Launching AI replies safely

Scrollmark can draft DM replies and comment responses using AI (internally called Mark). But "turn on AI" and "let AI send messages unsupervised" are two very different things. This playbook walks you through a safe ramp-up: start with full human approval, watch the output, and loosen the guardrails only when you trust the results.

The three layers of control

Before building anything, it helps to understand the three mechanisms that govern AI-generated messages:

All three work together. Brand Safety is the global switch; the Reviews queue is the human workflow; the per-node toggle is the fine-grained override.

Step by step

Step 1: Configure Brand Safety

Open the journey you want to add AI replies to (or create a new one). Click the Brand Safety & Variables button in the top bar and review these settings:

• Manual intervention for AI-generated messages — turn this on. Every AI-drafted DM will pause and wait for approval before sending.
• Manual intervention for AI-generated comments — turn this on if your journey replies to comments with AI text.
• Manual intervention for templated messages — leave this off for now. Templated (non-AI) messages are static text you wrote yourself, so they don't need review.
• Sentiment analysis — consider enabling this. It flags negative or hostile engagements so the journey doesn't reply to someone who's angry.
• Cooldown period — set a reasonable window (e.g. 3 days). Prevents the AI from re-engaging the same user repeatedly.
• Excluded usernames — add your own account and any team accounts so journeys don't reply to internal messages.

Step 2: Build the journey with AI nodes

In the journey builder, add your trigger and connect a Send a direct message action. In the message editor, toggle the Mark option to let AI generate the reply. Provide clear instructions:

Reply to the user's question about our product. Be friendly, keep it under 3 sentences. If you don't know the answer, say "Let me connect you with our team" and nothing else.

The special instructions field shapes how Mark writes. Be specific — vague instructions produce vague replies.

If the journey also replies to comments, do the same on the Reply to Comment action node.

Step 3: Review messages in the inbox

With manual intervention enabled, every AI-generated message pauses at the Waiting for Approval status. These show up in two places:

• Inbox → Reviews tab (Instagram) — a dedicated queue of pending AI drafts. You can approve, edit, or deny each one.
• Unified Inbox — pending messages appear with an approval badge on the conversation. You can approve directly from the conversation thread.

When you approve a message, Scrollmark sends it immediately. When you deny it, the message is discarded and the journey either continues to the next node or stops (depending on the flow).

Step 4: Watch, tune, trust

Run the journey for a few days with full approval enabled. Pay attention to:

• Tone — does the AI match your brand voice? Tighten the special instructions if not.
• Accuracy — is it inventing information? Add constraints: "Only reference information from the user's message. Do not make up product details."
• Edge cases — what does it do with sarcasm, spam, or off-topic messages? Add sentiment analysis or a condition node to filter these before the AI step.

Step 5: Gradually loosen the guardrails

Once you trust the output:

1. Turn off manual intervention for AI-generated messages in Brand Safety. Messages will send automatically without human review.
2. Keep sentiment analysis on as a safety net — it costs nothing and catches hostile engagements.
3. Keep the cooldown period so the AI doesn't flood anyone.
4. Monitor from the Unified Inbox. Even without approval, every AI-sent message is visible in the conversation history. Spot-check regularly.

You can also go hybrid: leave manual intervention on for comment replies (public, higher stakes) while letting DMs send automatically (private, lower risk).

Tips

• Start with one journey. Don't enable AI on everything at once. Pick your highest-volume, lowest-risk flow (e.g. a simple keyword-to-DM) and prove it there first.
• Write instructions like you're training a new hire. "Be friendly" is too vague. "Reply in 1–2 sentences, use the customer's first name, never promise a refund" is actionable.
• Use the human attention priority. On action nodes, you can set a priority level (Very High, High, Default, Low, Very Low) that controls where the conversation appears in the inbox queue. Set AI-generated messages to High so approvers see them first.
• Denied messages aren't wasted. When you deny an AI draft and rewrite it, you're training your own instincts about what the instructions need to say. Update the special instructions after every pattern you notice.

Related pages

• AI replies — how AI suggestions work in the inbox
• Journeys — builder basics and publishing
• Triggers and actions — message approval flow and action types