How to Use AI Reply Mode: Auto-send, Drafts, and Confidence-Based Review

The hardest part of running an AI agent is not training it — it is trusting it. The first week your AI is live on WhatsApp, you want to read every reply before it goes out. Six months in, when the agent has answered ten thousand questions correctly, you want it to ship replies instantly so customers do not wait.

Most chatbot platforms force you to pick one mode and stick with it: full auto, or fully manual. Botline lets you switch on a card click — and even mix the two with a confidence threshold so the AI sends straight when it is sure and pauses for your eyes when it is not.

This guide walks through all three modes, when to use each, and how to review drafts in the Triage inbox.

AI reply mode lives on the Settings → AI Configuration tab. The picker looks like this:

Pick whichever fits the level of trust you have today. You can change it any time, and the new mode applies to the very next message — no restart, no redeploy.

Auto-send is what most platforms call "production mode." The AI generates a reply, sends it on WhatsApp, and you see it in the inbox after the fact. This is the right mode when:

• You have already reviewed your knowledge base and tested edge cases (see our setup guide).
• You have run the agent in Always draft mode for a few days and the corrections you make are minor or rare.
• Speed matters more than perfection — for example, a busy ecommerce DM where customers expect replies in seconds.

The trade-off: if your AI hallucinates a price or makes a promise you do not honour, the customer sees it before you do. Pair this mode with the contact-pattern scrub and auto-pause on negative sentiment safeguards (both on by default in Pro plans) to limit blast radius.

In Always draft for review, the AI does the hard work — drafts a reply, attaches its reasoning and confidence — and parks the message in your inbox. Nothing goes to the customer until you click Approve, edit and approve, or Discard.

The Triage inbox theme is the best place to work in this mode. Each conversation with a pending draft shows a yellow Awaiting review badge, and the right pane previews the draft alongside the customer's last message:

Use Always-draft for the first 2-7 days after onboarding, after a major knowledge-base change, or when the stakes of a wrong reply are high (legal, medical, regulated industries). Most tenants live here for a week, then graduate to the third mode.

This is the sweet spot for most established tenants. The AI scores its own confidence on every reply (a number from 0 to 1). If the score is above your threshold, the message ships. Below the threshold, it lands in the draft queue.

When you pick this mode, a slider appears letting you set the threshold:

Recommended starting points by industry:

• Ecommerce / general retail: 0.80 — most product questions are fact-checkable, and confidence scores are reliable.
• Restaurants / food delivery: 0.75 — menu and hours questions are easy; allergen-related queries should still draft.
• Bookings / appointments: 0.85 — date/time mistakes are costly. Higher threshold pushes more borderline replies to your queue.
• Legal / medical / financial: 0.95 — rare auto-sends, almost everything drafts. Some teams stay in Mode 2 entirely for compliance.

The confidence score is tenant-aware — Pro plans use Claude Sonnet 4.6 which produces calibrated scores, while Starter plans use Nova Pro which is slightly more conservative. Both work. Adjust the threshold based on what you observe in the first week.

1. Open Settings from the dashboard sidebar.
2. Click the AI Configuration tab.
3. Find the AI reply mode card (the sparkle icon).
4. Click the radio next to the mode you want. The card you select highlights green; the mode applies to the very next inbound message.
5. If you picked Auto-send if confident, drag the threshold slider that appears below.

There is no save button. Botline persists the change as soon as you click. You can verify by sending yourself a test message on WhatsApp — drafts show up in the inbox immediately, auto-sends fire on the customer's chat within 2 seconds.

When you have one or more pending drafts, the dashboard surfaces them in three places:

• Inbox sidebar: a yellow chip ⏱ Pending 2m on each conversation with a draft.
• Triage theme: a dedicated "Awaiting your review" lane at the top of the conversation list. Switch via Settings → Conversation theme → Triage.
• Browser tab title: the count appears in the page title (e.g. (3) Conversations · Botline) so you do not miss them while working in another tab.

For each draft you can:

• Approve & Send — sends the AI's text exactly as drafted.
• Edit — opens the draft in the reply box. Edit any wording, then send.
• Discard — drops the draft entirely. The conversation returns to "needs reply" state and the AI is paused on it for the next 30 minutes (so it does not just regenerate the same reply).

Drafts that sit unreviewed for longer than your Response SLA setting (Settings → Response SLA, default 120 minutes) light up red in Triage, so urgent ones never slip.

• Start in Mode 2 (Always draft) for the first week. You learn what your AI gets wrong before any customer sees it. Treat it like training data.
• Move to Mode 3 (confidence-based) once you are approving 90%+ of drafts unedited. The signal is clear: the AI is safe enough to ship most things on its own.
• Use Mode 1 (Auto-send) only after Mode 3 has run cleanly for a month. The confidence-gate is a free safety net — turning it off should be a deliberate choice, not the default.
• Lower your threshold during sales / promotional periods. If you launch a Friday flash sale and want fast replies, drop from 0.85 to 0.75 for the day. Bump it back up Monday.
• Pair with the Triage theme. The classic theme works for Mode 1 (auto-send) but Triage is purpose-built for review queues. Switch when you switch modes.
• Pin two team members to draft review. One person answering drafts means a 2-hour outage when they go to lunch. Two people share load and create accountability.

• Setting the threshold to 0.99 in Mode 3. The AI rarely scores that high — you will end up reviewing nearly every reply, which is just Mode 2 with extra clicks. Start at 0.80-0.85.
• Switching to Auto-send the day you launch. Without a baseline of how your knowledge base performs, you have no way to spot regressions. Always start drafted, always test for at least three days.
• Editing every draft "to make it sound better." If you find yourself rewriting tone consistently, the system prompt is wrong — not the draft. Update the agent's prompt under Agents → [Your Agent] → System Prompt; do not micro-edit drafts.
• Discarding instead of editing. Discarding loses the AI's draft entirely and pauses the agent for 30 minutes. If the draft is 80% right, edit and send. Save Discard for outright wrong replies.
• Forgetting to set Response SLA. Without an SLA, drafts sit invisibly for hours. The default 120 minutes works for most SMEs; restaurants and on-demand services should drop to 15-30 minutes.

AI reply mode is the dial that turns Botline from "fully automated chatbot" into "AI-assisted human team." You decide where on that dial you want to live, and you can move the dial as your trust in the system grows.