Skip to main content
Every night, Aide compares its drafts with the replies your team sent and works out what to do better. It changes one thing on its own, how the agent writes, and only after testing the change. Everything else it learns is a suggestion for a person to apply, unless you let it add to its own knowledge. See the results in AI Agent > Quality.

​
When it runs

The run starts two hours after your business hours close, in your schedule’s time zone, so the day’s last replies are in. On a day with no hours, such as a weekend or a holiday, that is 02:00. With no schedule set, it runs at 02:00 UTC. See Business hours and holidays. A workspace where Aide wrote nothing in about the last two days is skipped, unless it still has open knowledge suggestions or suggested topics to check.

​
What it does, in order

  1. Reviews the day’s drafts against what your team sent: sent as is, edited a little, rewritten or not sent. It also grades a sample of your teammates’ replies and of the replies Aide sent. These feed the Quality page (see Quality: see whether Aide’s replies meet the bar).
  2. Groups what went wrong into themes, and turns them into eval cases.
  3. Checks your knowledge for gaps: questions your team answers that your knowledge doesn’t (see Find and fix gaps in your knowledge).
  4. Replays the agent’s eval cases against its current setup (see Evals: the tests Aide replays every night).
  5. Looks for a better way to write, from how your team rewrote Aide’s drafts.
It also reviews suggested topics; see Topics Aide adds on its own.

​
What Aide changes on its own

Tone & identity. Aide may change an agent’s Tone & identity, at most one change a day in a workspace, when the new wording does all of these:
  • It changes no number in the current wording, and drops nothing about confirmation, verification, disclosure, consent or handing off to a person.
  • On recent conversations it didn’t learn from, its drafts would have been sent as is or with minor edits at least 10 percentage points more often than the current wording’s.
  • Every eval case that passed that night still passes, three times in a row.
Each change is listed in AI Agent > Quality > Learned changes with the text Before and After, marked “Applied by Aide”. Knowledge, if you switch it on. In AI Agent > Quality > Issues > Missing knowledge, an owner or admin can turn on Apply what Aide learns automatically. It is off by default, and nothing changes at the moment you turn it on: Aide applies what qualifies from its next nightly run. With it on, Aide adds to its own knowledge without asking:
  • a new internal snippet with a fact your team keeps giving, or with how your team handles a kind of request
  • an addition or a correction to a snippet Aide added itself, as long as nobody has edited it since. Aide does this instead of writing a second snippet on the same subject.
Just before applying a change, Aide looks again at the replies it is drawn from:
  • Enough of your team said it: at least two teammates, in at least two conversations. On a team of one, three conversations. A reply that is Aide’s own draft sent as written doesn’t count.
  • The replies are recent: if the newest is more than 14 days old, Aide checks the change against your knowledge again that night before applying it.
  • A dated fact has a last day: a sale’s dates, an order deadline or a limited-time offer is applied only when your team’s replies give its last day. The snippet then ends that day (see Give an article or snippet an end date). Without a last day, the suggestion waits for you under “This looks time-limited”, where you can set the day or add it as it is.
  • It doesn’t go against what you wrote: if your team’s replies say something different from your Instructions or your Tone & identity, Aide asks you which is right, as a policy question, instead of drafting a change.
Aide applies at most 10 changes a night, those resting on your team’s most recent replies first. The rest wait for the next night, so a long list is never applied in one go. Snippets Aide adds are used in its replies and are not shown in your help center. They are listed under Applied by Aide on the same page, each with the conversations behind it. After a night in which Aide applied something, owners and admins get one notification in the bell, such as “Aide applied 7 changes”. Opening it takes you to Applied by Aide. It is one a night however many changes there were, and you can turn it off in Settings > Notifications.

​
What always waits for a person

Aide never changes on its own what a person wrote or what your customers can read. These stay suggestions until someone applies them:
  • a correction to an article (Apply edit)
  • publishing a draft article (Publish)
  • a change to a snippet someone wrote or edited
  • anything for your Instructions (Add as instruction)
  • policy questions, where your team’s replies disagree

​
Undo

  • Learned changes: select Undo to put the previous Tone & identity back. The change moves to the Undone filter.
  • Applied by Aide: select Undo to remove the snippet or put the passage back.
Your team’s edits always win. If someone edited Tone & identity or the snippet after Aide’s change, Undo is refused and says so, rather than overwrite their work. Edit it directly instead.

​
When Aide undoes its own change

For 3 days after a Tone & identity change, Aide watches two numbers. If the share of drafts your team sends as is or with minor edits falls by 10 percentage points or more, over at least 20 reviewed drafts since the change, or the eval pass rate falls by 10 points or more, Aide puts the previous wording back. Owners and admins get an in-app notification, Aide undid a writing change. See Notifications and email preferences.

​
What it costs

Nightly learning uses Aide’s own budget. It doesn’t count towards your AI replies or your daily AI usage limit. It doesn’t run while AI is paused on your plan.