Mandate applies your rules to each message before it reaches an AI tool.

A signed record your auditor checks on their own computer, and approval by someone you name before a message is sent. The same rules apply on AI websites and in your own software, and Microsoft’s tools keep running beside it.

Yes, in either of two ways. Proxy settings send eight AI websites, such as ChatGPT, Claude and Gemini, to Mandate, and your gateway keeps inspecting everything else. Or a gateway that speaks ICAP hands AI traffic to Mandate to be checked, so your gateway stays the one place that traffic is decrypted. Either way, Mandate adds approval by someone you name before a message is sent, and a signed record your auditor checks on their own computer.

Mandate copies Microsoft 365 Copilot’s own log, from Outlook, Teams, Word and Copilot Chat, into your record afterwards, and marks what your rules find there as a warning. The separate Microsoft Copilot website, copilot.microsoft.com, is checked before a message leaves, like the other AI websites.

If your rules refuse it, the message is not sent: the AI website shows an error instead of an answer, and an email gives the person your rule’s reason and a reference for the administrator. They can take out what the rule found and send it again. If your rules hold it, the AI website shows an error too, and an email says it is waiting for approval. A second email gives the decision, and an approved message goes through when the person sends it again within 24 hours of the approval. If nobody decides within 24 hours, the message is never sent, and an email says so.

Measured on a Mandate server in Canada, checking adds under 60 milliseconds to half of all requests and under 150 milliseconds to 99 in 100.

AI websites wait until Mandate is back, so nothing leaves unchecked. Email and other websites carry on.

Yes. ID numbers, card numbers and client names are found whatever the language, and people’s names in many languages. You add your own terms in each language you need. Mandate’s screens are in English and French, and you write the email a person gets in each language you use.

No. A policy says what is allowed, but it cannot stop a message or show what happened. Mandate applies it to each message before it leaves, and the signed record shows each decision.

You can start from one of Mandate’s ready-made rule sets. Tell your staff before IT sets it up: our free AI acceptable use policy template includes a notice that their AI use is recorded.

Try a draft rule on pasted text or a batch of your own files first, and see what it would catch before anyone is affected. The test uses the same checks as live messages, and what you try is not kept.

Yes. Give it a set of typical documents. It finds what they contain and proposes a draft rule set you can adjust and switch on, and the documents are not kept.

Only if you want rules that catch client names. You paste it in, one name per line, and it stays in your own workspace, apart from every other customer, and is used only to check messages. The record notes that a client name was found, without saying which, and an export shows how many names the list holds, not the names, unless an administrator chooses to include them.

IT sends a certificate and proxy settings to your Windows and Mac computers with Microsoft Intune, Group Policy, Jamf, a remote management tool or whatever it already uses. Chrome, Edge, Firefox and Safari are all covered. Signing in with Microsoft Entra ID and copying Microsoft 365 Copilot’s log each need an app registration in Entra ID, which a Microsoft 365 administrator approves. Your own software needs one new address and key.

Point your software at your Mandate address with a Mandate key, in the OpenAI, Anthropic or Gemini format; Cohere and Mistral use the OpenAI format. Mandate sends each request on with your own provider key. OpenAI’s Chat Completions and Responses, Anthropic’s Messages and Gemini’s generateContent work as before, streaming included, and a checked answer still streams. Tool calls and tool results in a request are checked like any other text. A refused request gets an error that carries your rule’s reason, and a held one gets a reference your software checks. Once someone approves it, Mandate sends it on and the check returns the answer. Each organization has its own request limit, and a request over it gets a standard “too many requests” error.

client = OpenAI(base_url="https://your-mandate-address/v1", api_key="your Mandate key")
client = Anthropic(base_url="https://your-mandate-address", api_key="your Mandate key")

Through Mandate, over MCP with OAuth 2.1 sign-in. An agent reaches only the tool servers you register and the tools you allow. Hidden characters are removed from what a tool sends back, and your rules can mask what they find there, before the agent reads it.

A held message’s text is kept until someone decides it, and never past 24 hours. Otherwise, nothing people type, upload or get back on AI websites is kept, and text from your own software, AI agents and the AI logs Mandate copies is kept only if you turn that on, for the period you set. The record keeps who or which agent, which tool, when, the rule, the kind of detail found and a SHA-256 hash of the request.

A signed record of which rule applied to each message, what it did and when. Any change after signing fails the check. Check a sample record →

Yes. Mandate applies your rules to each message before it leaves, and the signed record shows each decision. Export the period your insurer asks about, report included, and they can check it on their own computer.