Approvals and gates

There’s a sliding scale between “the AI does what I ask and shows me the result” and “the AI silently sends the email to my client.” This chapter is about where on that scale you want each kind of work to sit.

Jootle’s defaults are conservative: anything with real-world consequences (sending an email, posting to a calendar, creating a contact, writing to an external system) goes through an approval gate by default. You see what’s about to happen, you approve or revise, and then it happens.

The conservative defaults are right for the first weeks. Once you trust the patterns your AI handles well, you can loosen specific gates so that routine work doesn’t need clicking through.

The kinds of actions that gate by default

Out of the box, your AI will pause for approval on:

Outbound messages. Emails to be sent, Slack messages, SMS, anything that leaves your instance.
External writes. Creating a calendar event, adding a row to a Google Sheet, updating a CRM record, anything that touches a third-party system.
Spend. Anything that costs money beyond AI usage itself (paid API calls, services where you’ve configured limits).
Irreversible deletes. Removing a contact, deleting an artifact, archiving a project.
Identity actions. Anything that changes who can access what (adding users, changing permissions, rotating credentials).

Actions that don’t gate by default:

Internal writes (creating a task, attaching a list to a project, adding a note to an idea).
Research and read-only fetches.
Drafts (your AI happily writes drafts; sending the draft is the gated step).
Asking questions of you.

What an approval gate looks like

When your AI is about to do something gated, you’ll see a panel in chat or a notification that says:

Ready to send this email to Mike. Subject: “Quote accepted, when can we start?” Body: [draft body]. [Approve and send] [Revise] [Cancel]

You can approve, revise (which loops back to a draft), or cancel.

On Telegram, the same gate shows as a quick reply: “Send this email? (yes/no/edit).”

On email, your AI can send you a preview-to-send: “Reply yes to send, no to cancel, or send me edits.”

Loosening a gate

When a pattern is reliable, you can loosen its gate.

“From now on, when I ask you to send Lily an update, just send it. No approval needed.”

That installs a standing rule:

Audience: Lily
Action: send email
Approval: not required

Future “send Lily an update” requests go through directly. The artifact is still saved, and you can see what was sent in the activity log, but you don’t have to click approve.

You can scope these rules to:

A specific recipient. “Skip approval for emails to Lily.”
A specific kind of action. “Skip approval for calendar event creation.”
A specific project. “Inside the kitchen project, skip approval for emails to Mike.”
A specific channel. “When I’m on Telegram, skip approval for SMS reminders” (because you’ve effectively already chosen to be quick).

Combinations work. “Skip approval for calendar events under 30 minutes inside the household program” is a perfectly valid scope.

Tightening a gate

The reverse is also possible. If something is happening automatically that you’d rather review:

“From now on, always check with me before sending anything to a client.”

That tightens the gate. Even if you previously loosened it, this rule overrides for client recipients.

Tightenings always win over loosenings, on the principle that “ask first” is the safer default when rules conflict.

High-stakes always-gates

Some actions cannot be loosened by default. They always require explicit per-action approval:

Spending real money. Even if you loosen spend approval generally, charges over $100 (or whatever threshold you set) always confirm.
Sending to a “guarded” contact. Specific contacts marked as guarded (legal counsel, your spouse’s parents) always confirm.
Bulk actions over a threshold. “Email all 200 customers” always confirms, regardless of how loose your email-sending rules are.

These exist because the cost of a mistake at scale is enough that the friction of one extra click is worth it.

You can adjust the thresholds in settings, but the gate itself can’t be removed.

Reviewing what passed through gates

The Activity view shows everything your AI has done, including what it sent and to whom. Loosened gates produce entries you didn’t have to click for; they’re still recorded.

If you notice a pattern of “I wish I’d reviewed that one”, tighten the relevant gate. The system errs toward more friction, not less, when you express doubt.

Working with team members

If your instance has multiple users (see Users and permissions), approval gates can route to specific people.

A team member sends an email; the approval gate fires to you (the instance owner).
A team member’s request to invite a new user gates to you plus an admin.
A team member’s request to spend money over $50 gates to you.

You can also delegate. “The marketing manager can approve emails to anyone on the customer list without checking with me.” That delegation is itself an action you’d approve once and then it’s standing.

What happens when an approval times out

A gate that hasn’t been approved in a reasonable time (default 24 hours for routine actions) is auto-cancelled. The draft is preserved as an artifact; the send didn’t happen. You can come back to it if you want.

Urgent gates (“the lawyer needs this signature today”) can be marked priority by your AI when context suggests urgency, and you’ll get a follow-up reminder on your default channel as the window closes.

A small note on trust

Customers sometimes try to loosen everything in the first week to feel less friction. We’d encourage the opposite: keep the gates tight for two weeks, watch what your AI is choosing to send, and only loosen the patterns where you’ve seen ten in a row and they were all what you would have sent.

Trust earned this way doesn’t blow up. Trust granted up front, before you’ve seen what gets produced, can.