> ## Documentation Index
> Fetch the complete documentation index at: https://docs.blindsight.io/llms.txt
> Use this file to discover all available pages before exploring further.

# Import a policy

> Turn the AI policy your company already wrote into Endpoint policies you review and approve.

**Import a policy** reads the AI policy your company already wrote and proposes
Endpoint policies from it. Nothing becomes policy until you approve it.

Importing the policy you already wrote is one of the two main ways to add a policy.
**Endpoint → Policies** opens with both: **Import the policy you already wrote**, and
**Write one here** (**New policy**). Select **Import a written policy** to start. Once
you have imported before, the line **Last import** shows when, the file name and what
it created, with **Open** to go back to it. A person who may only write policies sees
**Write a policy**, and one who may only import sees just the import.

You need `endpoint.import.manage` to upload and review,
and `endpoint.manage` to apply. See [Permissions](https://docs.blindsight.io/endpoint/permissions.md#the-catalog).

## What is kept

Your document is read once to find the rules in it. It is not saved. When reading
finishes, the file and its text are discarded. Blindsight keeps the rules you
approve, the sentence each one came from, and a record of who approved what.
Proposals you do not apply are deleted when you apply or discard the import.

## Import a document

**1. Upload**

Choose **File** and drop in your policy in Word, PDF, a spreadsheet or plain text,
or choose **Paste text** and paste it. Written policies, access matrices and
permission tables all work. You can leave the page while the document is read.

**2. Coverage**

Before any rule is shown, you see how much of the document the platform can
enforce:

| Bucket | What it holds |
| - | - |
| **Enforced** | Statements that become policy rules, once their groups are matched. |
| **Not enforceable** | Real rules that are outside what Endpoint controls. They are listed so the count is complete. |
| **Refused** | Statements that were proposed but did not pass checking, with the reason. |

Open a bucket to read the sentences in it. Select **Export the coverage report**
to keep a copy. Export it here, because the quoted sentences are deleted with the
rest of the import.

**3. People**

Match each group the document names to a group from your directory, or to
**Everyone in this workspace**. If the document names nobody, every rule applies
to everyone.

**4. Review and apply**

Each proposal shows the sentence it came from and what it changes on your
devices. Edit a proposal in place if you need to. The summary confirms that
nothing becomes less restrictive and that no roles or permissions change. Select
**Apply** to create the policies. Devices pick them up like any other policy
change.

## What an import can turn into a rule

| In your document | Becomes |
| - | - |
| A named AI service, for example "Engineering cannot use DeepSeek" | A rule for that service's hosts, taken from the [AI services catalogue](https://docs.blindsight.io/endpoint/ai-services.md). One name can cover several hosts. |
| An explicit domain | A rule for that domain. |
| A category, for example "Image generators are prohibited" | A block for that category of AI services. |
| A subject people must not ask AI about | A [topic rule](https://docs.blindsight.io/endpoint/topic-rules.md). |
| Fail open or fail closed, file scanning, and what to do with a file that cannot be redacted | The matching policy settings. |
| Inspection settings for a single service | The policy's settings for that service. Applying them also needs `endpoint.protections.configure`. |

A host comes only from a service the document names or a domain it writes out.
An import never guesses a host.

## An import never proposes a control the document did not state

A proposal always comes from a sentence that says it. Blindsight does not turn a
sentence into the nearest control that happens to exist.

* A sentence about a kind of data becomes personal data protection only when it
  names personal data. Another kind of information, such as source code, forecasts
  or contracts, is a different control. It is held with its sentence for a person to
  decide, and nothing is switched on for it.
* When the reader was not sure what a sentence asks for, nothing is proposed. The
  sentence is held for a person.
* Words that mean the whole workforce, such as "employees" or "all staff", apply to
  everyone in the workspace. They are not matched to a group in your directory. A
  group named like "Engineering staff" stays a group.

## What an import holds back

An import does not turn every sentence into a rule. It holds back a sentence it
cannot turn into a rule, and lists it under **Not enforceable** with the reason.
For example:

* Rules about something Endpoint does not control, such as who may use the console.
* Approving or allowing a whole category of AI services. An import never loosens
  policy.
* A rule assigned to a single device rather than to people or groups.

## Import a revised policy

When your policy changes, import the new version. Blindsight compares it with the
rules applied from the last import:

* A rule that has not changed is hidden.
* A rule whose sentence changed is shown as before and after.
* A rule that is no longer in the document is listed under **No longer in the
  document**, so you can retire it.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.