Use the AI assistant

This section explains how to open the AI assistant, run supported tasks, review and apply proposed changes, configure preferences, and work within current limitations.

Get started

After being enabled, the AI assistant is available on every page of Authorization Hub. Contact your system admin if you cannot see the AI assistant button at the bottom right of the page.

note

The AI assistant works within a project. Questions about policies, attributes, and tests are answered only after a project is opened. Outside a project, for example on the project selection page, only general ABAC and ALFA questions are answered.

Ask your first question

1. Open a project and go to the Policies page.

2. Open the policy you want to ask about. You can ask targeted questions about specific parts of the policy by selecting them in the editor.

3. Click the AI assistant button at the bottom right of the page.

   The assistant panel opens on the right, and the page behind it remains usable.

4. Type your question and press Enter.

   The answer appears in real time as it generates.

tip

Words such as "this", "here", or "it" are interpreted as your selection or, without a selection, as the policy, policy set, or rule under the cursor. For example, select a rule and ask "Explain this".

Example questions

Task	Example question
Understand a policy	Explain what this policy does in plain language
Find where a rule belongs	Which policy should hold a rule for contractors?
Check for errors	Is this policy valid?
Test behavior	Test this policy for all permit cases
Test a scenario	Can a manager in Sales approve an expense of 5,000?
Write a rule	Draft a rule that lets auditors read every record
Review dependencies	What attributes does this policy depend on?

Panel controls

Control	Action
Send or Enter	The message is sent. Add a new line with Shift+Enter.
Stop	The answer in progress is stopped.
New conversation	A new conversation starts, and earlier messages are discarded.
Apply	The proposed change is written to the open policy as an unsaved change.
Undo	The most recently applied change is reverted.
Auto-apply	Proposed changes are applied without clicking Apply.
Minimize	The panel is hidden. While an answer is in progress, the AI assistant button shows a spinner.
Esc	The panel is closed.
Info icon	The AI disclaimer is shown.

Supported tasks

The task is recognized from the message. Tasks can be combined in one message, for example "test and fix this rule".

Task	Example	Result
Explain	Why is a contractor denied here?	The policy is read and, if needed, a request is tested to trace the decision. Each statement is supported by a line or a tested decision.
Locate	Where should a rule for auditors go?	The project's policies are reviewed, and one policy, policy set, or rule is recommended. If none fits, the reason is given.
Validate	Does this compile?	The policy is compiled, and each error is listed with its line.
Test	Can a nurse view records outside their shift?	The requests are evaluated against the policy by the decision engine, and each decision is reported.
Change	Allow managers to approve expenses up to 10,000	The change is written and offered as a proposed change.
Look up	Which attributes can I use for the department?	The attributes in the project's Dictionary, or the projects visible to you, are listed.

Proposed changes

When the assistant proposes a policy change, the answer displays a block that reads "Proposed change to {file_name}" with an Apply button.

The system automatically validates each proposed change by compiling and running it against the decision engine. The result appears below the change, even if compilation fails or verification isn't possible. Review this result before applying the update. To protect your work, proposed changes will never delete content unless you ask to remove it.

Apply a change

1. Review the proposed change and the result below it.

2. Click Apply.

   The change is written to the open policy as an unsaved change.

3. Review the result in the editor, then save it as usual. A saved change is staged in My workspace until it is saved to the project.

warning

Apply replaces the open policy with the proposed version and overwrites any edits made in the editor since sending your question. Click Undo to revert.

note

Apply is available only if the policy was open in the editor when you sent the question.

Undo a change

Click Undo at the bottom of the panel to revert the most recently applied change. Only the last apply can be reverted.

Auto-apply

Select Auto-apply at the bottom of the panel to apply changes without clicking Apply. With Auto-apply:

• A change is applied only when the answer is complete and contains exactly one proposed change.
• A change is not applied if the policy was edited after the question was sent.
• A change is not applied while the panel is minimized. It is applied when the panel is opened again.

The setting is off by default and resets when you reload the page.

Information used by the assistant

Each message is sent with the open policy file, your selection, and the earlier messages of the conversation. While an answer is prepared, the project's data is read with your permissions, so only the data visible to you in Authorization Hub is used. This information is sent to the model provider configured by your administrator. Read the How the AI assistant works section for the full list.

tip

Select the part of the policy you are asking about to get a more precise answer. Read the Policy size section for long policy files.

Conversations

Each conversation is private and tied directly to you and the current project. No other users can view your chat history.

• One message at a time: Wait for the assistant to finish responding before sending another message.
• Context window: Long conversations focus only on recent messages. Click New conversation when changing topics.
• Expiration: Inactive conversations are automatically removed. Read Conversation limits to learn more.

Preferences

Preferences let you set custom instructions for all AI assistant responses, eliminating the need to repeat yourself. For instance, you can automatically ask for test results in table format or limit responses to five sentences.

Preferences are private to you and tied directly to the active project. Other users cannot view them, nor do they apply to your other projects.

Preference commands

Type the commands in the assistant panel with a project open. The system processes commands directly without sending them to the language model.

Command	Action
/preferences add <text>	A preference is saved and applied to every answer.
/preferences add <task>: <text>	A preference is saved and applied to one task only. The task is explain, locate, validate, test, or change.
/preferences list	Displays a numbered list of your preferences for this project.
/preferences delete <n>	Preference number n is deleted from the list.
/help	The commands are listed.

To change a preference, delete it and add it again.

Preference examples

/preferences add Keep explanations under five sentences.
/preferences add test: Show test results as a table with columns Case, Expected, Actual.
/preferences add change: Add a short comment above every new rule.
/preferences list
/preferences delete 2

tip

A test: preference applies when a test is requested but not to tests that run while a change is made. To cover both cases, add the preference without a task.

Preference limits

Limit	Value
Preferences per user, per project	10
Length of one preference	1,000 characters
Total length of your preferences in a project	4,000 characters

A preference is not saved if it is a duplicate, if it contains characters that are not visible, or if you are not a member of the project.

What preferences can't do

Preferences dictate how the assistant processes tasks and formats responses. Preferences cannot:

• Override the assistant’s scope or safety rules.
• Add, remove, or relax policy rules, conditions, or obligations that were not requested in the current prompt.
• Skip change validation. Authorization Hub verifies all proposed changes regardless of preferences.
• Expose system configurations or credentials.

If a preference conflicts with these rules, the assistant will apply the compliant parts and explain any skipped instructions in its response.

warning

You must not store passwords, keys, and other secrets in a preference.

Limitations

The following limits apply to the AI assistant in this release.

Scope

• As of Authorization Hub release 26.2, the AI assistant can only modify ALFA policies that are currently open in the policy editor on the Policies page. It can read attributes, but cannot modify them. Furthermore, it cannot read or modify attribute connectors, domains, cache configurations, or Policy insights templates and reports. Support for these features may be added in a future release.
• The AI assistant never saves, publishes, or deploys changes. Read the Apply a change section for details.
• Apply is available only if the policy was open in the editor when you sent the question.
• Only the open project is used. Policies cannot be compared or copied across projects. Outside a project, only general ABAC and ALFA questions are answered.
• The AI assistant answers only questions related to authorization policies and Authorization Hub. Other topics are declined.
• Only OpenAI-compatible model providers are supported. Read the Supported models section for details.

Conversation limits

• One message is answered at a time in a conversation.
• A message can be up to 8,000 characters long.
• Only the most recent messages of a long conversation are used.
• Conversations are removed after a period of inactivity.
• Previous conversations cannot be browsed.
• An answer is stopped after a time limit. Read the Response time limits section for details.

Policy size

The open policy file is included with each message, up to a limit of 32,000 characters. For larger files, only the beginning of the file and the lines surrounding your selection are sent. This character limit applies strictly to the active file. It does not depend on the length of your message or conversation history. Furthermore, policies read directly by the assistant are always read in full.

The assistant reads a large file that is saved in full before proposing any changes. If a long file contains unsaved edits, save it before requesting a change. This limitation applies to the initial release and is planned for improvement in a future update.

Safety

The assistant's scope and safety rules are instructions to the language model, not an enforced control. Read the How the AI assistant works section for details.