The AI playbook, sync health and going live

These steps happen in the partner console. The AI assistant's settings need an admin; the rest any role can read.

What the AI assistant works from

For each visitor message the AI assistant may answer, it reads:

• the profile's public and AI-only fields, as your sync sent them (guide 2), and never a private one;
• your site's live playbook;
• the profile owner's own instructions, if they gave any;
• that one conversation's recent messages.

Anything in these may reach the visitor: the AI assistant may repeat an AI-only field or an instruction word for word. Before each sentence goes out, the platform checks it. A price, phone number, email address or link that the profile, the playbook or the instructions do not contain stops the reply, and the conversation goes to the owner.

Every AI reply carries the label "<assistant name>'s AI assistant", which nothing in the playbook can remove or change.

The AI playbook

Admins shape the AI assistant for each site on the AI playbook page. Changes go to a draft, which you test and then publish as a new version.

1. Save draft. Nothing changes for visitors yet.
2. Run test. The draft answers your sample questions as the chosen profile. Visitors see nothing. Each test counts towards your monthly AI budget, and a site has a limited number of test runs an hour.
3. Publish draft. The platform first checks the draft on your sample questions and on its own test cases. If it passes, it becomes the live version. If not, nothing is published, and the page names the sample questions that failed.
4. Roll back to any earlier version from the version history. A copy of it is published as the next version.

On the same page, admins can lower the site's daily cap: the most AI replies each profile gets in a day, at most the platform's own cap. When a profile reaches it, new messages wait for the owner until the next day. Recent AI replies lists each reply with the model and the versions that produced it, never its text.

Elsewhere in the console:

• The Sites page sets each site's default for the AI switch: whether the AI assistant replies for profiles whose owners have not chosen. Owners can always switch it off or on for their own profile.
• The Channels, AI and rules page picks the model for all your sites, from the ones the platform allows. Until you pick one, the platform's default applies.
• The Usage page shows this month's AI spend against the monthly budget, which the platform team sets. When the budget runs out, the AI assistant stops replying and visitors see that the owner will reply.

Owners narrow all of this from their inbox (guide 5): a tone, frequent questions with their answers, and topics to avoid. They cannot allow a topic you forbid or switch off a hand-off rule you turned on.

When the AI assistant hands over

The conversation goes to the owner, and the visitor is told the owner will reply, when:

• the profile has no answer, or a hand-off rule of your playbook applies;
• a sentence fails the platform's checks, or the AI service does not answer;
• the profile reached its daily cap, or the monthly AI budget is spent;
• your site has not synced its profiles for too long (see below);
• the profile has no public assistant_name or name that fits the label.

For the first two, the AI assistant takes the conversation back 10 minutes after the hand-off or the owner's last reply. After a daily cap, it waits for the next day. For the rest, the owner hands it back.

Sync health

The Sync status card, on the Integration page and on the Profiles page, shows:

• Profiles imported: how many profiles the platform holds for the site, not counting deleted ones, and the total your last batch declared;
• Owners: how many owners your profiles name. Owners are counted, never listed;
• Last profile event received;
• AI assistant answers until: 7 days after your last sync request that counts, unless the platform's operator set another period. In the last day the card warns you. Once it passes, the card says the AI assistant stopped answering, and your next sync request that counts, even a reconciliation read, starts it again (guide 2, step 9 says which count);
• Rejected changes, last 24 hours.

The Profiles page goes into detail:

• Rejected changes of the last 30 days, newest first: the time, the profile ID, the change, the version, the field when the reason concerns one, and the problem. Never the content.
• Field visibility: each field name with how many profiles sent it as public, AI-only, private or without a flag, counted over your 1,000 most recently synced profiles. A field sent without a flag is private, and the page warns about it, since it is usually a mistake.
• Recently synced profiles, with their public name, or "Name kept private", and the deleted ones marked.

The Integration health card on the Integration page lists user tokens that failed a check, by reason (guide 4).

Going live

Going live is not open yet. It comes with a later release of the platform, together with the go-live checklist and request in the console. Until then every site stays in the sandbox. The Integration page already shows a go-live checklist, but only its sync step reflects your site, and its request button sends nothing yet.

To be ready for that day:

• sign widget tokens with the audience messenger-widget, since the platform refuses messenger from the first go-live (guide 4);
• list every origin that shows the messenger or your owner dashboard among the site's allowed origins (guide 1), since the platform checks them from go-live;
• keep the daily reconciliation running, so the AI assistant keeps answering.

Webhooks, such as needs_owner when a conversation needs the profile owner, and paid messages also come in later releases. Neither has a shape you can build against yet.