AIDataVeil
AI does the work. You set the rules. AIDataVeil protects every exchange routed through it, and keeps the proof.
AIDataVeil sits between your company and the AI models it uses, whether the request comes from an employee, a software tool or an agent.
The AI never needs to know who Emily Carter is.Her name, address, account number and every other detail that identifies her become stand-ins. AIDataVeil puts the real ones back in the answer.
Three ways to handle sensitive content
Masking deletes meaning. Stand-ins keep it.
Most security controls limit what can be sent. AIDataVeil lets authorized work happen by transforming the flow, and refuses the rest.
Other tools
Security tools block the exchange.
The work does not reach the model.
Redaction tools remove information.
The model reasons on what remains: [NAME] must pay [AMOUNT].
AIDataVeil
AIDataVeil transforms, then restores.
The model reasons on coherent stand-ins, and the authorized values come back. What cannot be protected is not sent.
Redaction removes information. AIDataVeil uses coherent stand-ins to keep a usable context, then restores the authorized values: the substitution is reversible and governed by your policy.
How AIDataVeil does it: seven properties
- Where content is processed
- On your machines or your host's, in memory.
- Where mappings live
- In the history, stored encrypted. The online services hold keys, access records and opaque identifiers, not content.
- Operator access
- The publisher and the vault host hold no readable content. Content is processed in memory, on the machine that runs AIDataVeil, under your control or your host's.
- Readable retention
- No readable business content at rest within AIDataVeil's scope. History is kept for a period your administrator sets: 30 days, 90 days, one year or the length of the subscription. It is destroyed when the subscription ends.
- Agents
- They reason on stand-ins; a real value appears only in a decision the mandate allows.
- One policy
- The same policy for people, software and agents.
- Independent verification
- A signed receipt for each exchange, verifiable offline with the published verifier, without the publisher.
What changes
Stop preparing data for AI.
Your teams work on the real file. AIDataVeil controls what the model receives. Nothing to develop: people open the workspace, compatible software and agents change one address.
Without AIDataVeil
- Teams paste cleaned-up snippets into AI, or nothing at all.
- Whole use cases stay forbidden by legal and security teams.
- Agents cannot touch sensitive records.
- Every AI provider and every tool needs its own controls.
- Security sees the rules, not what actually left.
With AIDataVeil
- Teams work on the actual document.
- One policy governs chat, software and agents.
- Security sees what left, and where it went.
- Compliance receives evidence without content.
- Changing AI models does not mean rebuilding the controls.
Bring a file that looks like yours. The demonstration shows what the AI would have received, what it receives instead, and what comes back to your screen.
Request a demonstrationThe first question everyone asks
Why trust AIDataVeil if you don't trust your AI provider? You don't have to take our word for it.
AIDataVeil is software that runs where you decide, on your machines or your host's. It is designed so your team can verify what it does.
The protection engine runs on your side.
On your computers or on your own host.
You can watch what goes out.
Detected identifiers travel as stand-ins, and only to the AI models your policy allows.
No readable business content kept at rest.
History is stored encrypted, with the keys managed separately from the data.
Check it yourself, without the publisher.
Plant your own marked test values, then search the device, the logs and the counters.
How to verify
The protection engine runs on your side. Your files are processed there, in memory. They are not sent to the publisher.
You can watch what goes out. Your team can verify the destination and inspect the protected payload before it leaves.
No readable business content kept at rest. The publisher keeps no readable business content at rest within AIDataVeil's scope.
Check it yourself, without the publisher. Verify each signed receipt offline with the published verifier.
Real values are handled only in memory, on the machine that runs AIDataVeil: yours, or your host's.
AIDataVeil works with supported cloud, private and local AI models, under the same policy.
AIDataVeil makes every AI data flow enforceable and independently verifiable.
AI data governance
One set of rules for people, software and agents.
AI data governance that applies itself to every exchange. The same engine sits in front of a person, a software or an agent.
- Who
- People, teams, software and agents, each with its rights.
- What
- Each kind of data gets its treatment: stand-in, kept, or refused.
- Where
- Which AI may receive it: cloud, private or local, as your organization's AI data governance policy decides.
- Spaces
- Each discussion, client and mission has its own stand-ins.
Emily Carter is Rachel Hayes here, Grace Miller there.
People work in a chat workspace.
Anyone can select a word to protect it, on top of the organization's lists.
Software goes through a gateway compatible with supported AI APIs.
No development: compatible software only changes its AI address. Your lists and field rules protect what it sends.
Agents act under mandates, with badges and an MCP gateway.
Route the agent through AIDataVeil and sign its mandate in the console, with its own words to protect.
How the policy is applied
Software goes through a gateway compatible with supported AI APIs. Real values are resolved only where the authorized action requires them.
Every version of the policy is signed, with a way back to the previous one. A correction request goes back to the same AI that answered; a request moves to another model only when your policy sends it there.
A file AIDataVeil cannot read reliably, such as handwriting or a blurred scan, is not sent. The message stays calm and offers a way forward.
One console for your lists of clients and terms, the field rules, the authorized AI and the agents' mandates.
Every version of the policy is signed, and every exchange leaves a receipt.
When a file cannot be protected, it stays on your side and the person sees why.
This document was not sent.
This file needs a clearer copy before it can be protected.
site-inspection-notes.pdf · 3 pages
The passage
The problem is transmission: what a file reveals when it is sent to an AI.
The AI receives the file it needs, without the identities it doesn't need to know.
Identifiers are detected on your side and replaced by stand-ins that hold together, so the model still has a file it can reason on.
Real values come back only in the authorized context, or at the moment of an authorized action.
A signed receipt records what happened, without the business content.
The answer
The answer comes back restored and checked.
AIDataVeil tells personal details apart from what the work needs, in messages, documents, spreadsheets and scans.
Replaced
People and companies, addresses, emails, phones, ID, tax, health and bank numbers, cards, plates, IP addresses, secret keys, file numbers.
Kept
Dates of events, amounts, durations, roles.
Adapted
Birth dates keep the age, postcodes keep their city.
Your rules
Lists of names and terms to protect and words allowed to stay, set by the organization, plus any word a user selects.
Formal notice to Elena Marquez, 14 Elm Grove, Leeds LS6 2AQ, e.marquez@northgate.example: rent of 2,450 overdue since 3 March.
Formal notice to Sofia Lindqvist, 9 Mill Lane, Bristol BS1 5TR, s.lindqvist@northgate.example: rent of 2,450 overdue since 3 March.
Dear Ms Marquez, the rent for 14 Elm Grove has been overdue since 3 March. Unless 2,450 is received within ten days, the lease terms apply.
What the check does before you read the answer
Before it is shown, AIDataVeil checks the restoration: unresolved stand-ins, mismatched identities and unexpected protected values are caught according to the controls supported by that version. This check does not certify the factual accuracy of the AI's answer.
Same gender, a city that matches its postcode, an email built on the stand-in name. The amount is not an identifier, so it stays. The authorized real values are restored in the answer, streaming included.
The check corrects what it can locally first; if the answer still fails, AIDataVeil asks the same model once more.
If a point still cannot be confirmed, AIDataVeil shows the answer with a clear note of what was checked.
The action
The model reasons on stand-ins. Real values appear only where an authorized action needs them.
An agent can prepare, classify and decide with protected identities. When an approved tool must act on the real person, AIDataVeil resolves the value at that boundary, under the agent's signed mandate. The model does not need the real value back.
The agent's model decides on stand-ins: “Send the approved notice to Sofia Lindqvist”.
With a valid mandate, AIDataVeil resolves Sofia Lindqvist to Elena Marquez at the boundary, for this action only.
The mail connector acts on the real person: e.marquez@northgate.example.
The receipt records each state separately: authorized, attempted, and confirmed only when the integrated tool returns confirmation. It is signed and does not contain the notice.
The guardian
The history and its keys are kept apart.
Business data is not kept in readable form at rest: the history is encrypted, the keys are kept separately, and proofs do not contain the file.
The vault holds encrypted blocks and no keys.
It is run by the hosting partner, who holds the blocks and no key.
The guardian holds the keys and no history.
The publisher's key service holds the keys and no block.
A purge removes the key material first, then the blocks.
A purge is designed to destroy the key material required to decrypt that conversation.
How keys and history are handled
The guardian holds the keys and no history. Neither can decrypt the stored history alone.
A purge removes the key material first, then the blocks. Separately, encrypted blocks are physically removed according to the retention period set by policy.
Zero readable data does not mean zero processing: AIDataVeil processes what protection requires, and the model receives the authorized context it needs to work. Closing an account withdraws its keys: what was stored can no longer be read.
The proof
A receipt for every exchange handled by AIDataVeil.
Each receipt is signed and carries no content: no names, no text. Your auditor checks it offline, without the publisher.
- Program
- 8d1e 4ab7 c2a9 a7fc listed in the public register
- Space
- a41f e9c2 7d3e 5b18
- Policy
- version 14
- Destination
- Private model · hosting partner
- Requested provider retention
- none where the provider's API supports it
- Content
- none
- Signature
- valid checked offline
This receipt contains no content. Values shown are a specimen.
It states which program processed the exchange, which version of your policy applied, where the request went, and which retention was requested from the provider.
- Offline. The verifier for Node and Python checks the signature and the chain up to the published root, without asking the publisher.
- Public register. The fingerprint of every program release is published, so the program on a receipt can be matched.
A receipt proves what was applied: which program, which version of your policy and which destination. It does not judge the content of the answer.
The deployment
One protection layer. Wherever AI enters the work.
The same protection engine on workstations, virtual desktops, servers and containers, plus a gateway for software.
- Workspace
For people: the chat workspace.
- Gateway
For software: one gateway in front of the models you authorize.
- Embedded
For software vendors: the gateway placed inside your own product's architecture.
- Agents
Mandates, badges, decision receipts and the MCP gateway.
- Sovereign
A private model run by your hosting partner, under the same policy and receipts.
It is deployed with single sign-on and device-management packages, or through IT services partners who install and support it.
Built for
For the files that cannot go to a model as they are.
AI is blocked first where files carry other people’s secrets. The same layer then applies to every organization, whatever its sector or size.
Where data blocks AI today
- Law firms and legal teams
- Insurers and banks
- Healthcare
- Accounting and audit
- Human resources
- Industry with sensitive IP
- Software vendors adding AI agents
Files every organization holds
- Customer and patient records
- Employee and HR files
- Contracts and legal matters
- Financial figures and forecasts
- Product plans, research and source code
- Supplier and partner information
Questions asked by IT and privacy leads
The hard questions, answered plainly.
How do you know the protection works?
Each version is measured before release on real-world documents, spreadsheets, scans and trick cases, and ships only when it passes its tests. Your own team can check it too, with its marked test values and the signed receipts.
Do we need to develop anything?
No. A compatible software only needs its AI address changed, and an agent is routed through AIDataVeil with a mandate signed in the console.
Why not an enterprise plan from the AI provider?
An enterprise plan sets one provider's storage and training terms, and that provider still receives your files as they are. AIDataVeil sends it stand-ins for detected identifiers, applies one policy across the models you authorize, your software and your agents, and gives you receipts you verify without the provider or the publisher. The two work together.
Is this anonymization?
No. AI works on stand-ins, then the real values come back in the answer: the substitution is reversible and governed. That reduces identity exposure, but the rest of the file can still be confidential, which is why AIDataVeil also controls which AI a file may be sent to.
What exactly does a receipt prove?
It proves what was applied: which program processed the exchange, which version of your policy applied and where the request went. It contains no content, and your auditor verifies it offline, without the publisher.
Which AI models does it work with?
The models your policy authorizes, reached through a connector or through the gateway compatible with supported AI APIs, including a private or local model. Any application that lets you set the address of its AI can be routed through AIDataVeil.
How fast can an organization deploy it?
It uses the tools IT teams already know: one protection engine for workstations, virtual desktops, servers and containers, packages for device management and a single sign-on connection.
Demonstration