When AI replies vs when you should take over

The AI handles 80% of replies competently. The other 20% need a human. Direct REI's job is to know the difference and tap you in at the right time.

First decision: auto-send or draft?

Every email/SMS campaign has an AI-reply mode you pick when you create or edit it — the "When the AI drafts a reply" setting:

• Send automatically (default) — the AI writes the reply and sends it 2-15 minutes after each inbound, with no action from you. Best for high-volume outreach where you trust the AI to carry routine conversations.
• Draft + notify me — the AI writes the reply but does not send it. It lands in your Drafts inbox for you to approve, edit, or skip — as soon as their reply is picked up, with no 2-15 minute wait, since you decide when it sends. It still reads their reply the same way: it marks them Qualified when your Reply Goal is met, flags anything that needs you, and saves what they told you to the contact. Best for high-value campaigns or while you're still learning to trust the AI's voice.

You can run different campaigns in different modes — e.g. auto-send a big cold-email push while keeping a niche partnership campaign in draft mode. The choice only affects routine replies; the red-flag rules below override both modes.

What the AI handles automatically (auto-send mode)

When a prospect responds to an auto-send campaign via email or SMS, the AI:

1. Reads the full conversation thread
2. Drafts a reply in your voice using your campaign's Reply Goal
3. Waits 2-15 minutes (randomized — never instant)
4. Sends from the same mailbox/number that received the reply
5. Extracts any new data (buy box, budget, timeline) into the CRM

For routine back-and-forth — "yes I'm a builder", "we focus on land", "what's your price range" — the AI carries the conversation without you involved.

Train the AI's answers (per campaign)

The AI only ever answers from what it has on file — the campaign's Reply Goal, the contact's property and buy-box details, and the thread itself. Anything else it refuses to guess at. So when a lead asks "how did you get my number?", "what are closing costs?" or "are you a real person?", the honest default is a dodge. Trained answers fix that: every email, SMS and AI Calls campaign has its own list of question-and-answer pairs.

1. Open the campaign on the Email or Texting page and press 🎓 Train AI replies in the AI settings card (it is also on the Edit campaign form, and on the AI Calls cards under Voice Outreach).
2. Add a row per question: When they ask or say… on the left, The AI should answer… on the right. Up to 40 pairs, 200 characters per question, 600 per answer.
3. Copy in the Q&As from another campaign if you already trained one — duplicates fold automatically.
4. Press Save answers. Email and SMS campaigns use them on the very next reply; AI Calls campaigns pick them up on their next outbound call.

How the AI uses them: it matches the lead's message by meaning, not exact words, and treats your answer as the source of truth — every fact, number, price and commitment stays exactly as you wrote it, while the wording adapts to the conversation (texts stay short, so long answers get condensed). A question you haven't trained is still never invented. On AI calls, the platform rules still win: a trained answer that states a guaranteed price or a binding offer gets softened to "a person confirms the specifics".

Give the AI a conversation script (optional)

Trained answers cover what the AI says when a lead asks something. A conversation script decides the path: what the AI says first, and what it says next depending on the answer — "if they say yes, say this; if they say no, say that". It is optional: a campaign with no script (or a switched-off one) replies exactly the way it always has.

1. Open it — the 🗺 Conversation script button sits beside 🎓 Train AI replies on the Email or Texting page's AI settings, on the Edit campaign form, and on each AI Calls campaign card on Voice Outreach.
2. Build the tree — press Start from an example for a ready-made yes / maybe / no script, or add the first step and type what the AI says. Press + If they say… on a step to add a branch: a short label for what the lead might say ("Yes / interested", "Not right now", "No"), and under it the step the AI moves to. Branches fan out below each step like a flowchart; wide scripts scroll sideways, and on a phone they stack.
3. End a path — a step with no branches can carry up to three endings, picked from the Then menu on its card: hand off to a person (the contact is flagged for you, with a reason you write; on a call the AI offers a callback), mark not interested & stop (status → Not interested, the AI stops replying; on a call it wraps up politely), mark as qualified, add a tag you name (an automation can pick that tag up — create a task, enroll a campaign, post in Team Chat), add a note to the contact in your words, or turn the AI off for this contact. Up to 40 steps, 8 levels deep, 6 branches per step.
4. Save script — the Use this script switch at the top keeps a script on file without using it. On texts and emails it applies to the next AI reply; on AI Calls, from that campaign's next outbound call.

How the AI follows it: on texts and emails, your sequence's message 1 goes out exactly as you wrote it (the editor shows it at the top, read-only) — the AI doesn't send it. Step 1 is the question that message asks, so the AI knows what the lead's first reply is answering; every step after it is something the AI writes back. Each step holds up to 1,000 characters and keeps its formatting — press Enter for a new line or leave an empty line between paragraphs, and the AI's reply uses the same breaks (handy for an email step). A scripted step can also run longer than the AI's usual short reply; it sends the whole step. The ⤢ button opens the editor full screen. On AI calls, the AI asks step 1 right after its greeting. At every step it keeps your wording (lightly adapted to the conversation), picks the branch that matches what the lead said by meaning, and moves on. If the lead says something no branch covers, it answers briefly from your trained answers and the facts on file — never guessing — then brings the conversation back to the current step. It never reads step numbers aloud or mentions a script, and the platform rules still win: a stop request always ends the conversation, and a scripted line never turns into a binding offer.

The Drafts inbox (draft mode)

When a draft-mode campaign generates a reply, it doesn't go out — it stages in the Drafts inbox and you get notified. Open it from the Drafts awaiting review KPI on the Home page, or the drafts icon in the top bar (it shows a count badge when drafts are waiting).

For each draft you can:

• Send — fires the reply as-is from the mailbox/number that received the inbound.
• Edit then send — tweak the text in the box first; your edited version is what goes out.
• Discard — the draft won't be sent. (Future inbounds from that contact still queue — to fully pause a contact, toggle AI Replies OFF on them.)

Data extraction (buy box, budget, timeline) still happens in draft mode — only the send waits for you.

What triggers the red flag

When the AI sees one of these signals in an inbound message, it stops auto-replying and flags the contact red:

• Call request — "can we hop on a call", "what's a good time to talk", "give me a ring"
• Specific deal mention — "I have a 5-acre lot at 123 Oak St, what would you pay"
• Complaint or hostile language — "stop emailing me", "this is spam", "I'll report you"
• Opt-out request — "unsubscribe", "remove me", "don't contact me again"
• Legal/serious topic — anything mentioning litigation, regulators, contracts that need signing
• Inbound call — any phone call gets flagged automatically (calls always need a human)

Red contacts appear with a 🚩 in the CRM, in the "Needs attention" filter, and as a count on the nav bar.

How to handle a red contact

1. Open the contact — the attention banner shows what triggered the flag
2. Read the conversation history (always there in the timeline)
3. Take the action they want — book a call, send the deal info, honor the opt-out
4. Click Clear flag when handled, or set status to qualified / notinterested

The AI does NOT auto-reply to any further messages from a flagged contact until you clear the flag. This prevents the bot from saying something dumb to someone who's clearly upset or asking for human help.

The attribution context on red contacts

The banner includes WHY the contact reached out — "Inbound call from John Smith. Last outreach: mailed May 15." This is critical context before you call back.

If they got mailed AND emailed last week, that's a strong-intent lead. If they're flagging because they're annoyed, you need to know that before responding.

How to disable AI replies per contact

If you want to handle a specific contact entirely manually (e.g., a high-value lead you don't trust an AI with):

1. Open the contact's detail view
2. Toggle AI replies off

That contact's replies will go straight to red-flag status. You handle all responses manually until you re-enable.

How to take over a conversation

Once a contact is flagged, you can:

• Reply directly from your email/SMS — the thread continues in your normal inbox. Direct REI logs your replies via IMAP polling and adds them to the timeline.
• Reply from inside Direct REI — open the contact, hit Send in the detail view, type your message, send. Goes through the same mailbox.

What's next

To understand which campaigns are producing red-flag leads: Reading cross-channel attribution.