Autonomy
Decide what your agents may do without asking you first, one class of action at a time. Drafting is on. Anything that reaches a person or changes your CRM waits for you until you say otherwise. Everything an agent does on its own is listed, and most of it can be undone.
Overview
An agent that can only draft is a writing tool. An agent that can act is staff. Settings, Autonomy is where you decide which one you want, and you decide it per kind of action rather than per agent, because the risk lives in the action.
Nothing changes until you change it. A new account drafts freely and asks first about everything else, which is the safe posture and the one you keep by doing nothing.
The four things an agent can do
Four classes, ordered by how far the blast radius reaches. A taxonomy with thirty verbs is one nobody configures correctly.
- Write drafts. Produce a draft, a brief or a document for you to read. Nothing leaves your account. This is on by default, because it is the product.
- Send to people. Send an email, or post to your Slack. A person receives it. An email cannot be unsent; a Slack message can be deleted, which is why Slack is the easier one to let an agent do unattended.
- Update your CRM. Write a next step, an activity or a field back to your CRM. This changes the data your forecast is built from.
- Share externally. Publish or share something with people outside your company. Public or customer-visible, and hard to take back.
A bad draft wastes a minute. A bad CRM write corrupts the numbers every forecast reads and is often not reversible from inside the product. That is why the two are separate settings even though both are writes.
The three answers
- Ask me first. The agent prepares the action and it waits in your queue. You approve it or discard it. This is the default for everything except drafting.
- Just do it. The agent acts on its own, and the action appears in the log below with an undo next to it where one is possible.
- Never. The agent is not offered the ability at all. It is not refused at the last moment, it never has it, so it does not spend a turn discovering that and does not write around it.
Setting one role differently
The buttons at the top set every role at once, which is what most accounts want. Under Per role you can override any single cell, and the narrower setting wins: allowing CRM writes for your CS agent and not your BDR is two clicks.
Choosing "Use the setting above" removes the override rather than writing the same value again, so the page can keep telling a deliberate choice apart from an inherited one.
What your agents did on their own
Below the settings is the record: every action an agent took without asking, newest first, with the role and skill that took it and when. Failures are in the list too. An autonomous write that errored is exactly the thing you need to see, and a feed of successes only would be a record of good news rather than a record of what happened.
This is not decoration. It is half of what makes turning autonomy on reasonable, which is why it is on the same page rather than somewhere you might never open.
What can be undone, honestly
- A drafted email. Yes. The draft is deleted from your mailbox.
- A Slack message. Yes. It is deleted from the channel. If your workspace does not let the app remove its own posts, you are told that rather than shown a success.
- A logged activity. Yes, on every CRM. Salesforce deletes the task, HubSpot the note, Pipedrive the activity.
- A next step. Depends on your CRM, and the page tells you which case you are in. On Pipedrive a next step is an activity, so it deletes cleanly. On Salesforce and HubSpot it overwrote a field, and the value that was there before was never recorded, so there is nothing to put back. Writing an empty string over it would be a second change you did not ask for, not an undo.
Where an action cannot be taken back, the undo button is not offered at all rather than offered and then refused. That difference is also the reason CRM writes keep asking first by default: the class the product cannot always take back is the class it should not take unasked.
An undo can also fail for an ordinary reason: you disconnected the tool, or the draft was already deleted by hand. The page says which, because "undo failed" with no reason is the least useful thing to tell someone who has just found an action they did not authorize.
It writes to whichever CRM you connected
Your agents do not have a favourite CRM. When one proposes a next step or logs a call, it says what it means and the platform sends it wherever your book actually lives: Salesforce, HubSpot or Pipedrive.
Each one spells the same thing differently, and that is handled for you. A next step is a field on Salesforce and HubSpot, and Pipedrive has no such field, so it becomes an activity there instead, marked not yet done. A logged call is a task on Salesforce, a note on HubSpot and a completed activity on Pipedrive.
If you have connected more than one, writes go to the first of Salesforce, HubSpot, Pipedrive that you have linked. If you have connected none, the action waits in your review queue and says so, rather than failing with something you would have to decode.
When something goes wrong
If an autonomous action cannot be applied, it does not disappear and it is not retried in a loop. It lands in your review queue and waits for you, which is the same place it would have been if you had never turned autonomy on. The failure of an agent acting alone is a request for your attention.
If this page cannot read your settings it says so, rather than showing you the defaults. Those are different facts, and being told your agents are asking first when nobody knows what they are doing would be worse than an error message.
Some of these have nothing behind them yet
All four classes are listed even when no agent can perform one, because the setting still applies the day one can, and because seeing the whole grid is how you understand your posture rather than only your changes.
Where nothing can perform a class, the page says so under it. Today that is Share externally: nothing publishes on your behalf, so turning it up changes nothing. The line disappears on its own when an agent gains the ability, so it can never claim more than is true.
Where to find it
Settings, then Autonomy. It sits after Connections, because what an agent may do only becomes a real question once there is something connected for it to act on.