Audiences, automations and sending with the AI agent

Writing emails is the obvious use for an AI agent. The less obvious one is everything around the email: who gets it, what happens after they get it, and when it goes out.

Nothing happens without your confirmation

Everything on this page is proposed, not done. The agent works out the details, shows you a card with exactly what it will create — the rules, the steps, the recipient count — and waits for you to hit Confirm or Cancel.

That card is a real preview, built against your actual data. An audience proposal shows how many people match and a sample of who they are. A send proposal shows the live recipient count. Confirming doesn't cost extra credits.

Audiences

Describe the segment; the agent builds the rules.

"People who've opened at least half my emails but never clicked."

"Everyone who signed up through the pricing page form in the last 30 days."

"Contacts in Germany tagged 'customer' who haven't been active in 90 days."

It uses the same rule types the audience builder offers, including your own custom fields and tags, plus Shopify and Stripe behavior when those integrations are connected.

Engagement with one specific email is worth calling out — it's the segment people most often want, and the fiddliest to build by hand:

"Everyone who didn't open the Black Friday email."

"People who clicked the link in last week's newsletter."

That matches on the email itself, across every time you've sent it — so a resend counts — and "didn't open" means people who received it and didn't open, not people who never got it. Bot opens are excluded, so the numbers agree with your reports.

You can edit an existing audience the same way: "Change my Engaged audience to 60 days instead of 30." Confirming an audience counts against your plan's audience limit, so a workspace at its limit gets told rather than silently failing. The agent can’t delete an audience — do that from your People page.

Automations

Describe the flow and the agent builds the whole thing — trigger, waits, tags, and the email bodies drafted for you.

"Build a three-email welcome series for new subscribers, two days apart."

"When someone abandons a cart in Shopify, wait an hour and send a reminder with their items, then follow up two days later with a discount."

Triggers cover everything the automation builder supports: new subscribers (optionally from one form), entering or leaving an audience, Stripe and Shopify events, and your own custom events. If you name a custom event that doesn't exist yet, the agent proposes creating it and tells you that nobody enters the flow until your app starts sending it.

New automations are created inactive, so you can read through everything before turning it on. One automation can hold up to five emails; ask for more and the agent trims the flow and tells you it did.

Reading one back. Open an automation and ask what it does, and the agent loads the real structure and summarizes it — including what each email actually says.

"What does my onboarding automation say in each email?"

Editing one. You can restructure a flow in a single request:

"Add a final email offering a discount, move the tag step to the end, and make email two shorter."

Steps you keep stay the same steps, so people currently moving through the flow keep their place. If the automation is live and the edit changes its structure, confirming pauses it so you can review before turning it back on. Deleting one shows you how many people are mid-flow before you commit.

A couple of limits: an automation with conditional branching can be summarized but not edited via chat, and SMS, webhook and update-field steps can be moved or removed but not rewritten.

Contacts

Single changes:

"Add sam@example.com with the first name Sam and tag them 'beta'."

"Tag jo@example.com as 'vip'."

Adding a contact preserves any tags they already have. Tag names are matched exactly — the agent won't guess at a near-miss and mislabel real people.

Bulk work over your existing list:

"Tag everyone who signed up in the last 30 days as 'new'."

"Remove the 'trial' tag from everyone tagged 'customer'."

"Suppress all the hard bounces."

"Suppress anyone who hasn't been active in 12 months."

You see the count before you confirm. Large bulk changes are applied in the background, so they may take a moment to show up on your People page.

Scheduling a send

This is the highest-stakes thing the agent does, and the most carefully gated.

"Schedule this to my subscribers audience for Tuesday at 9am."

"Send this to everyone tagged 'customer' but exclude the trial list, tomorrow morning."

The card shows the email, who it's going to, who's excluded, the from identity, subject and preview text, the timing, and a live recipient count. Times are read in your browser's timezone, including across daylight-saving changes — "Tuesday at 9am" means 9am where you are. If your browser doesn't tell us the timezone, the agent asks for one rather than guessing.

If anything's off, Edit send opens the normal send modal seeded with those values, so you can adjust it with the full form and its validation before confirming.

Multi-step requests

When a request has genuine stages, the agent proposes an ordered plan and works through it, using the result of each step in the next.

"Write a win-back email, build an audience of people with no activity in 60 days, and schedule it for Friday morning."

Steps it can just do — creating a draft, writing the body — it does. Steps that need your confirmation pause the plan until you confirm, then it picks up where it left off. You can watch progress from the agent bar even when it's minimized.