Labelling conversations automatically, and handling Facebook and Instagram comments in one place

The problem is not the number of messages, it is that you do not know what is in them

Every inbox shows the same list: a name, a last message, a time. That is enough for one person who opens every conversation anyway. It is enough for nothing beyond that — and it falls further short when the list becomes gathered from three channelsrather than one.

• A manager cannot answer a simple question: “what do our customers ask about most?”. The answer lives inside thousands of messages, and nobody has time to read them.
• A complaint looks like a price question in the conversation list. The first needs a reply within minutes and the second can wait an hour, and the list does not tell them apart.
• Nobody knows where the time goes. “Average reply two hours” is a number that fixes nothing. “Average reply to cancellation requests six hours” is a number that fixes something today.
• Manual tagging is abandoned within a fortnight. Any field an agent is asked to fill in after every conversation is filled for the first two days and then left — and the report built on it becomes worse than no report.

The answer is not a new field for the team to fill in. The answer is for the enquiry type to be read from the customer’s own text, the moment the message arrives.

Ten types, and the platform is what decides them

Every conversation carries one type, visible in the list, drawn from a fixed list of ten. The list is deliberately commercially neutral: a clinic, a shop and an events venue all read the same vocabulary and find it correct.

The four families are not decoration. The colour in the interface stands for the family rather than the type, because ten distinguishable colours cannot actually be told apart by somebody with colour blindness — whereas four families can.

The one rule that stops the tag going wrong

The dangerous thing about automatic labelling is not that it gets one wrong, but that it replaces a correct tag with a doubtful one. So writing to the conversation is governed by two conditions only, and everything else means: leave what is there.

Sentiment (positive / neutral / negative) is read in the same call and recorded on the message. But it is not shown as a verdict on a particular conversation — only as a trend across dozens of them — because tone is harder to read than enquiry type, and courtesy in the Gulf is ordinary politeness rather than satisfaction.

What a manager can see and could not before

Once every conversation has a type, the conversation list becomes filterable by type, and behind it sits one report answering three questions: what they asked about, how long they waited, and how they felt.

Every number drawn in the report is repeated in a table at the foot of the page, and both come from the same calculation rather than two queries — because a table built with a second query is how a chart and a table begin to disagree. And the whole report exports to PDF from those same numbers.

• Ready periods: the last 7 days, 30 or 90 — with a manual date range available.
• Filtering by channel and by enquiry type, in the same vocabulary the inbox uses.
• If you ask for a period further back than your plan’s analytics retention, the period is shortened and the page tells you, rather than quietly showing incomplete numbers.
• An automatic reply inside a flow counts as a reply — because a customer who received an instant answer considers themselves answered.

The second front: comments left unanswered for a whole day

The customer who writes “how much?” under an Instagram post is the same customer who, had they messaged you on WhatsApp, would have had a reply within a minute. The difference is that a comment sits somewhere nobody opens regularly, so it stays visible to everybody with no answer — and the ten people who read it saw a page that does not reply.

The comments inbox deals with that by gathering Facebook and Instagram post comments into one page inside the platform, which they reach the moment they are written through Meta’s notifications, not when somebody opens the app.

Comments from customers are read the same way conversations are, but in a vocabulary better suited to a public post: six intents rather than ten enquiry types.

The page itself is divided into ready folders that collapse the sorting: “needs follow-up”, “interested”, “questions”, “complaints”, “assigned to me”, “done”, “hidden” and “all comments” — each with a live counter. And every comment carries a set of actions: a public reply, a private message, hide or unhide, mark as done, delete, assign to a colleague, and add its author as a lead. There are optional AI helpers too: an editable draft reply, a translation of the comment, and a summary of the whole discussion under the post.

From a public comment to a private conversation: the route that works

The pattern that works is neither “reply in public” nor “move it to private”, but both, in that order. The public reply shows the other readers that the page answers; moving to private takes the details of the request — the name, the number, the amount — out of a place everybody can see.

This route is the same logic as click-to-WhatsApp ads — a move from a public audience to a private channel you own — except it starts from content you published for nothing rather than a click you paid for.

What is actually available today, and what is waiting on Meta

This section exists because the article loses all its value if it leads you to expect a screen you cannot open. Both capabilities are built and tested, but they are not equally available.

The practical order in the meantime: turn conversation labelling on for your WhatsApp channel now — it is most likely where most of your volume sits — and read the report for a fortnight before you change anything about how work is distributed. And if you are connecting your existing number for the first time, start from the route that connects without losing the WhatsApp Business app.

The practical summary

• The tag is read from what the customer said rather than asked of an agent — because every manual field is abandoned within a fortnight.
• The governing rule is “do nothing”: a greeting, a photo, an unconfident reading and a technical failure all leave the previous tag in place.
• An agent’s correction locks the tag permanently. That is what makes correcting it worth the trouble.
• The number you act on is “first reply time by enquiry type”, not the overall average.
• Sentiment is read as a trend over time, not as a verdict on a customer — and Gulf courtesy is not satisfaction.
• A comment is answered in public and then moved to private, and the window for moving is seven days and one message only.
• The comments inbox is built but closed until Meta approves the comment permissions; Instagram and Messenger direct conversations are available today, according to plan.