Chat SDK's Instagram adapter: the 24-hour window is the spec

The Chat SDK Instagram adapter shipped today, and the headline feature is not the one that should change your plans. Bots can now send and receive DMs and media, render quick replies and link buttons, show typing indicators, receive reactions, and handle story replies through Meta's Instagram Messaging API.

That's the announcement. I read the Vercel changelog entry, and the sentence that actually decides your architecture is near the bottom: Meta enforces a 24-hour messaging window.

---

🔒 The 24-hour window is the product spec, not a footnote

Per the changelog, bots "can only reply within a day of the user's last message." That single rule quietly deletes about half of what people imagine when they hear "Instagram bot."

What you might want to build	Inside the 24-hour window?	Verdict
Answer "how much is delivery to Kandy?"	Yes	Works
Take an order and confirm it in the same chat	Yes	Works
"You left something in your cart" three days later	No	Dead
"Your parcel shipped" 48 hours after ordering	No	Dead
Daily tuition-class reminder	No	Dead
Win-back campaign to last year's customers	No	Dead

So this is a reactive support and sales channel, not a marketing channel. Every conversation is started by the customer, and your bot's job is to finish it before the clock runs out.

Key takeaway: Design for one continuous session, not a relationship. If your flow needs to reach the user tomorrow, that step has to move to email, SMS, or a channel with its own messaging rules — and you need to collect that contact detail during the 24 hours.

---

⚡ Buffering means your latency budget replaces your streaming UX

The other line worth reading twice: "Messages are buffered, so streamed responses send as one message when the stream completes."

If you've built web chat, you've spent time on token-by-token streaming, because watching text appear makes a slow model feel fast. In a DM, none of that exists. The user sees nothing, then sees everything. You are back to plain request/response, and perceived speed becomes entirely a function of total reply time.

These are my own rules of thumb, not numbers from the changelog:

Total reply time	How a DM feels	What I'd do
Under 2s	Normal	Ship it
2–5s	Noticeably slow	Fire the typing indicator immediately
5–15s	User assumes it's broken	Send a short acknowledgement first, answer second
Over 15s	User closes the app	Don't route that intent through a slow model

This is exactly why the adapter supports typing indicators, and why I'd treat them as mandatory rather than decorative. It also argues for a small, fast model on the first hop with escalation to a bigger one only when the question needs it.

---

💰 Every DM is an inference call, and DMs are cheap talk

Instagram DMs are high-volume and low-intent. "Available?", "price?", "delivery to Galle?", a sticker, then silence. A web chat widget gets a handful of serious sessions a day. A busy Instagram shop gets hundreds of one-word messages.

Three cost realities to price in before you build:

1. Every inbound message is a model call unless you filter. Simple keyword or classifier routing in front of the LLM cuts the bill more than any prompt trick.
2. Multi-turn chat is not linear. The whole history is resent as input on every turn, so a 20-message conversation costs far more than 20 single replies. Our AI conversation cost calculator shows how badly the naive estimate undershoots.
3. The changelog says nothing about Meta's own messaging fees for Instagram. Don't assume it works like WhatsApp. If you're comparing the two channels for a Sri Lankan business, our WhatsApp chatbot cost calculator gives you the WhatsApp side of that comparison, including the local rate card.

If your margin per order is Rs 300 and your bot burns tokens on 40 tyre-kickers per sale, the bot is the expense, not the sales channel.

---

🛠️ Where this genuinely fits a Sri Lankan small business

Instagram is the storefront for a large slice of small commerce here. Clothing pages, cake orders, phone repair, guesthouses, tuition classes. Many of them have no website at all, and the DM inbox is the checkout, the support desk, and the CRM.

Good fits for a bot built on this adapter:

• Price and availability questions at 11pm, when the owner is asleep. Same-session, high volume, boring to answer manually.
• Story replies routed into a real conversation. The changelog calls this out explicitly, and for a shop posting daily stories that's the single biggest inbound source.
• Structured intake using quick replies: size, colour, city, payment method. Buttons beat free text when you need clean data.
• Handoff to a human with the context already collected, so the owner opens a chat that's already 80% done.

Poor fits:

• Anything that must be delivered later. See the 24-hour rule.
• Payment collection. Push to a link button and take it off-platform.
• Sinhala and English code-switching without testing. Assume your model handles it badly until you've checked with real messages, not sample ones.

---

❓ What the changelog doesn't tell you

I'd want answers to these before committing a client to this stack, and the announcement doesn't provide them:

• What the professional Business or Creator account requirement means in practice for review, approval time, or permissions.
• Whether Meta charges anything for Instagram messaging volume.
• Rate limits, and what happens when a shop gets a viral post.
• How the adapter behaves when the 24-hour window expires mid-generation.

The changelog points to documentation and an adapter directory. Read those before you quote anyone a delivery date.

---

🎯 What this means for you

If you're a builder here, this lowers the cost of putting an AI agent in front of the channel your customers already use, and that's real. The adapter handles the plumbing — media, cards, reactions, typing — that used to eat your first week.

But build it as a fast, reactive, session-bound assistant. Answer inside the window, capture a phone number or email while you have permission, and move anything time-delayed to a channel that allows it. The teams that get burned will be the ones who ship a follow-up sequence, watch it silently fail, and only then read the paragraph about the 24-hour window.