPT, Claude, Gemini, Copilot, Perplexity and DeepSeek. STAFF MEMBER TYPES SSN 123-45-6789 TLS SAFE-DESK GATE 01 Check 02 Swap for [SSN_1] 03 Encrypt, send swap table sealed chat not stored OWNER'S LOG + phone alert rule names only STOPPED passwords, keys sent nowhere TLS · SSN [SSN_1] AI PROVIDER sees only [SSN_1] TLS · reply with [SSN_1] SAFE-DESK GATE 04 Put the real values back TLS STAFF MEMBER READS SSN 123-45-6789 CHROME EXTENSION same gate in ChatGPT, Claude, Gemini and more
Client data is swapped for placeholders before anything leaves, so the AI provider only sees [SSN_1]. The real values go back in on the way home. The owner sees which rules fired, never the values. Passwords and keys are stopped.
01

Checked before it leaves

Every message is read against your industry's rules. Card, routing and IBAN numbers are verified with their checksums, and account numbers are only caught next to words like "account", so ordinary numbers pass untouched.

02

Swapped for placeholders

Client data becomes a placeholder like [SSN_1], and only that version goes to the AI. Passwords and keys are stopped outright.

03

Put back in the answer

The real values go back in before your staff read the answer. The same value keeps the same placeholder all conversation.

Industries

Rules written for the work you do.

Pick your industry at sign-up. Every template also covers cards, SSNs, ITINs, passwords, keys, emails, phone numbers and dates of birth.

Law firms

Text marked privileged or attorney work product stopped. A-numbers, passports, case and docket numbers protected and reported to the owner.

ABA Model Rule 1.6 · Formal Opinion 512

Accounting and tax

EINs, routing and account numbers, AGI, refunds and wages protected and reported to the owner.

IRC 7216 · FTC Safeguards Rule

Insurance agencies

Medicare numbers, health plan IDs, claim numbers, diagnosis codes and credit scores protected and reported to the owner.

GLBA · NAIC Model Law #668

Medical and dental

Patient identifiers, record numbers, diagnosis codes and dates of birth protected and reported to the owner. Safe-Desk does not sign a Business Associate Agreement, so it is for everyday office work and staff must not submit patient information.

No BAA · not for patient information

Financial advisors

Account numbers, IBANs, balances and net worth protected and reported to the owner.

SEC Regulation S-P · GLBA

Real estate and title

Bank details protected and reported to the owner; any mention of changed wire instructions is sent as written and reported to the owner. Property addresses are sent as written and reported to the owner unless you set them to Protect.

GLBA · closing wire fraud

HR and payroll

Direct-deposit details, A-numbers, passports and salary figures protected and reported to the owner; medical and leave wording sent as written and reported to the owner.

ADA · state SSN laws · I-9

Auto dealers

License numbers protected. Bank details and credit scores protected and reported to the owner.

FTC Safeguards Rule · FCRA

Home services

Card and bank details, and gate, alarm and lockbox codes, protected and reported to the owner. Customer names and job addresses are sent as written and reported to the owner unless you set them to Protect.

PCI DSS · customer security

Industry guides: AI for accounting and tax firms · AI for law firms · AI for insurance agencies · AI for financial advisors

Built to hold nothing

We never keep your messages. The gate keeps getting better.

For the owner

You see what was caught. Never the value itself.

Pricing

One price per person, with $5 of AI each included.

$19 per staff member a month
  • 14-day free trial. After it, Stripe bills monthly and sends invoices and receipts automatically.
  • Each seat includes $5 of AI inside Safe-Desk (Mercury today; Claude and ChatGPT are coming) every month, pooled across the firm. AI beyond that is billed at the model's cost plus 30%. Each answer shows what it cost. New firms start with a $50 monthly AI spending limit, a limit on AI charges beyond your included allowance, which the owner can change in Account.
  • Using your own ChatGPT or Claude, through the connector or the extension, or your own AI key, costs nothing extra
  • Protect and restore, the Chrome extension, the AI connector and the Gate API are included in the seat
  • Optional setup: we set it up for $149 one time for up to 10 staff, $10 for each extra person.
  • All nine industry templates, owner log, phone alerts
  • Cancel any time from the Account tab. It takes effect at the end of the paid month. A firm that returns does not get a second trial.
Start free trial

Straight answers

Questions firms ask first.

Does the AI provider ever see my clients' data?

Not the data the rules catch. The AI only ever receives placeholders for those values. For chat and the extension, the swap table that turns placeholders back into values stays in your staff member's browser, sealed; our server uses it for one request and never stores it. For files uploaded for the AI connector, the sealed swap table is kept on our server for 7 days so the finished document can be filled in. Encrypted copies in our daily backups are gone within 14 days after that.

Does it catch names and addresses?

Yes, in two ways. Paste in your client list and those names are swapped out wherever they appear as written, in any capitalisation and in the possessive ("Dana Reyes's"); the list is kept only as one-way fingerprints. Beyond the list, Safe-Desk detects names, addresses and dates from context, like "client Dana Reyes" or "1420 Oak Street". That detection is a best effort, because names have no checksum, and you choose per category whether to flag, protect or ignore.

Does this make us compliant?

No tool does that on its own. Safe-Desk helps you meet confidentiality rules like IRC 7216 and GLBA by keeping client data out of AI tools and keeping a record. It does not certify compliance. Safe-Desk does not sign a Business Associate Agreement (BAA), so it is not for patient information (protected health information).

Is it SOC 2 certified?

Not yet. We are building to the SOC 2 criteria and have not been audited. The security page lists what is in place today.

Can staff still use ChatGPT in another tab?

Yes, in a browser or app Safe-Desk doesn't cover. The extension checks what staff paste or send on ChatGPT, Claude, Gemini, Copilot, Perplexity and DeepSeek. Once it is listed in the Chrome Web Store, your IT person can force-install it so staff can't remove it; until then each person installs it from Settings. A written AI policy covers the rest.

Guides

Straight answers to what owners ask about AI.

All guides and industry pages: Safe-Desk guides.

Give your staff AI you don't have to worry about.

Start free trial