AI Script Authoring & Review

When no existing script in your library fits, the Breeze AI assistant (in chat) and background AI agents can write one, get it independently reviewed, and — after a human approves it, or your organization has explicitly opted in to unattended runs — execute it on the devices you’re working on.

Nothing the AI writes ever runs unless a human reads a plain-language summary of it and approves, or your partner and your organization have both opted in to the unattended lane described below and the script cleared every safety check.

The three review layers

Every AI-written script goes through three checks, in order, before it can run:

Automated scan. The script is scanned for known-dangerous patterns (disabling security tools, deleting system files, exfiltrating credentials, and similar) before anyone — human or AI — sees it. A match on the most dangerous pattern list stops the process immediately; nothing is shown to a reviewer or an approver.
Independent AI review. A second AI model — one that never sees your conversation with the assistant, only the script itself and the devices it targets — reads the script and produces a plain-language summary, a risk tier (Low, Medium, High, or Critical), a list of findings, and a recommendation to approve, request changes, or reject.
Human approval. You see a readable approval card: what the script is for, what it does, the reviewer’s summary and risk tier, and the code itself — never raw JSON. Low- and Medium-risk scripts can be approved by one qualified technician; High- and Critical-risk scripts require a second, independent approver (“four eyes”).

The approval card

When the assistant or an agent proposes a script, you’re shown:

Goal and expected effect, in the AI’s own words
Risk tier and the reviewer’s summary
Findings the reviewer flagged, with severity
The script itself, exactly as it will run
Rollback notes, if the AI provided any

Approving runs the exact script you read — there is no “edit and approve” option. If you want changes, choose Request changes with a note; the AI gets your feedback and the reviewer’s findings, and proposes a revised script that goes through all three review layers again.

Some scripts also require an extra acknowledgement before you can approve: any script that matched a “strict” pattern (for example, a registry change under HKLM) needs the approver to hold the Scripts permission and complete a fresh multi-factor authentication check — the same requirement as approving a strict pattern on a script from your library.

Provenance

Every script version and every execution that comes from this feature carries its origin:

The script library shows an Origin column (Human, AI proposal, Imported, or System) and a “Reviewed” / “Edited since review” badge — editing an AI-authored script by hand clears the review badge, because what was reviewed is no longer exactly what would run.
Script detail has a Provenance panel: who authored it, the reviewer’s summary and risk tier, who approved it and how, and links to every execution.
Device Activity on a device’s page lists AI-authored script runs with an AI badge and a link to the originating proposal — opening that link shows the goal, status, risk tier, and the reviewer’s summary.

If the organization that produced a proposal is later merged into another organization, or its data retention period ends, the proposal’s own record may be erased. The script version and execution it produced remain visible; the provenance panel and the device-activity link both show “evidence erased” instead of a broken page.

Promotion to the library

An AI-authored script is not added to your library automatically. After it has run and its author-specified verification claim has been independently confirmed (not just “the command exited 0” — an actual check, like confirming a service is running), Save to library becomes available on the proposal. Saving requires the Scripts permission and a fresh MFA check, exactly like creating a script by hand. From then on it’s an ordinary library script, usable anywhere a library script is — including in an AI agent’s autonomous “Act mode” toolkit.

The unattended lane

For narrow, low-risk, high-confidence cases, Breeze can run an AI-authored script without waiting for a human — but only when every one of the following is true:

Your partner (MSP) has raised the ceiling that allows the unattended lane for its organizations at all.
Your organization has separately and explicitly turned the lane on. A partner-level ceiling never turns it on by itself.
The independent reviewer rated the script Low or Medium risk (never higher), found it matched the stated goal, is reversible, and has an adequate verification plan.
The script touches only a small, configurable allow-list of resource classes (for example: services, temporary files, DNS cache, printing) — never credentials, security tooling, disk, boot configuration, user/group accounts, firewall rules, shell evaluation, or anything the scanner can’t confidently classify.
It targets exactly one device, has a timeout of 5 minutes or less, and — on Windows, for the resource classes that warrant it — a System Restore point is captured immediately before the run.
Your organization hasn’t had two unattended runs in a row fail their independent verification — if it has, the lane pauses itself for that organization until an administrator resets it.
An hourly cap on unattended runs, set by your organization within your partner’s ceiling, hasn’t been reached.

Any script that doesn’t clear every one of these goes to a human instead — there is no partial credit. On Linux and macOS, the resource classes that would need a restore point (registry, services, filesystem changes) are not eligible for the unattended lane at all in this release, because only Windows can take one.

Admin setup

Your partner administrator raises the partner-wide ceiling (maximum risk tier, allowed resource classes, hourly cap) under partner-level AI settings.
An organization administrator with the Approvals permission and a fresh MFA check turns the lane on for that organization, optionally tightening any of the partner’s limits further (but never loosening them).
The AI Risk Dashboard’s Script Proposals panel shows unattended-run counts and the lane’s current state (open/closed) so you can monitor it.

FAQ

Can the AI run any script it writes? No. Every AI-authored script is reviewed by an independent model and then either approved by a human or, only if your partner and your organization have both opted in, cleared through the unattended lane’s fixed safety checks.

What if I disagree with the reviewer? You decide. The reviewer’s risk tier and recommendation are advisory to you (though they set the deterministic floor for the unattended lane) — you can approve a script the reviewer flagged as Medium risk, or reject one it recommended approving, and the AI Risk Dashboard tracks how often that happens.

Does this replace the Script Builder AI in the editor? No — Script Builder AI is a separate, in-editor assistant that helps you write scripts by hand and only runs a script with your explicit click. This feature is about the chat assistant and background agents authoring and running a script mid-conversation or mid-task, with independent review in between.

Can I turn this off entirely? Yes — script proposing can be disabled per organization independently of the unattended lane.

Scripts — the script library, execution history, and security patterns this feature builds on
AI Features — the assistant, its tool tiers, and the AI Risk Dashboard
AI Agents — background agents, Act mode, and how they use this feature
Approval & Assurance — four-eyes approval and MFA step-up