PACT AI
Use the world's most capable AI on your most proprietary work, without giving it away. PACT AI bridges external AI like ChatGPT and Claude with your own data sovereignty, so your business secrets stay protected on your infrastructure, and remain yours.
Most tools force a trade-off: use a powerful external model and expose your data, or protect your data and settle for a weaker tool. PACT AI removes the trade-off.
The problem it solves
Today, to get real help, teams paste sensitive records, contracts, and strategy into external chatbots. The moment they do, that data leaves the company. Banning AI loses the upside. Using it openly loses the data. Neither is acceptable.
Use the smartest external models on everyday work, and reason on your crown jewels in-house. You get the intelligence and keep the data, with the trade-off gone.
Grounded knowledge (encrypted RAG)
Answer from your private documents while they stay encrypted. Relevant context is retrieved without decryption, so the model is grounded without ever reading your files in the clear.
Autonomous action
Agents read, reason, and act within governance, automating sensitive workflows without exposing plaintext.
Every action accountable
Retrieval and every agent action are verifiable and audited through Moduli Chain, meeting enterprise governance and compliance.
Build agents your way
Assistive agents
- A co-pilot over encrypted private data
- Human stays in the loop
- Fastest path to value
Workflow agents
- Automate defined steps on ciphertext
- Connects to the systems you run
- Every action logged and provable
Autonomous agents
- Multi-step action across systems
- Acts without exposing plaintext
- Verifiable and governed end to end
How PACT AI works
Answering and acting run as one encrypted flow. Data stays encrypted from source to result, and every step is on the record.
Connect data & tools
Wire encrypted data sources (TorusDB, documents) and the systems you already run.
Tokenis
Tokenis detects and tokenizes sensitive entities in both questions and documents, so only non-sensitive, tokenized content is ever sent to an external model.
Encrypted retrieval
Relevant grounding is retrieved over ciphertext with no decryption. Analytics such as aggregation and filtering run directly on the encrypted data.
LLM gateway
Plaintext never leaves. External models like ChatGPT or Claude see only tokenized or de-identified content, or everything runs locally.
Reason in-house or de-identified
Reasoning happens inside your boundary (the in-house model), or on a de-identified version (external). No confidential value is exposed either way.
Grounded answer or action, audited
A grounded answer or an executed action. Every step is verifiable and logged on Moduli Chain.
How is my sensitive data protected?
See it in two situations. The path changes with what is actually sensitive. The most sensitive data never leaves the company.
PACT AI routes by how sensitive the content is.
Fields such as names, account numbers, and IDs are tokenized by Tokenis before anything reaches an external model, then restored inside your boundary. You keep using the strongest external models, and the underlying values are never exposed.
You get two safe answers at once. The in-house model reasons on the full document inside your walls, while external AI like ChatGPT and Claude reasons on a de-identified version rephrased locally. Where they agree, you have confidence. Where they differ, you have insight. Nothing confidential leaves either way.
PACT AI's in-house model works on your full context, and it can be tuned on your own material to fit your domain. An approach that only sends de-identified text outside cannot build that depth.
Tokenis detects and tokenizes sensitive entities in both questions and documents, so only non-sensitive, tokenized content is ever sent to an external model.
PACT AI is the AI add-on to TorusDB, the encrypted data platform, so your data stays encrypted at rest, in transit, and in storage, on your own infrastructure, including CPU-only and air-gapped networks where cloud AI cannot go.
- You
- PACT AIClassifyTokenisGateway
- TorusDB vault
- Local AI · in-house
Hi Hong,
Sorry your order #12345 is delayed.
You asked for a refund.
Wondering if this fits your environment?
Take 15 minutes with a security engineer, against your real workload.
Why this is different
The crown jewels never leave
Not a gateway that masks and forwards. A layer where the crown jewels never leave, and you still get the reasoning power of external AI like ChatGPT and Claude.
Two answers for your hardest questions
Full-context in-house plus de-identified external, so privacy and intelligence stop pulling against each other.
Runs where cloud AI cannot
Encrypted data, your own infrastructure, CPU, air-gapped.
Related use cases
See where this runs in production, across AI, finance, and core enterprise systems.
Frequently asked
Quick answers about this product.
No. Your documents stay encrypted end to end. Relevant chunks are retrieved as ciphertext, and an LLM gateway ensures no plaintext reaches an external model. The model receives only the grounding it needs to answer, never the raw text.
An LLM gateway sits between your data and any external model and prevents plaintext from leaving the encryption boundary. Retrieval and grounding happen inside that boundary, so the model gets only what it needs. Nothing sensitive leaks.
Agents act only within pre-defined permissions and policy. Every action is recorded so it can be verified and audited after the fact. You can always see what an agent did and why it was allowed to.
No. Agents operate without plaintext exposure: sensitive data is tokenized at the gateway or processed in-house, and only de-identified content ever reaches an external model. Automation never requires handing over plaintext, so you avoid the usual trade-off where automating more work widens your exposure surface.
Automate and answer on data you couldn't expose.
Let's connect your documents and design an agent for your most sensitive workflow.