The 4-tier classification that runs the inbox
Every inbound DM falls into one of four tiers. Tier 1 — Hot lead. Someone asking about your offer, your pricing, your availability, or how to work with you. ~2-5% of DMs. These get a human reply within 5 minutes.
Tier 2 — Qualified question. Specific, on-topic question from someone in your target audience. ~10-15% of DMs. These get an AI-drafted reply within 5 minutes; you approve in 60 seconds.
Tier 3 — Generic engagement. Compliments, fire emojis, "great post." ~40-50% of DMs. These get a templated heart/thumbs/short reply via auto-respond; you do not need to see them.
Tier 4 — Spam / scam / collab pitch. Self-promotional, scammy, or unsolicited pitches. ~30-40% of DMs. These get muted/archived automatically. You see only the ones the AI flagged as ambiguous.
The math: 1,000 weekly DMs collapse to ~20 hot leads + ~150 qualified questions. The hot leads you handle yourself in ~15 minutes. The qualified questions you approve in ~20 minutes (AI draft + your edit). Total active triage time: ~35 minutes/week instead of 8-10 hours.
The triage time arbitrage
Manual DM handling at 1,000 DMs/week: ~10 hours/week × 52 = 520 hours/year. AI-triaged DM handling: ~35 min/week × 52 = 30 hours/year. Net recovery: ~490 hours/year, or roughly 12 full work weeks.
The unified inbox is non-negotiable
You cannot triage 4 separate platforms — IG DMs + TikTok + Twitter/X + YouTube comments + email — by jumping between apps. The cognitive cost of context-switching destroys the system before it starts.
A unified inbox tool (Front, Missive, Spike, or a custom n8n/Zapier setup) pulls all inbound into one place, where you triage once. Tier 1 hot leads get tagged and routed to the top. Tier 4 spam gets auto-archived. You see one prioritized list, not 6 notification badges.
The 2026 specialty tools — Manychat (IG/FB), Inflact (IG/TikTok), Snipfeed (creator-specific) — are improving fast but still platform-specific. For most solo creators with 25K+ followers across 3+ platforms, a unified general-purpose inbox + AI agent layer beats specialty single-platform tools.
AI-drafted replies that do not sound like AI
The default failure mode of AI DM replies is the "Hi there! Thanks so much for reaching out!" tone. Audiences detect it immediately and feel unheard. The fix is voice configuration plus narrow scope — AI replies only handle the question; humans handle the relationship.
Voice configuration means feeding 30-50 of your past DM replies as few-shot examples. The AI learns your specific tics: lowercase casual vs sentence-case professional, emoji frequency, sign-off style, how you handle pushback. After voice config, AI drafts feel like yours.
Narrow scope means the AI does not try to be friendly or build relationships. It answers the question, points to the next step (your link tree, your booking page, your free guide), and signs off. Warmth comes from your human follow-ups on the hot leads — not from a chatbot performing friendship.
AI DM reply prompt skeleton
- Role: You are a customer-service assistant for [creator name], who [niche/credentials in one line].
- Voice examples: [paste 10-20 of your past DM replies; one per line]
- Rules: Reply in [creator name]'s voice. Keep replies under 60 words. End with a clear next step (book a call / read this guide / DM keyword X). Do not improvise on pricing, availability, or commitments. If the DM is ambiguous or sensitive, output [ESCALATE] instead of a reply.
- Input: [paste the inbound DM here]
- Output: Either a reply draft OR [ESCALATE] with a one-line reason.
The 5-minute SLA + the escalation rule
The whole system depends on a strict SLA on tier 1 + tier 2 DMs. If the hot lead does not get a reply within 5 minutes, the system has failed. To enforce this, the unified inbox + AI agent route tier 1 + tier 2 DMs to your phone via SMS or push notification with the AI draft already attached. You approve or override in seconds.
The escalation rule that keeps you sane: anything mentioning refund, lawyer, threat, specific person's name (not yours), or "I am" + a sensitive identity claim — gets escalated to human, even if the AI is confident. The cost of an AI mishandling these is ~100× the cost of a 10-second human override.
Tier 3 + tier 4 do not have an SLA. Tier 3 auto-replies fire instantly with a templated heart/thumbs and a "you might like X" link. Tier 4 auto-archives. You glance at the muted folder once a week to confirm no real leads got buried.
Compliance: AI-DM disclosure rules in 2026
Instagram and TikTok are rolling out platform-level AI labeling for commercial DMs in 2026. The requirements vary and change quickly. The safe pattern is a one-line footer on AI-generated first-touch replies: "Replies sent via [creator] AI — a human will handle next steps." This satisfies disclosure expectations and explicitly sets the right user expectation.
FTC: 1:1 customer-service replies are not commercial endorsements and do not trigger FTC disclosure rules. If your AI agent recommends a specific paid product or affiliate, the recommendation itself triggers the same disclosure as a human-recommended affiliate — #ad or "affiliate link" is required.
AI DM disclosure quick rules (2026)
Platform: IG + TikTok require labeling AI-generated DMs in commercial contexts in 2026 (rollout ongoing; check current platform docs). FTC: no disclosure for 1:1 customer service; full disclosure if AI recommends affiliate/sponsored products. Safe practice: one-line footer on AI first-touch replies, full disclosure on any recommendation.