Using AI sessions — Getting started

Throughout the app, you'll see places where you can chat with an assistant to get help with your work. This is called a "session." This guide explains what a session is, where to find one, and how to get the best results out of it.

What a session is

A session is a conversation with the app's built-in assistant. You type what you need in plain language, and the assistant can look things up for you, draft content, or — when you ask it to and approve the result — make changes in your account. Think of it like messaging a capable assistant who has access to your contacts, campaigns, and settings, but who always checks with you before anything goes live.

You don't need any special phrasing or commands. Just describe what you want the same way you'd explain it to a colleague.

Finding things you've already made

Ask for content by what it's about, not just by its exact name — "find the post about pricing objections", "which template is our welcome email", "attach the Q3 board report". The assistant searches titles and summaries across blog posts, emails and templates, files, web pages, social posts, images, segments, campaigns, marketing events and automations, and shows you what matched before opening anything. You can also ask for content by the conversation it came from — "the deck from the webinar session" — and the assistant lists the main things that session produced. What it found comes back as clickable chips — including the conversation itself, so you can open the session the content came from straight from the reply. Questions about numbers — how many, what rate, which trend — are answered from your data directly, and you can combine the two: "how many blog posts cover pricing?" finds the posts, then counts them.

Where to start a session

You can start a session in a few places, and it works the same way in all of them:

Pulse. Click "New report" on your home dashboard to start a session. This is a good place to start when you're not yet sure what you need — you can ask general questions, get a summary of what's happening in your account, or kick off something new.
A campaign's chat. Open any campaign and you'll find a chat alongside it. Use this when your request is specific to that campaign — for example, asking for a draft of one of its emails, or a change to its schedule.
A segment's chat. Open a segment and you'll find a chat there too. Use this to describe in plain language who belongs in it — the contacts of a contacts segment, or the organizations of an organization segment — and the assistant will help translate that into the conditions that define the segment.

Wherever you start, the assistant behaves the same way and follows the same rules: it drafts, it asks when something is unclear, and it never sends or publishes anything without your review.

Finding all your tasks

Every session you start anywhere in the app — from Pulse, a campaign, or a segment — is saved as a task you can come back to later. Click All tasks to see every one of them in a single list, across all of your campaigns, with whichever task had the most recent activity at the top.

Each task in the list shows the artifacts it produced — an email, an image, a landing page, and so on — as small chips underneath it, so you can jump straight to what the assistant made without reopening the whole conversation. The list loads more tasks automatically as you scroll, so you don't have to wait for everything to load up front.

To find a specific task, use the search box in the top-left of the page. It matches against a task's title and its most recent message, searching across all of your tasks as you type.

The navigation panel lists the campaign's tasks beneath the campaign while you're in a task — up to seven, then View more opens the full list inside the panel, where a filter box narrows it. Clicking another task keeps that list open. Use Back in the top bar to return to the campaign.

Seeing who did what

When more than one person on your team works in the same session, each message shows the name of the teammate who sent it, right next to the time it was sent. The same applies when someone answers one of the assistant's clarifying questions, or approves or rejects an action the assistant asked to take — the transcript records who made that call. This makes it easy to see, when you reopen a session later, who asked for what.

A few things worth knowing:

Attribution follows the person, not the session. If you start a session and a colleague later picks it up, their messages are attributed to them, not to you — every action is credited to whoever actually took it.
The assistant's own automated activity (for example, a scheduled routine running on its own) is shown as the automation rather than a person.
Older sessions from before this feature was introduced may show some messages without a name.

Pointing at your own data

Type @ anywhere in the message box and a search list appears. Pick something from it and it becomes a chip in your message — a specific contact, organization, segment, email template, campaign, automation, image, saved file, social post, web page or form, or an email you've already sent.

This is the difference between the assistant guessing and the assistant knowing. "Send an email to our newsletter people" makes it work out who you mean and probably ask. "Send an email to @Newsletter subscribers" points at one exact segment, so it just goes.

Reach for it whenever your request refers to something that already exists:

Who — "send email to @Newsletter subscribers using template @Welcome email"
What to build from — "create an email template that includes the information from @Q3 research" (a saved file)
What to reuse — "create a LinkedIn post about our launch. Reuse the image @Launch hero rather than generating a new one."
What to copy — "Clone @Monthly newsletter into this campaign and adapt it"

Throughout this documentation you'll see examples written as @[Segment name] or @[Template name]. That's a stand-in for the chip you'll create by typing @ — not something to type literally.

How to write a good prompt

The clearer your request, the better the result. A good prompt usually covers three things — and two more that are easy to forget:

Who it's for. Which contacts, segment, or audience is this message or piece of content aimed at? @-mention them and there's no ambiguity.
What you want. What should it say or do? What's the goal — get a reply, drive a click, share an update?
Any constraints. Tone, length, which channel (email vs. text), any specific offer or detail that must be included, anything to avoid. For anything visual, layout counts too — "one column, no images" produces a very different email template from the default.
Name it. The name of a template, form, page or file is yours to choose, and choosing it up front saves renaming later: "create a web form called 'Contact Us' that captures…".
Ground it. If the material already exists, point at it rather than re-describing it — a URL, or an @-mention of a saved file, template or image. "Summarize the posts at [url] into a newsletter template" beats pasting the posts in.

Here's the difference in practice:

Vague: "Write an email."

Better: "Draft a re-engagement email for @Dormant 60 days. Keep the tone friendly and casual, not salesy. One clear call to action: book a call with us. Under 150 words."

The second version tells the assistant who the email is for, what it should accomplish, how it should sound, and how long it can be — so the first draft is much closer to something you can actually use.

You don't have to get every detail right the first time. If your first message is short, the assistant will often ask a follow-up question rather than guess.

Starting from an example

Every message box has a + button beside it — except the journey assistant's, in the rail next to the journey canvas. Open it and choose Examples to browse a list of starter prompts grouped by what they produce — email, email design, SMS, web pages, social, images, content, presentations, research, insights and more — with the ones relevant to the screen you're on listed first. There's a search box if you'd rather type than browse. (The same + menu is where you'll find Fields and Import contacts file — except in a segment's chat (where you describe or refine a segment), whose + menu has Examples and Fields only.)

Each example shows its name and a preview of the prompt underneath. Clicking one drops the full prompt into your message box, where it's yours to edit before sending. Most contain bracketed placeholders like [pick a segment] or [topic] — replace those with the real thing (and where an example says to pick a segment or contact, that's a good moment to type @).

Clarifying questions are normal

If your request leaves something important open, the assistant may ask you one question before it starts — for example, which segment to use, or what tone you're going for. This isn't a sign anything went wrong. Just answer the question, and the assistant will pick up where it left off and continue the work.

Stopping a task that is taking too long

While the assistant is working, the message box shows a Stop button. Press it and the assistant stops where it is — nothing further is created, and anything it already saved stays saved as a draft. You can then send a new message in the same session to steer it differently.

Long tasks are normal: research, a multi-step campaign, or a page with several generated pieces can legitimately run for several minutes, and closing the tab does not interrupt the work — reopening the session picks the progress back up.

During a long quiet stretch — say, while a big email design is being composed — the activity panel shows a small "Still working" line with a running timer, so you can tell the difference between a task that is genuinely busy and one that has died. If the AI service behind the app is briefly overloaded, you may also see a note like "AI provider is busy — retrying (attempt 2 of 5)": the assistant is waiting out a temporary slowdown on the provider's side and will continue on its own; there is nothing you need to do. Likewise, if the app briefly can't reach the AI service at all, the line reads "Reconnecting to the AI service… (attempt 2 of 6)" — the assistant keeps trying for about two minutes before it gives up on that step.

But a task will not run forever. If one is still going after about twenty minutes without finishing, it is stopped automatically and the session shows an error instead of a spinner, so you are never left watching a task that has quietly stopped making progress. When that happens, a narrower request usually succeeds: ask for one thing rather than several, or split the work across two sessions.

If the connection drops

Now and then a session can't get through — either your own connection to the app drops (for example your Wi-Fi cuts out), or the app reaches you fine but briefly can't reach the AI service behind it. When that happens the session shows a small card explaining which of the two it was, with a Try again button, instead of a raw error.

You don't lose your place. The conversation so far, and anything the assistant already produced, are saved — and if the drop was only on your side, the assistant often keeps working in the background while you're disconnected. Press Try again and the session picks up where it left off: it reconnects to a task that is still running, or restarts the step that failed.

If the app had confirmed your message (the reply had started), Try again never sends it again — it only reconnects or restarts the failed step.
If the app never confirmed it, Try again sends your message again. If the app did in fact receive it, it recognises the message and reconnects to that task instead of starting a second one. If an earlier task is still running in the session, your message is kept and the card stays up — press Try again once that task has finished to send it. In a rare case — the connection dropping at the exact moment the app was setting the task up — the same message can still be handled twice.

If the connection drops just as a task is starting, the card says This task didn't start — your message is saved, and Try again starts it.

If a step failed because the AI service couldn't be reached, the card comes back when you reload the page (for up to an hour, and only for your latest message), so you can still press Try again from there. Sessions that generate campaign plan items in the background don't offer Try again — regenerate the item from the plan instead.

If the app itself still can't be reached when you press Try again, the card says so — "Still can't reach the server — retrying in 5s…" — and tries again on its own a few more times over the next minute. If none of those get through, it shows "Couldn't reconnect. Try again in a moment." and the button is yours again. You can also dismiss the card and carry on by typing a new message.

What the "Working" panel shows

While the assistant works, the activity panel has a Working section you can open to follow along. When the assistant runs on Sonnet 5.5, it fills with the model's summarized notes: short progress updates, such as that it is about to look up last month's send results, and brief summaries of its reasoning. It is a running log of what the assistant is doing, not a transcript of everything it thought. The reply itself appears in the conversation; the last entry in the panel often repeats it as the assistant wraps up.

One entry per step. Each note is its own entry, typically one for each round of work, so a longer task reads as a short list of steps rather than one long block of text.
Specialists appear too. When the assistant hands part of a task to a specialist, the specialist's notes show up as entries of their own, alongside the assistant's.
Nothing to switch on. The notes appear on their own, and showing them doesn't change what a request costs.
Live only. The notes are there while you watch the assistant work. When you reopen the session later, you see the conversation and its results, not the notes.

If the assistant declines a request

Once in a while the AI model behind the assistant declines a request. Its provider runs a safety check on every request, and that check occasionally flags ordinary work by mistake. When that happens the session shows a message saying the AI model declined this request, instead of an answer. If the check says what it flagged, the message names the area — "cybersecurity", for example — and it sometimes ends with a note from the provider.

Nothing is lost: anything the assistant already saved in the session stays saved, and the declined step creates nothing. Rephrasing often helps — say plainly what the work is for, for example "a product launch email for our customers". If it keeps happening, contact support.

A few things worth knowing:

A backup model can step in. If a backup model is set up for your workspace, the assistant usually switches to it on its own and carries on, so you won't see the message. You still see it when the backup model declines too — the message then says so and names the backup — and for a few areas, such as biology, the assistant doesn't try a backup at all, so the message appears straight away. Contact support if you'd like a backup model set up.
One step of a bigger task. When the assistant hands part of a task to a specialist — a research step, say — and only that step is declined, the task doesn't end there. The assistant is told which step was declined, so it can carry on with the rest and tell you which part it couldn't complete. If several steps are declined in a row, the assistant stops and you'll see the message.
Notes already shown stay in the activity panel. The model can decline partway through, after it has already written a few working notes in the "Working" panel. Those notes stay there; they are not taken back, and they are not part of the answer. If a backup model takes over, you'll see the first model's unfinished note followed by the backup model's notes.

Everything is a draft until you say so

No matter what you ask for, the assistant never sends an email or text, publishes content, or turns on an automation by itself. Whatever it produces — a message, an image, a set of conditions for a segment, a change to a journey — is a draft for you to review. You decide whether to edit it, approve it, or discard it. Nothing goes live, goes out, or gets activated until you take that step yourself.

Sending an email or text, or publishing a social post — whether right now or scheduled for later — always pauses for a clear approve-or-reject on that exact send. You're asked once per send; approving is what lets it go out, and rejecting stops it with nothing sent. (The email and text panels later show who approved each send — see "Seeing who did what".)

The same kind of pause comes before the assistant moves contacts from one organization to another, unlinks contacts from an organization, or merges organizations: the card lists each thing the assistant is waiting to do (for example "Move 3 contacts from “Initech” to “Globex”. Moving: “Ann Lee”, “Bo Kim”, “Cy Ray”."), and nothing changes unless you choose approve — any other answer, "yes" included, rejects it. If a question and an approval are waiting at the same time, the approval card is the one shown, only approve approves it, and the assistant asks its question again afterwards. See Organizations.

This means it's safe to experiment. If a draft isn't right, you can ask for changes, start over, or just close the session — nothing will have happened in your account.

What can you ask it to create?

The assistant can help you produce several different kinds of content and information — emails, text messages, images, web pages, social posts, blog posts, presentations, and more. For a full rundown of what's available and what each one is useful for, see Artifacts overview.