SELLIGNITE FIELD GUIDE

Published by the SELLIGNITE team. A practical operating model with examples and checks that show whether it is working. Examples are illustrative unless stated otherwise.

THE SHORT ANSWER

AI reply triage means an AI agent reads every reply to your outbound campaigns, sorts it into a fixed set of categories, takes the safe actions on its own and drafts an answer for a person to approve. Unsubscribes, bounces and out-of-office replies are handled automatically. Interested replies, questions and objections reach a named owner within minutes, with a draft ready, so no warm prospect waits a day for an answer.

KEY TAKEAWAYS
  • Use a fixed list of about nine reply categories, each with one owner and one next action.
  • Let the agent act alone only on low-risk replies: unsubscribes, bounces, out-of-office and auto-replies.
  • Every message sent to a prospect is drafted by the agent and approved by a person.
  • Send uncertain classifications to a person instead of guessing; a confidence threshold makes that rule explicit.
  • Measure time to first human answer on interested replies and the share of labels a person corrects.
THE PROCESS / AT A GLANCE
  1. 01Define the reply categories
  2. 02Write routing rules per category
  3. 03Connect the inbox, sequencer and CRM
  4. 04Run in review mode for two weeks
  5. 05Automate the safe categories

What is AI reply triage?

AI reply triage is the step where an AI agent reads each reply to an outbound campaign, labels it, routes it to the right person or action and prepares a draft answer. It replaces the manual habit of an SDR scrolling through a shared inbox twice a day and deciding what each message means.

Triage is a narrow job, and that is why it works well with AI. The input is short text, the possible outcomes are known in advance and a wrong label is easy for a person to spot and fix. The agent does not need to sell. It needs to read carefully, follow rules and hand the conversation to a human at the right moment.

Reply triage sits between your sequencing tool and your CRM. In a Predictable Revenue style outbound system, it is the bridge between the people who generate replies and the people who run meetings, which is also where many teams lose warm prospects. Our SDR to AE handoff checklist covers the next step once a meeting is booked.

Why does outbound reply handling break as volume grows?

Outbound reply handling breaks because replies arrive at random times, in mixed formats, across several inboxes, while the person responsible is busy with other work. At ten replies a week a founder can answer each one personally. At fifty, the replies that need thought wait behind the ones that need none.

  • Interested replies wait. A prospect who writes "sure, send times" on Monday afternoon and hears back on Tuesday has had a day to cool down. The same logic as speed to lead on inbound applies to outbound replies.
  • Stop requests get missed. An "unsubscribe" that is not processed means the next step of the sequence still goes out, which damages trust and sender reputation.
  • Referrals are lost. "Talk to Anna, she runs this" is one of the most valuable replies you can get, and it often ends up in no one's task list.
  • Reasons are not logged. Without a label on each reply, you cannot learn which segments or messages produce objections.

For well-targeted outbound we plan for an 8-12% reply rate, with about one in three replies becoming a booked meeting. Those are planning targets, not guarantees, and the second number depends heavily on how quickly and how well replies are answered.

How to classify cold email replies: nine categories

Classify cold email replies into a short, fixed list of categories, each tied to one owner and one next action. Too few categories hide important differences; too many make the agent and the team inconsistent. Nine is a practical starting point.

CategoryExample replyNext actionWho acts
Interested"Yes, open to a call next week."Answer with times, book the meetingPerson approves draft
Question"Does this work with Pipedrive?"Answer from approved facts, propose a stepPerson approves draft
Referral"Not me, try our COO."Thank, ask permission to mention them, add new contactPerson approves draft
Not now"Ask me again in Q2."Confirm, set a dated follow-up taskPerson approves draft
Objection"We already have a vendor."One question, per objection playbookPerson approves draft
Not interested"No thanks."Stop sequence, log reasonAgent, person reviews weekly
Unsubscribe"Remove me from your list."Suppress on every channel immediatelyAgent
Out of office"Back on the 24th."Pause, resume after return dateAgent
Bounce or left company"Anna no longer works here."Mark contact invalid, flag account for new contactAgent

Some replies fit two categories, such as an out-of-office message that names a colleague to contact instead. Tell the agent which category wins (here, Referral), and add an "Unclear" label for anything it cannot place. Unclear always goes to a person.

Which replies can the AI handle alone, and which need a person?

The agent can act alone when the action is low-risk and reversible, and it must hand over whenever a message goes to a prospect. That single rule keeps the speed of automation without letting an agent speak for your company unsupervised.

Safe for the agent to complete

  • Unsubscribe and stop requests: suppress the contact in the sequencer, the CRM and any LinkedIn steps at once. This one should never wait for a person.
  • Out-of-office and auto-replies: read the return date, pause the sequence and resume a few days later.
  • Bounces and job changes: mark the contact invalid and create a research task for a replacement at the same account.
  • Clear "not interested" replies: stop the sequence and log the reason, without sending a reply.

Drafted by the agent, approved by a person

  • Every answer to an interested prospect, a question, a referral, a timing reply or an objection.
  • Anything that mentions price, contract terms, integrations or delivery dates. The draft may only use facts from an approved list; if the fact is missing, the draft says a colleague will confirm.

Objection drafts follow the same five-step pattern described in our guide to cold outreach objections: acknowledge, ask one question, answer briefly only if earned, offer a small step, log the reason.

How fast should you respond to cold email replies?

Aim to send a human-approved answer to every interested reply within one working hour, and to every other reply that needs an answer within the same working day. These are our working targets, chosen because a prospect who just replied is reading their inbox now; they are not industry statistics.

AI reply triage makes those targets realistic because the slow parts disappear. The agent labels the reply within minutes of arrival, pulls the account context from the CRM, writes the draft and sends a notification to the owner in the tool they actually watch, such as Slack, Microsoft Teams or the CRM task list. The owner's job shrinks to reading, editing if needed and pressing send.

Set a fallback for absences. If the owner has not approved an interested-reply draft within an hour, the notification goes to a second named person. A rule like this matters more than the model you choose.

How to set up AI reply triage in five steps

Set up AI reply triage by fixing the categories and rules first, connecting the tools second and widening the agent's autonomy only after a review period. The order matters: an agent with unclear rules produces unclear results quickly.

  1. Define the categories. Start from the nine above, then read your last 100 replies and check that each one fits. Adjust names to the language your team already uses.
  2. Write routing rules. For each category, write the owner, the next action, the response-time target and the CRM fields to update. Keep the whole set on one page.
  3. Connect the tools. The agent needs read access to the reply inbox, the ability to pause and stop sequences in your sequencing tool (for example Instantly, Apollo or HubSpot sequences) and write access to a small set of CRM fields. Our CRM hygiene guide lists fields worth having.
  4. Run in review mode for two weeks. The agent labels and drafts everything, but a person confirms every label and every action, including the automatic ones. Record each correction.
  5. Automate the safe categories. When corrections on a category are rare and the mistakes are harmless, let the agent complete it alone. Keep spot-checking a sample every week.
GO DEEPER / PUT THE IDEAS TO WORK

Example: one morning of replies for a small outbound team

Imagine a 30-person company selling an accounting add-on to finance teams, with one SDR and a founder who runs sales calls. The scenario is illustrative. Overnight, 14 replies arrive to three running campaigns.

  • Six are handled before anyone logs in: three out-of-office replies are paused with resume dates, one unsubscribe is suppressed on email and LinkedIn, one bounce is marked invalid with a research task for a new contact, and one plain "no thanks" is stopped and logged.
  • Two are interested. The agent drafts answers with two proposed time slots and notifies the SDR. Both are approved with small edits within the first hour of the day.
  • One is a referral to a finance director. The agent drafts a thank-you, adds the director as a new contact linked to the original thread and drafts a short first message that mentions who suggested the contact, held until the referrer agrees.
  • Three are objections or timing replies. Drafts follow the objection pattern, and one "ask me in March" reply gets a dated task in the CRM.
  • One question asks about a specific ERP integration that is not on the approved facts list. The draft says a colleague will confirm, and the founder answers it personally.
  • One is labeled Unclear, a one-word reply: "Interesting." The SDR decides it is mild interest and sends a short question.

The SDR's morning work on replies is reading and approving eight drafts instead of reading, sorting and writing fourteen messages from scratch.

Prompt rules that keep the agent accurate

The instructions you give the agent matter more than the model. These rules prevent the most common triage errors:

  • Classify first, draft second. Ask for the category, a one-line reason and a confidence score before any draft. A person can check the reason in seconds.
  • Use a confidence threshold. Below the threshold you set, the label becomes Unclear and goes to a person, whatever the agent's best guess.
  • Quote, do not invent. Drafts may only state facts from an approved company sheet: what you offer, how you work, which integrations exist. No prices in email.
  • Read the whole thread. A reply of "yes" means nothing without the question it answers.
  • Match the prospect's length. A two-line reply gets a two-to-four-line answer, not a pitch.

For a wider view of where agents fit in the SDR role, read what an AI SDR can and cannot do. If you want help designing and running reply triage inside your own stack, see our AI agents service or our outbound sales systems.

Your implementation checklist

  • Read your last 100 replies and test the nine categories against them
  • Write one page of routing rules: owner, action, response target, CRM fields
  • Create an approved facts sheet the agent may quote from
  • Set a confidence threshold and an Unclear category that always goes to a person
  • Run two weeks in review mode and log every correction
  • Automate only the categories with rare, harmless corrections
  • Set a backup owner for interested replies not approved within an hour

What to measure: Track time from reply to first human-approved answer on interested replies, the share of labels a person corrects per category, and the share of replies that become booked meetings, weekly.

Frequently asked questions

What is AI reply triage in outbound sales?

AI reply triage is the use of an AI agent to read every reply to outbound campaigns, assign it to a fixed category and route it to the right action. Low-risk replies such as unsubscribes and out-of-office messages are handled automatically. Replies that need an answer get a draft that a person approves before it is sent.

Can AI answer cold email replies without human review?

AI can safely complete actions that do not send a message, such as suppressing an unsubscribe or pausing a sequence for an out-of-office reply. Answers to prospects should be drafted by AI and approved by a person, because the reply speaks for your company and can contain mistakes. Many teams keep this approval step permanently for interested replies.

How many categories should I use to classify cold email replies?

A list of about nine categories works for most B2B teams: interested, question, referral, not now, objection, not interested, unsubscribe, out of office and bounce. Add an unclear label for replies the AI cannot place, and send those to a person. Fewer categories hide important differences, while many more make labeling inconsistent.

How fast should a sales team respond to an interested outbound reply?

A practical target is a human-approved answer within one working hour for interested replies and within the same working day for other replies that need an answer. The prospect is usually still in their inbox right after replying. AI triage helps by labeling the reply and preparing a draft within minutes.

What tools does AI reply triage need to connect to?

AI reply triage needs read access to the inbox that receives replies, the ability to pause or stop contacts in the sequencing tool and write access to a few CRM fields. A notification channel such as Slack, Microsoft Teams or CRM tasks tells the owner when a draft is waiting. The agent should not have wider permissions than these tasks require.

WANT THIS RUNNING IN TWO WEEKS?Book a free AI Audit ↗