A claims file is a person's bad day in writing — often with medical and financial details attached. You shouldn't have to redact it by hand before asking an AI for help — and then un-redact the answer afterwards. That's what we do for you, automatically, before anything leaves your computer.
No card required. You land straight in a simple chat — drop a file and ask.
Three steps. No setup, no training, nothing to install. If you can attach a file to an email, you can use this.
Drag in the claims file, medical report, correspondence or policy schedule you're working on. Before anything leaves your computer, emails, phone numbers, dates of birth, bank and card numbers and health-insurance numbers are swapped for neutral placeholders automatically when their format matches a pattern rule. Policyholder and claimant names are flagged by a name-detection model that runs on your device too — it finds most, not all (95.3% of person names and 81.7% of company names on our published benchmark), so review the highlights before you send; and if the model isn't ready yet, the send is held rather than risk a name slipping through.
"Summarise the medical report in this claim." "List the open questions in this correspondence." "Draft a first response to this complaint." The AI does its work seeing only the placeholders — not the details we detected.
The reply comes back with the real names restored — automatically, exactly where they belong. No find-and-replace, no un-renaming, no copy-paste gymnastics. Just the finished answer.
We can't leak your policyholders' data — we keep no copy of it. The swap happens before your file leaves your machine, and we keep nothing readable afterwards: no clear-text copies of your documents or the answers. The only content-adjacent state is tokenised and transient — response buffers purged hourly, and opt-in multi-turn maps, stored client-encrypted (data map). There is nothing in the clear on our side to breach, subpoena or lose.
Contact details, dates of birth, bank and card numbers and health-insurance numbers are matched by pattern rules. Policy and claim references have no dedicated rule — they are caught only when their format matches one of those. Free-form policyholder and claimant names use a small name-detection model that also runs on your device; recall is strong on common names but not perfect (unusual or non-Latin names can be missed). For your most confidential files, install the desktop app — it tokenises the file itself entirely on your own machine before anything is sent to the AI, keeps your API key and logs local, and stores tokens in an encrypted vault.
Insurers and brokers handle health records, finances and personal histories under GDPR, insurance-secrecy rules and their own conduct standards. Here is what the architecture actually does, so you and your own advisers can evaluate AI use against them — this is not legal advice.
The workarounds people use today are slow — and fragile.
The manual rename. "Policyholder A", "Claimant X", find-and-replace before pasting — then reversing it all in the answer, hoping you didn't miss one instance on page 14.
The retyped summary. Typing an anonymised version of the document by hand because uploading the real one feels wrong. Half an hour gone before the AI has even started.
The quiet rule-break. Pasting the real thing and hoping it's fine. It's the fastest option — and the one that can cost a policyholder's trust — or a regulatory finding.
The blanket ban. The firm forbids AI entirely, and the productivity gain everyone else is getting goes to your competitors instead.
The duty of confidentiality isn't unique to insurance. If people trust you with their private affairs, this is built for you too.
Plain answers, no fine print surprises.
The leading AI models — the same ones behind ChatGPT, Claude, Gemini and others. The difference is what they're allowed to see: with us in between, they get your question and a version of your document with the identifying details we detect replaced by placeholders. They do the thinking; they never see the details we replaced.
The swap list — which real name maps to which placeholder — is created on your device and used to restore the answer on the way back to you. It's applied automatically, so the document you download reads exactly as if the AI had seen the real names all along.
Every request comes with a signed, tamper-evident receipt: a small certificate recording when the request happened and that the protections were applied. You can verify it independently, months later, and keep it in the claim file.
No clear-text copies. We keep no readable copy of your documents, your questions or the answers — the only content-adjacent state on our side is tokenised and transient (response buffers purged hourly; opt-in multi-turn maps, stored client-encrypted). Once your download completes, the readable version exists only on your machine. That's the point: we can't leak what we can't read.
Yes. The identifying details — names, dates of birth, policy and claim numbers, contact details — are replaced regardless of what kind of document they appear in, a medical report included. What the AI receives is the substance of the report without a person attached to it. And because nothing is stored, the report exists only where it always did: in your claims system.
No. Create a free account and you land directly in a simple chat. Drop a file, ask your question — that's the whole workflow. Prefer to stay in the tools you already use? After you register, you can install a small browser add-on for Chrome or Firefox and keep using ChatGPT, Claude, Gemini or Grok as you do today — names, emails and account numbers are replaced inside your browser before anything is sent, and the add-on downloads from your account page. (If your firm has an IT team that wants to connect it to internal systems, there's a full technical setup for them too, but you never need to touch it.)
Free to start, no card required. Drop a document, ask your question, and see the answer come back with the names where they belong.
Questions first? Write to hello@cloakapi.io — a human answers.