Social: AI on comments and DMs

Two channels, two very different engines

Instagram and Facebook arrive in Zuzana as two separate things, and the single most useful thing to understand before you switch anything on is that they are not handled by the same AI.

• Direct messages (Instagram DM, Facebook Messenger) run the full AI, the same one that answers your chat widget. Knowledge base, policies, workflows, order lookups, escalation to a human. A DM is treated as a real support conversation, because that is what it is.
• Post comments (Instagram and Facebook) run a separate, deliberately narrow agent. It has no knowledge base, no order data and no workflows. It gets exactly three choices per comment: reply publicly, take it to a DM, or leave it alone.

That difference is on purpose. A public comment thread is the worst possible place to be wrong, and the worst possible place to be right about something private. So the comment agent is built to be socially competent rather than knowledgeable, and the moment a comment turns into actual customer service it is supposed to get out of the public thread.

Switching it on 🔌

There are two layers, and both have to be right. This is the usual reason someone reports that nothing is happening.

Layer 1: the connected account. Under Settings → Instagram or Settings → Facebook, each connected account has a Configure button with:

• Market. Everything from that account lands in this market, which is what decides the market instructions and brand name the AI uses. There is no language detection step in front of it.
• Direct Messages and Post Comments. These control ingestion: whether the messages become tickets in Zuzana at all. Comments are off by default.

With only this layer on, you get a working social inbox and no AI. That is a perfectly good place to sit for a week.

Layer 2: the AI toggles, further down the same settings page:

• Auto-reply to DMs. On or off. That is the whole control.
• Handle comments with AI, plus a Response mode (Draft, Semi-Automated, Fully Automated).
• Escalate to Direct Messages / Messenger. Lets the comment agent move a case into a private DM.
• Like comments (Facebook only). Lets the AI like positive comments as the page.

Note: your team's master AI responses switch sits above all of this. If it is off, no social AI runs at all, on either channel.

What the response mode changes for comments 📝

Draft, Semi-Automated and Fully Automated mean what they always mean in Zuzana: who presses send, and what happens to the ticket. On comments they also change what the AI is allowed to say, which is unusual and worth knowing.

• Draft. Nothing is published. The reply waits on the ticket for a human. Because a person will read it first, the agent is allowed to attempt factual answers about products, shipping and services from its context, and is told to skip the comment rather than guess when it cannot.
• Semi-Automated. The reply is posted publicly and the ticket stays open for a human to look at.
• Fully Automated. The reply is posted publicly and the ticket is closed.

In both of the sending modes the agent is under a hard instruction to never answer factual questions. It has no knowledge base to answer them from, so anything it produced would be invention. Those comments go to a DM, or get passed on.

So a fully automated comment setup is not "the AI answers our comments". It is "the AI handles the friendly noise, and routes everything real into a private conversation". That is usually the correct ambition for a comment section.

Note: Escalate to DM is not available in Draft mode, and the toggle switches itself off when you pick Draft. The public half of an escalation says "we have sent you a DM", and that must not be a claim sitting in a draft nobody sent yet.

How the DM escalation actually works

When a comment needs real help and escalation is on, three things happen at once:

1. A short public reply acknowledges the person and tells them to check their inbox.
2. A private reply is sent on that comment, which opens a DM thread with them.
3. That DM gets its own full AI run, with the knowledge base, policies, workflows and order tools available.

The DM appears as a side conversation on the comment ticket, not as a separate ticket on the board, so the public comment and the private follow-up stay in one place.

Two platform facts are worth knowing here. A private reply is one shot per comment: after that first message the thread is an ordinary DM thread. And escalation needs a reachable recipient, so a comment from an account Zuzana cannot address privately will be passed on instead.

How to test it 🧪

The two channels are tested in two different ways, because only one of them can be simulated.

DMs: use the Playground. Open the Playground in chat mode. This is not an approximation, it is the same engine that answers a real Instagram DM, so what you see there is what a customer gets. Set a market, set a customer email if you want order context, and have the conversation. Anything worth keeping can be saved into a chat-mode test suite and re-run whenever you change a policy or a workflow.

Comments: test them live, on Draft. The comment agent cannot be reproduced in the Playground or in a test suite, because both of those run the chat and email engines. The safe substitute is Draft mode, which is the same idea as mirror mode: real comments, real AI, nothing published.

1. Set Response mode to Draft and turn Handle comments with AI on.
2. Comment on one of your own posts from a personal account. Your own page's comments are ignored, so it has to be someone else's account.
3. Open the ticket in the helpdesk and read the proposed reply. Edit and send it, or discard it.
4. Do this on ten real comments before you raise the mode. Ten drafts on your actual comment section tells you more than any amount of theorising about tone.

Every run on either channel is listed under More → Runs, which is where you go to see which tool the agent chose and why it did or did not reply.

How to adjust the replies 🎛️

This is where the split between the two engines matters most, because the levers are not the same.

The comment agent reads only your team and market configuration:

• Brand name, About the company and What we sell
• Tone of voice, emoji usage, verbosity, headline and bold-text settings
• Guard rails and Additional instructions
• The market instructions of the market that account is mapped to
• A description of the post's image, so the reply matches the picture people are commenting on

It does not read your articles, policies, workflows or product catalogue. So if the comment replies are too long, too formal, too eager, or keep saying something you would rather they did not, the fix goes in Additional instructions or a guard rail. Writing a policy about it will change nothing, because the comment agent never sees policies.

DMs, by contrast, are tuned like every other channel. Articles, policies and workflows all apply. If you want something to behave differently in a DM than in an email, scope it with a ticket rule on Source (Instagram DM, Facebook Message) rather than building a separate setup.

Note: inflow rules do not scope social. On email they decide which tickets the AI may touch; on comments and DMs the per-channel toggle is the whole gate. Your finer-grained controls are the per-account ingestion toggles, the comment response mode, and source-scoped policies and workflows on the DM side.

A sensible rollout

1. Connect the account, ingestion only. DMs and comments become tickets, humans answer them, you see the volume and what it actually consists of.
2. Turn on DM auto-reply. DMs are private, one to one, and backed by the full knowledge base, so this is the lower-risk half despite feeling like the bigger step.
3. Turn on comments in Draft. Read the drafts, fix the tone in Additional instructions, not in the individual replies.
4. Raise comments to Semi-Automated with Escalate to DM on. Now the friendly comments get handled publicly and everything real moves to a private thread.
5. Fully Automated only once the pattern is boring, and remember it also closes the ticket.

Good to know

• One comment is one ticket. A reply to your reply arrives as a new comment, so it becomes a new ticket rather than continuing the old one.
• Your social signature is appended to public comment replies too, not just DMs. A three-line email-style sign-off looks wrong under an Instagram post, so keep it short or leave it empty.
• Liking comments is Facebook only. Instagram does not offer it.
• The first comment on a brand new post can be answered before Zuzana has finished describing the post's image. Later comments on the same post have it.
• Names. The AI greets people by first name only when it holds a real one. It will not greet anyone by their handle, so no public "Hi vinfan_92".
• Private information is never posted publicly. Order details, addresses and account data are barred from comment replies regardless of mode.
• A claimed ticket is left alone. If a colleague has taken a social conversation, the AI stays out of it.
• DMs have no draft mode. The DM AI sends. If you want a human in front of every reply on a channel, that channel is comments on Draft, or DM auto-reply off with agents answering by hand.