# AI autonomy

AI autonomy is one page that lists every kind of action an AI agent in your workspace may take, such as drafting a follow-up, creating a task, linking a conversation to a contact, or changing a CRM record. For each one you choose **Allowed**, **Ask first** or **Never**. Use it when you are ready to hand a kind of work to the AI, or when you want something to stop for good.

## What you'll need

- A **tenant admin** login. The page is at **Settings (the gear icon) → Agents & AI → AI autonomy** *(sign-in required)*.

## Choose what each kind of action may do

1. Open **Settings (the gear icon) → Agents & AI → AI autonomy**.
2. Find the kind of action under **CRM automations** or **Record actions**.
3. Choose **Allowed**, **Ask first** or **Never**. The sentence under the kind's name changes at once to say what it will do with your choice.
4. Click **Save changes** in the bar at the bottom of the page. **Discard** puts back what was saved.

| Choice | What happens |
|---|---|
| **Allowed** | The action runs without asking. It is recorded and listed under **Handled for you** in the AI Inbox, where it can be undone or marked as wrong. |
| **Ask first** | The action waits for a person in the AI Inbox. This is the default for every kind, and how every workspace worked before this page existed. |
| **Never** | The action does not run. Each time it would have, it is recorded as declined by this setting, so you can see what was stopped. |

While a kind asks first, **Once it has earned trust** offers **Keep asking** or **Run after the review window**. The second lets the action run on its own once it has earned trust, after a review window during which anyone can decline it. A linked conversation can also **Link right away**. A kind earns trust once enough of its decisions are in and most were accepted. A run of declines, undos or wrong marks moves it back to asking. The page's right-hand column gives your workspace's numbers.

**Record actions** (updating a field, creating a record, moving a stage, proposing a booking, adding or removing records on a list) are listed so you can decide before an agent uses them. The first in use is **Add or remove records on a list**: when you select contacts or leads and ask the assistant to add them to a list, take them off one or make a new list of them, **Ask first** shows a card in the chat to confirm, **Allowed** makes the change straight away, and **Never** declines.

Every list change that ran, confirmed or allowed, is listed under **Handled for you** in the AI Inbox, with **Undo**:

- An addition takes off only the records it added. Records that were on the list before stay.
- A removal puts back the records it took off.
- A new list is deleted, but only while nobody has renamed it or changed who is on it. The records themselves are kept either way.

Undo needs the same edit permission as the change. The person who asked sees their own changes there even if they cannot open the CRM approvals. Undo is not offered for a change of more than 5,000 records, and the row says why; change the list on its page instead.

## Let messages send without asking

Under **Messages to your contacts**, **Let messages send without asking** is off by default. While it is off, every email, text or call an automation would send to a contact asks a person first, even for a kind set to Allowed; that row says so in amber. Turn it on, and save, to let message kinds set to Allowed send. The page shows who changed it and when.

## Count decisions on what the AI filled in

When the AI fills in a form or a suggestion (a contact read from a call, a logged call's notes, a list described in your words, bulk tasks each worded for its record, an owner, stage or tag suggested on a record), the person who saves it keeps, changes or clears what it proposed. That decision is recorded under the kind of action it matches: creating a record, updating one, changing a record's stage, changing a list, or creating a task.

**Count people's decisions on AI-filled forms and suggestions toward earned trust** is on by default. On, these decisions count toward each kind's earned trust alongside decisions on what the AI proposed by itself. Off, they are still recorded, but do not count. Either way, each kind's row says how many of its decisions came from forms, and how many were accepted. The section shows who last changed the setting and when.

## Choose which AI suggestions show on records

Under **AI suggestions on records**, each kind of suggestion has its own choice: **Suggested owners** (on contacts and leads), **Suggested stages** (on contacts, and on leads among lead stages: moving a lead to a contact stage is **Convert**) and **Suggested tags** (on contacts, leads and deals). The section shows who last changed it and when.

- **Don't suggest**: no chip is shown for that kind.
- **Suggest** (the default): suggestions appear as chips to people who can edit the record, and nothing changes until someone accepts one.
- **Apply on create** (owners and stages only): when someone creates a contact and leaves the owner or the stage empty, a sure suggestion fills it in. It applies only what the person creating the contact could have set themselves: they need permission to edit the record, and an owner must be a current member of the workspace. Anything they picked on the form stays, and so does any change made before the suggestion is ready. The record then shows a line such as "Owner set to Sam from a suggestion: 5 closed-won deals with 3 similar contacts", with **Undo**. The change is also on its timeline, and in the AI Inbox under **Handled for you**, where anyone who can edit the contact can undo it too. **Undo** puts back what was there, unless someone has changed the value since. A suggestion that is not sure enough is still offered as a chip.

**Apply on create** is a setting on top of the kinds below: if **Update a field on a contact, company or deal** (for owners) or **Move a deal or lead to another stage** (for stages) is set to **Never**, nothing is applied, and the section says so beside the choice. A value applied this way is recorded on the AI's trust record as allowed by your setting, so it never counts toward earned trust; an **Undo** is recorded too.

Under **Bulk tasks** in the same section, **Offer "Word each task for its record"** (on by default) decides whether the bulk task form offers each record's stored next step as its task title, and **Start with it ticked** (off by default) whether the form opens with it already chosen. Turned off, the form offers one typed title for every task.

If the owner or stage changed after a suggestion was made, accepting it changes nothing: the chip says **This changed since the suggestion**, names the current value, and the suggestions refresh.

## What no setting changes

Some rules hold whatever you choose, and the page lists them in **What no setting changes**:

- **Consent.** An automation never sends to a contact who opted out of that channel.
- **High-value deals.** An action on a deal at or above your high-value bar always waits for a person.
- **Contact touches per day.** A message past the daily number of touches for a contact (emails and texts sent in the last day, and scheduled ones) waits for a person.
- **Quiet hours.** A message allowed to send without asking waits until your quiet hours end.
- **Helpdesk floors.** Forbidden actions, sensitive topics, the AI master switch and "Talk to a human" are set on the helpdesk.

## Helpdesk and campaigns

- **Helpdesk** actions are shown on this page so every kind is in one place, but they are set on the helpdesk's own AI policy (**Helpdesk → Setup → Helpdesk Settings**, the **AI governance** section), per ticket action. A helpdesk kind whose actions are set differently reads **Mixed**. See [AI and governance](/help/helpdesk/ai-and-governance).
- **Autonomous-campaign messages** read **Set on each campaign**: each campaign's own approval setting decides when a send waits. Their approvals, edits and declines count on the AI's trust record like every other kind.

## Troubleshooting

**A kind is set to Allowed but its actions still wait in the AI Inbox.** If it sends messages, check **Let messages send without asking**. Otherwise a floor held it: the action's row in the AI Inbox names it (a high-value deal, an opted-out contact, or the day's touches for that contact).

**Something you set to Never still appears in a queue's history.** That is the record of what the setting stopped. It did not run.

## Related

- [AI and governance](/help/helpdesk/ai-and-governance): the AI Inbox, Handled for you, Undo and Teach.
- [Brain governance](/help/knowledge/brain-governance): autonomy for what agents publish to your knowledge base.