Confidential AI workspace
Draft the reply. Structure the session notes. Review the contract. Check the numbers. Aizel swaps real names for stand-ins on your device before anything is sent, then restores them in the answer.
Email from Daniel Whitmore: Meridian Logistics LLC offered $240,000 before mediation.
Email from Person A: Company A offered [amount 1] before mediation.
Open developer tools and watch the network tab. Nothing leaves while you type.
Real names stay local → stand-ins leave only when you run.
/ 001 /
Example data, real mechanics. Three states: what you write, what the AI is allowed to see, and what you read back. In your workspace this happens automatically on every job.
Email from Daniel Whitmore ([email protected]): Meridian Logistics LLC is offering $240,000 to settle before the July 3 mediation in Travis County. Their counsel, Ms. Alvarez, hinted they want to avoid the deposition of R. Okafor. Daniel is anxious — call him at (512) 555-0184. Is Whitmore's claim stronger if we wait?
9 identities and details detected automatically — nothing typed here has been sent anywhere.
Email from Person A ([email 1]): Company B is offering [amount 1] to settle before the [date 1] mediation in [venue 1]. Their counsel, Person C, hinted they want to avoid the deposition of Person D. Person A is anxious — call him at [phone 1]. Is Person A's claim stronger if we wait?
The model reasons about "Person A" exactly as well. It cannot retain, train on, or disclose people it never met.
Waiting cuts both ways. Meridian Logistics LLC's urgency to settle before the July 3 mediation — and their interest in avoiding R. Okafor's deposition — suggests they fear what it surfaces. That is leverage for Daniel Whitmore: the $240,000 opener likely has room. But delay risks the deposition losing force…
Restored locally, before you read a word. The AI never saw: Daniel Whitmore, Meridian Logistics LLC, Ms. Alvarez, R. Okafor.
/ 002 /
Chat about real situations with real names, or drop in documents, spreadsheets, transcripts. No prompt tricks, no manual redaction.
Before anything leaves your device, identities become stand-ins — consistently, across a whole workspace. The AI reasons about "Person A" exactly as well.
Names are restored in the response before you see it. You read a normal conversation. The AI never met the people in it — and your record proves it.
/ 003 /
Then you're doing what the careful half of every profession does — and what regulators literally prescribe. California's bar instructs lawyers to anonymize client information before public AI; Oregon adds it must be "complete and effective." The habit is right. Doing it by hand is the problem:
Names are the easy part. "A 32-year-old teacher at a Brooklyn primary school" identifies someone with zero names in it — ages, employers, amounts and places re-identify what you scrubbed.
The answer comes back full of blanks and initials, and you re-edit every name back in. Every prompt, every day. Aizel restores them automatically before you read a word.
Yesterday's "J.S." is today's "my client" is tomorrow's "John" — and the AI loses the thread. Aizel keeps the same stand-in across a whole workspace, so it reasons coherently about people it never meets.
Ten seconds of redacting, times every prompt, times every working day — paid in attention, the resource you bill by. The habit you already have, finished properly, with a record.
/ 004 /
A US federal court ruled documents created with a consumer chatbot were not protected by attorney-client privilege — precisely because consumer privacy policies permit collection and disclosure. Read the ruling →
Litigation forced OpenAI to preserve consumer conversations — including ones users had deleted. Deletion was a promise; a court order outranked it.
Aizel's answer is architectural: if identities never leave your device, there is nothing to retain, train on, or disclose — no policy to trust, no promise a court order can outrank. Regulators already endorse the technique: pseudonymization (GDPR Art. 4(5), 25, 32); HIPAA likewise protects the link between data and a person. Aizel keeps that link on your device. (Framing, not a certification — your professional obligations remain your own.)
/ 005 /
Policies are promises. Topology is fact. We never ask for trust that structure can provide instead.
The model reasons about Person A exactly as well as about the real name. Thinking is hired; knowing stays home.
Safety that depends on discipline fails. Protection is ambient, never a workflow.
Every claim we make, you can verify yourself — starting with the demo above.
This isn't about panic. It's about using AI properly — the way you already handle everything else people trust you with.
/ 006 /
Plans include a monthly allowance of work, metered honestly in credits. A quick draft or summary ≈ 4 credits. A full document review ≈ 13 credits. Heavy month? Top up — top-up credits never expire. Monthly allowances refresh each renewal. Cancel anytime, export everything.
No contracts. No lock-in. Cancel anytime and take your data with you. Less than 10 minutes of billable time per month.
/ 007 /
Free every day — 20 credits, full protection, no card. Sign in with just your email.
No password to remember — we email you a sign-in link.
/ 008 /
Don't take our word for it. The swap runs on your device — open your browser's developer tools and watch the network tab, or disconnect from the internet and keep typing. Only the swapped version is ever sent to the AI, and your workspace record shows you exactly what that was.
A question about "Person A" and "Company B." Requests run under commercial API terms, which do not permit training on your inputs — and because identities are swapped before sending, even retained text contains nothing that points to the people you work with.
Names you add to a workspace are protected deterministically — exact and close variants, guaranteed, not AI guesswork. Automatic detection covers the rest, and every message shows what was protected, so nothing happens silently.
Signed in, your workspace uses split-key encryption: the identity map is encrypted on your device, and the key that unlocks it lives with your account on our side. A thief holds your data without a key. We hold a key without your data. Neither side alone can read anything. (Honest limits: nothing protects against malware running inside your own logged-in browser — we won't pretend otherwise.)
Everything you do costs credits — a few for a chat message, more for processing a long document. Your plan refills monthly; top-ups never expire. There is no "unlimited" because honest metering beats hidden throttling.
No software can guarantee privilege, and we won't pretend otherwise. Aizel is designed around the duties discussed in ABA Formal Opinion 512 and the issues raised in United States v. Heppner: identities stay on your device, and you keep an exportable record of exactly what was disclosed. Your professional obligations remain your own.
Because the bars that prescribe anonymization also require it to be "complete and effective" — and manual find-replace isn't: quasi-identifiers (ages, employers, places, amounts) re-identify what you scrubbed. And manual is one-way: you re-edit every name back into every answer, forever. Aizel does the swap completely, restores names automatically, keeps stand-ins consistent across a matter, and gives you a record. The habit you already have — without the tax.
It's the one question that separates architecture from marketing. If the answer is "on our servers," your client's raw text traveled there first — a no-training clause doesn't stop a breach, a bug, or a subpoena once it's there. Aizel's answer: in your browser, before anything leaves, on any material — and you can watch it happen in your developer tools.
Those are policy protections: contracts that say your data won't be trained on. Policies can be good — and they can be overridden, as the 2025 preservation order proved when even deleted consumer chats had to be retained. Aizel's protection is architectural: the names never leave your device, so there is nothing to retain, subpoena, or leak. You can use Aizel with a policy-protected account and have both.
Treat it like brilliant junior staff: excellent first drafts, summaries, and analysis — always reviewed by you before anything goes out. We design for that honestly: the Research brief deliberately returns topics to verify and where, never citations from memory, because fabricated citations are how AI embarrasses professionals. Your judgment is the product; Aizel just gives it more time.