Playbooks · 5 min read
Social Support Desks: What Instagram Structurally Allows
The usual version of this question is whether AI will replace human support agents, and the usual answer is a hybrid model with a percentage attached. We do not know, we are not going to invent the percentage, and the question is the wrong shape anyway.
Here is a shape that can be answered. A support desk is not one thing. It is roughly nine things bolted together, and each of them can be checked separately against two questions: does the platform permit it, and does the software actually do it. Some of the nine are structural on Instagram and will still be true in a long time. Some do not exist here at all. Knowing which is which is more useful than a forecast.
If you want the version of this argument about the role rather than the desk, that post is separate. This one is about the machinery.
The nine parts, checked
Intake. Messages, comments and story replies arriving in one place. This exists, and it is the least interesting part.
An SLA. This is the strange one. In every other support channel, the response deadline is a policy somebody chose and can relax. On Instagram it is imposed by the platform: twenty-four hours to reply inside a DM thread, restarted only by their message, and seven days to reply to a comment on your own post. Nobody negotiates that, no plan tier extends it, and it is the hardest SLA most small businesses will ever operate under. It is also the single strongest argument for automating the repetitive part — not speed as a nicety, but a deadline you did not set.
Macros. Stored answers you wrote, sent as-is. Here they are templates, they are free and unlimited at any volume, and they are the correct answer for every question with exactly one correct answer.
Routing and assignment. Does not exist here. No ticket ids, no statuses, no queues you assign to a colleague, no round-robin.
Escalation. Exists, in one specific and deliberately crude form: takeover. The moment a human replies by hand in a thread, automation is out of that thread. No rule to configure, no escalation matrix, no state to set. It is the seam, and it is one-way.
Quality control before send. The review queue holds replies the system was not confident about, so a person approves, edits or discards rather than apologising afterwards.
An audit log. Activity records what was sent and, for what was not, which of the ten checks refused it and in what order. kill_switch, connection, takeover, window, dedupe, cooldown, quiet_hours, rate_budget, credits, content_safety.
Satisfaction measurement. Does not exist. No CSAT survey, no rating prompt, no sentiment score attached to a resolved conversation.
Reporting. Does not exist. There is no analytics product here — no first-response-time chart, no resolution rate, no volume trend. There are three stores you read yourself: Activity, Leads, and the credit ledger.
Four of nine exist, one is imposed by the platform, and four do not exist at all. That is a more useful sentence than any prediction about agents.
What would have to change for the missing four
Not a roadmap. A condition, for each, so you can judge for yourself how likely it is.
- Ticket state and assignment. Would need a durable conversation object with an owner, a status and a history that survives the reply window closing. That is a product decision, not a platform limit — which means it is the most buildable of the four and still is not built here.
- Real first-response-time reporting. Would need timestamps joined across intake and send for every conversation, and a definition of "responded" that counts a template the same as a human sentence. Ask any vendor showing you this number which of those two it counted.
- Satisfaction scores. Would need a survey mechanism inside a messaging window that exists to protect people from being messaged. Sending a rating prompt uses up the same twenty-four hours the customer's actual answer needed.
- Cross-channel resolution. Would need the same person identified on Instagram and wherever else they contacted you. There is no WhatsApp, Messenger or Telegram here, and no integration that joins identities. Leads leave as CSV; the joining happens in whatever you open the CSV with.
The split that holds regardless
Whatever happens to the tooling, one division has stayed stable and it is not a technology question.
Automate answers that are true regardless of who is asking. Hours, price, delivery areas, the link, what to bring to an appointment, whether you are taking new clients. These are facts about you, and a stored sentence is more accurate than a fresh composition.
Route to a person anything whose answer changes if you knew one more thing about the sender. Refunds, complaints, anything about someone's health or money, anything phrased as "should I". A model will produce a fluent and general answer to a question whose real answer was "it depends" — and it will not know it has crossed the line.
The honest bet
Not a forecast, just the direction the incentives point. The repetitive half of a support inbox has been getting cheaper to answer for years and there is no mechanism pushing back the other way. The judgement half has not moved, because the thing that makes it hard is not writing.
What we would not bet on is any of the four missing parts arriving because a model got better. None of them is a prose problem. The support-hours version of this covers what round-the-clock actually means under a twenty-four hour window, and where a bot should stop answering entirely is the line-drawing post.
