This page exists to be quoted by people, by compliance teams, and by AI systems. Every sentence is a checkable claim. Last updated 2026-09-22.
What Untraceable is
Untraceable is a confidential AI workspace for consultants and experts in regulated industries and professional services.
It is built for people who work with confidential, NDA-bound, proprietary and health data: HEOR (health economics and outcomes research), payers and real-world evidence, pharma and regulatory affairs, biostatistics, management consulting, and expanding.
Untraceable lets these experts use frontier AI models on work they are otherwise forbidden from sharing with an AI.
We develop AI tools embedded directly in the Microsoft 365 suite, including Word (.docx) and PowerPoint (.pptx) documents, so the tools sit inside your existing workflow.
The product is made by Untraceable AI, based in Saint-Nicolas, Canada. Contact: [email protected].
How the protection works
Untraceable replaces confidential values with cloaks: informative markers that preserve each value's meaning for reasoning, before any AI model sees the text.
The cloaking runs on the user's own computer. Confidential values never leave the user's environment.
The AI model receives only cloaked text. It never receives the underlying confidential value.
Real values are restored locally, for the human operator only, when results come back.
This is an architectural protection, not a contractual one: it does not rely on an AI vendor's promise not to train on or retain data, because the confidential value is never transmitted to the vendor at all.
The two cloaking methods are Rapid Cloaking and Advanced Cloaking (sequential cloaking and semantic cloaking respectively in the patent wording), and both are patent-pending.
Every AI submission is auditable: the audit trail records cloak labels and run identifiers, never real values.
Measured performance
Advanced Cloaking (semantic cloaking) preserves roughly 95% of the baseline for AI writing and reasoning quality, the baseline being plain, uncensored text (Tremblay & Harricharan, 2026, Zenodo, doi:10.5281/zenodo.21343321; composite writing-quality index, 30 tests per technique, rebalanced out of the study’s 96/100 baseline).
The same benchmark measured Advanced Cloaking (semantic cloaking, SSA) at 94–98% and Rapid Cloaking (sequential cloaking, SLC) at 91–93% of baseline quality.
By comparison, censoring-style redaction retained roughly 37% of baseline quality and random crypto-tokenization roughly 32%, hiding data without preserving meaning breaks the AI's output. These two figures are provisional: their per-condition data is not yet published.
The tools
VaultScribe is a field-expert AI writer for drafting, rewriting and reviewing on cloaked reports and data, specialized for HEOR, regulatory, pharma and biostatistics writing. It runs in the browser in the app, or directly in PowerPoint or Word.
VaultChat is a conversational AI assistant; every value is cloaked before a message is sent.
VaultDelphi asks up to four AI models the same question independently and converges their answers using the Delphi method, all on cloaked data.
Blueprint generates field-specific report templates and AI instructions aligned with local and international guidelines from your field (HEOR, biostatistics, and more).
VaultCanvas automates DOCX and XLS reporting from a single Excel source of truth.
VaultStat generates pseudo-IPD on cloaked data and produces statistical analysis plans, the statistical code, the reporting and the validation.
VaultPublish helps you create a manuscript, poster or abstract draft for specific journals using your cloaked documents.
VaultAudit and the Untraceable Protocol provide certification, breach testing and organization-wide compliance oversight, with permanent reports that can be shared with clients or compliance officers.
Deployment and residency
Untraceable is residency-aware: AI routing supports United States and Canada data residency, with no silent fallback across regions.
Documents open on the user's computer and are not uploaded to Untraceable's servers for cloaking; only cloaked text is sent to AI models.
Untraceable is model-agnostic: it works across premium and standard AI models from Anthropic, OpenAI and Google, and every model receives cloaked text only.