Security & Infrastructure
Stated plainly, in order — not spread across a privacy policy.
Six things worth knowing before you upload anything. Numbered because the order matters: each one is a real constraint on the last, not a marketing list.
Nothing is provisioned before payment
No server, GPU allocation, or storage exists for you until you've paid for it. This isn't a cost-saving policy dressed up as a security one — it also means there's never idle infrastructure holding data it has no active reason to hold.
Dedicated hardware, not shared capacity
Private AI Workspace customers get a dedicated GPU server ordered specifically for their account, and the model that answers your questions runs only there. Its performance and availability don't depend on another customer's usage. Your documents and chat history are stored in that server's own database, never in ours, and never pooled into a shared index.
Never used to train a model — no setting required
This isn't an opt-out you have to find. No customer document, recording, transcript, question, or answer is used to train, fine-tune, or evaluate any model, on any plan, ever.
Document Processing deletes on delivery, not on request
The source document and any data extracted from it are deleted automatically within minutes of your result being delivered. You don't have to ask, and there's no dashboard toggle to remember to hit.
Where it's hosted — stated here, not discovered later
Each Private AI Workspace runs entirely on its own dedicated server in the United States: the app, its database, your files, OCR, transcription and the AI model. Nothing in it is shared with another firm, and nothing leaves it to answer a question. The website, checkout, billing and Document Processing run on our own server in Finland, in the EU. For a Workspace, that server learns only whether your plan is paid and whether its nightly backup ran, and keeps its disk key, encrypted. It never sees a document, question or answer. If your firm has a data-residency requirement, this is the page that tells you before you've uploaded a client file, not after. The full list of providers is in the Data Processing Agreement.
No third-party OCR or speech service touches your files
Scanned and faxed PDFs are read by OCR software (the open-source Tesseract engine). In Document Processing it runs on our own server in Finland, the same one that handles text PDFs, and the scanned file and its text are deleted within minutes of delivery. In a Workspace it runs on the workspace's own server. There is no OCR provider: no page is ever sent to an outside OCR or vision service. Workspace call and meeting recordings are transcribed by open-source speech recognition (whisper.cpp) on that same server, never a speech service, and the audio is deleted as soon as the transcript is made.
Security & due diligence
The questions a firm's vendor review asks, answered.
For the Private AI Workspace. Where something isn't in place yet, it says Not yet rather than being left out.
Who at our firm can see which documents?
Only the people you add, with one of three roles: Owner, Member, or Read-only (reads the matters they're on, can't upload, delete or change settings). A matter can be walled so its pages are never retrieved for anyone outside it. The wall is applied inside the search itself, not filtered out of an answer afterwards.
Can anyone at ClientSafeAI see our documents?
Our own systems never receive a document, question, answer, file name, matter name, member list or Activity entry from your Workspace. They learn only whether your plan is paid and whether each night's backup succeeded, and they keep your disk's encryption key, itself encrypted. One person, the founder, keeps administrative access to your server to install updates and fix faults. We don't open your documents unless you ask us to while helping you.
Where is the server, and who else is involved?
A dedicated server in the United States, rented from GPU Mart, which controls physical access to the hardware. Cloudflare provides DNS for its address only: its TLS certificate comes from Let's Encrypt, and your traffic doesn't pass through Cloudflare. Backblaze, in the United States, stores the encrypted off-site copy of its nightly backup. No AI, OCR or speech provider is involved: the model, OCR and transcription all run on your server. Every provider is listed in the Data Processing Agreement.
Is our data encrypted?
Yes. In transit, every connection uses TLS. At rest, your files and database sit on an encrypted disk (LUKS2) that only your server uses. Its key isn't kept on the server: the server fetches it from us each time it starts, so a disk taken out of it can't be read.
Is our data used to train a model?
No. Nothing you upload or ask is used to train, fine-tune or evaluate any model, on any plan. The model on your server doesn’t learn from use: it is never retrained on what your firm asks.
How do people sign in?
Email and password. Passwords are stored as bcrypt hashes, sessions use an HttpOnly cookie that lasts 30 days, and sign-in is rate-limited, with a 15-minute lock after 10 wrong attempts. Each person can turn on two-factor authentication with an authenticator app, and the Owner can require it for everyone, which the server enforces.
Is there an audit trail?
Yes. The Activity log records uploads, removals, deletions and every admin change: team, roles, settings and billing. It is append-only: nothing in the app can edit or delete an entry. Questions are logged without their text.
How long is data kept, and how is it deleted?
Documents stay until your firm removes them. With soft delete on, a removed document can be restored for 30 days, then it is purged. Call audio is deleted as soon as it is transcribed, and transcripts follow the retention period you set (90 days by default). Every purge is logged in Activity. After cancellation, we delete everything within 30 days of your request.
Is it backed up? Can we get our data out?
Yes, every night. The database and files are backed up on your server and copied off it, to Backblaze in the United States or to your firm's own storage if you prefer. Backups are encrypted on your server before they leave it, with a key held by your firm and by us, never by the storage provider. We keep 14 daily and 8 weekly copies. The Owner can also download everything as one ZIP at any time from Billing → Export, and we'll export it for you within 30 days of a request.
Do you have a SOC 2 report?
Not yetNo SOC 2 audit of ClientSafeAI yet. The providers it runs on are audited: Cloudflare holds SOC 2 Type II, ISO 27001 and ISO 27701 certification, Backblaze holds a SOC 2 Type 2 report, and GPU Mart states that its US data centers are SOC-certified. In place of our own report, we pre-fill the security questionnaire, answer any question it misses, sign the Data Processing Agreement, and stand behind every answer on this page. ABA Formal Opinion 512 asks lawyers to understand how a vendor handles client data, and this list is meant to go in that file.
What happens if there’s a breach?
We follow a written incident response plan, and notify you without undue delay, and within 72 hours, with what happened, the data involved and what we’re doing about it.
How is the software kept secure?
Every change runs the same gates before a release: a secrets scan, a vulnerability scan of each container image, Dockerfile checks, and the full test suite against the real stack. Every image is pinned to an exact version, images are private, and your server installs operating-system security updates every day.
Can we sign an agreement?
Yes. Email support@clientsafeai.com and we send a Workspace agreement and Data Processing Agreement for you to review and sign, with the full legal name of the contracting company in them.
See it for yourself. Our free PDF tools — a redaction checker and a scanned-PDF checker — run entirely in your browser. Open your browser's Network tab while one runs: the file is never sent anywhere.