AI Insights | airlock docs

AI insights looks at how your organization uses its tools, skills and agents, once a week, and suggests what to do about it: skills worth creating, skills that need work, skills nobody uses, and agents that could be set up better. It only suggests. Nothing is created, changed or deleted until an admin decides to act.

AI insights is off by default. An admin turns it on for the whole organization.

Turning it on and off

AI insights is admin-only and lives under Settings → Organization → AI insights.

Open Settings → Organization.
Open the AI insights section and switch it on.

The section shows exactly what is read and what never is before you switch it on. Turning it off stops future weekly reviews; a review that is already running when you switch it off may still finish. While it is off the Skill suggestions card is hidden, but suggestions already made are kept and reappear when you turn it back on. Every change to the switch is recorded in the audit log.

What it reads, and what it never reads

Read	Never read
Which tools were used over the last 30 days, how often, in which order, and by how many people: counts only, never who	The content of tool calls: the arguments sent to your applications and the responses they returned, including those kept with approval requests and in the audit log
Your organization's skills: name, description and instructions	Personal skills of members
Your agents: name, description, instructions and model preference
Suggestions it made before, so it does not repeat itself

The review is done by an Anthropic Claude model on AWS Bedrock, through an inference profile that keeps processing in the EU. The model is given no tools, so it cannot change anything, and airlock does not use your data to train models. See the privacy policy and the Data Processing Agreement, Annex A, entry 19.

When it runs, and who hears about it

The review runs every Monday at 06:00, Brussels time, for every organization that has AI insights on.

When it finds something new, every active admin gets one email listing the new suggestions, grouped by kind, each with the usage pattern it is based on. A week with nothing new sends no email. The email's button opens the Skill suggestions card on the dashboard.

Acting on a suggestion

Open suggestions appear on the dashboard in the Skill suggestions card, for admins only, while AI insights is on.

Kind	What a click does
New skill	Create skill opens the new-skill setup, prefilled from the suggestion with a name, a description and an outline, and set to be available to everyone. Nothing is saved until you create it. Creating it as an organization skill marks the suggestion done; keeping it personal does not, because a personal skill reaches nobody else.
Improve	Open skill opens the skill's own page, with the suggestion shown above it. Often the suggestion comes with a proposed edit: open it to see the change as a diff against the skill's current text, then Accept to save it as a new version, Edit to load it into the editor and adjust it before saving, or Reject to dismiss the suggestion. Without a proposed edit, make the change in the editor as usual. Either way, saving marks the suggestion done. If you opened the skill another way, Link it to my next save on the suggestion does the same.
Retire	Open skill opens the skill's own page, with the suggestion and a Retire skill button above it. Retiring deletes the skill after the usual confirmation, and marks the suggestion done. A skill synced from GitHub has to be disconnected first.
Agent	Open agent opens the agent's own page, where you make the change. The suggestion is not shown there, so keep the card open or note it first. If the agent has since been deleted, the finding stays as text.

Dismiss removes a suggestion you do not want. The review is shown the suggestions it made before and told not to repeat them, so a suggestion that was dismissed or acted on does not come back in a later week. An exact repeat is always dropped; a reworded one is unlikely but possible, and you can dismiss it again.

Each run makes at most twelve suggestions and prefers a few worth acting on over many plausible ones, so some weeks bring none.

Proposed edits

A proposed edit is the whole skill with the suggested change made, written to change only what the suggestion is about. When a suggestion is about when a skill gets picked, for example trigger phrases its description is missing, the proposed edit changes the description instead of (or as well as) the instructions, and the diff shows each one separately. Long skills get one too: the review reads every skill's length, and a much-used skill that is far longer than its steps need is itself a suggestion, to shorten it. The edit for a long skill is written in a second step from its full text, a few minutes after the suggestion appears. It can also move reference material an agent needs only now and then (a lookup table, a long example, background) into new attachments, leaving a line in the skill that says which attachment to load. The moved text is copied exactly as it was, and the new attachments are listed under the diff. Accept saves the new text and the new attachments together, so if the save fails neither is kept. Edit is not offered for such an edit, because the editor holds the skill's text only. A week brings at most six proposed edits, and a suggestion can come without one when the review could not say precisely what to change.

A proposed edit is made against the skill as it was on the day of the review. If the skill has been changed since (its text, or its name, description or tags), the edit is marked as out of date and can no longer be accepted or loaded into the editor, because it could undo those changes. The diff still shows what it suggested, so you can make that change by hand. The same goes for unsaved edits of your own: discard them to accept a proposed edit, since saving them first changes the skill. Nothing is ever applied without an admin accepting it.

From an AI client

Admins can work through suggestions from any connected AI client as well. list_skill_suggestions lists the open ones, each with its id and the evidence behind it, and dismiss_skill_suggestion dismisses one. To act on a suggestion, pass its id as recommendation_id to the tool that does the work: create_skill for a new skill (as an organization skill), update_skill or update_skill_content to improve one, delete_skill to retire one. An improve suggestion with a proposed edit also carries proposed_content and proposed_content_revision: passing them to update_skill as content and content_revision, with content_is_complete: true, applies the edit. A proposed description comes as proposed_description, passed to the same call as description (with content_revision, and without content when there is no proposed_content) (an edit that moves text into new attachments also carries proposed_attachments, passed to the same call as new_attachments), and is refused if the skill changed since the edit was proposed (proposed_content_stale says so in advance). The suggestion is marked done once that change is saved. If your policy asks for approval on those tools, the suggestion is marked done when the approved change runs.

The same rules apply as in the Control Room: admins only, nothing while AI insights is off, and suggestions about skills or agents in a private, paused or archived space are left out.

Good to know

It reads organization-wide usage. A suggestion such as "three people run the same four tools every Monday" is about how your team works, and the review sees counts, never names or who did what. A pattern can still come from a single person's habits, but no suggestion says who that person is.
It judges from what it can see. Very long skills are shortened before the review, and the usage reports cover the busiest tools first. A suggestion says what it is based on, so you can check it before acting.
Turning it off keeps what was suggested. Existing suggestions are hidden while it is off and come back, as they were, when you turn it on again.