Reply classification — the five categories and per-category response.

The premise

A reply is not a binary event. The operator who treats it as one — every reply either meeting-or-suppression — discards the operational signal embedded in the language of the response itself. The reply that says “not now, we just signed with someone else for a 12-month term” is not a no; it is a no with an explicit timing constraint and a CRM stage. “Send me a deck and follow up in Q3” is not a yes; it is a soft-pass with a structurally low conversion rate that masquerades as forward motion in the upstream metrics. “Please contact Sarah on our procurement team” is not a no; it is a routed handoff that, executed correctly, has a higher base-rate conversion than the original outbound touch.

The five-category model — positive intent, negative intent, objection-with-signal, soft-pass, auto-responder — is the minimum taxonomy required to extract operational signal from production reply volume. Each category routes to a different downstream workflow, has a measurably different per-category conversion rate, and has a classification-accuracy ceiling that, in our observation, distinguishes operators converting reply volume at 15% from operators converting at 3%.

The five categories

(1) Positive intent — interested, ready to meet

A reply that affirmatively requests next steps: a meeting, a demo, a deck, a price, a written follow-up containing the contact’s availability. The defining characteristic is the absence of a qualifying clause — no “but,” no “send more,” no “follow up in.” The contact has read enough to commit to a calendar event.

Typical share of total replies: 15 to 25 percent. The high end reflects upstream targeting that has produced a self-selected high-intent population; the low end reflects broader-net campaigns where objection-with-signal is correspondingly larger.

(2) Negative intent — clear no

A reply that closes the conversation with no preserved future-engagement signal: “not interested,” “please remove me,” “stop emailing me,” a one-word “no.” The defining characteristic is the absence of a reason. A no with a reason is a different category entirely.

Typical share of total replies: 10 to 20 percent. The high end typically reflects mistargeted campaigns or list-quality issues, and is the cleanest leading indicator of an upstream segmentation problem worth investigating before the reply rate itself degrades.

(3) Objection-with-signal — “not now because X”

A reply that declines the immediate ask but preserves the qualifying constraint behind the decline. The four canonical B2B objections — timing, budget, authority, fit — each surface as a recognizable linguistic pattern, and each contains forward-looking operational signal: when the contact expects to revisit, who actually owns the decision, what the actual budget envelope is, what fit-criterion the contact is filtering against.

Typical share of total replies: 25 to 40 percent. The largest category in any production reply stream, and the category the binary-classification operator routes incorrectly with the highest frequency. Each objection has an empirically distinct conversion-rate-on-objection (Chapter 4); the cumulative loss from treating the entire category as negative is the largest pipeline leak in the stack.

(4) Soft-pass — “send me more info” or “follow up in Q3”

A reply that neither commits to next steps nor closes the conversation. The contact returns a deferral phrased as forward motion: “send me a deck,” “put me on your list,” “happy to take a look,” “circle back in the new year.” The defining characteristic is the substitution of a low-cost ask for a calendar commitment.

Typical share of total replies: 15 to 25 percent. The soft-pass conversion rate to qualified opportunity is, empirically, between 4 and 9 percent — meaningfully above the cold-touch baseline but a fraction of the positive-intent rate. The operator who routes soft-pass as positive-intent inflates the meeting-booked metric and pays for the inflation downstream at the show-rate and qualification stages.

(5) Auto-responder — OOO, role-change, redirect

A system-generated reply with no human in the loop at the moment of receipt. Structurally heterogeneous; the sub-classification (below) is where most of the operational signal lives. The routing decision is made entirely on the metadata and language of the auto-response itself.

Typical share of total replies: 8 to 15 percent. The high end corresponds to mid-Q4 and mid-summer sending windows where out-of-office volume spikes. Operators who do not sub-classify treat the entire category as noise and lose the role-change and redirect signal embedded in it.

Per-category routing rules

The per-category routing target is a function of the downstream operation each category requires. The five-target architecture below is the minimum-viable routing tree; the operational details of each target are addressed in subsequent chapters.

The four-hour SLA on objection-with-signal is the discipline that distinguishes the 15%-conversion operator from the 3%-conversion operator. The contact who raises an objection is, by the act of raising it, signaling engagement; the half-life of that engagement is, empirically, on the order of hours, not days. Chapter 3 addresses the triage workflow that holds the four-hour window at production reply volume.

The classification mechanics — three tiers

The classification step admits three implementation tiers, each with a measurably different accuracy ceiling and per-reply cost.

Tier 1 — keyword heuristics

A regex-and-token-match pipeline that routes on category-defining phrases: “out of office,” “unsubscribe,” “send me,” “follow up,” “not interested,” “no longer with.” Cheap, deterministic, fully auditable. Accuracy ceiling, in our observation, ~70 percent. The 30-percent error band concentrates in objection-with-signal, where the linguistic surface is most varied.

Tier 2 — LLM classification

A single-shot classifier prompt returning one of the five categories plus a confidence score and, for objections, a sub-category (timing, budget, authority, fit). Two to three orders of magnitude more expensive per classification than the keyword tier, with accuracy ceiling ~90 percent. The 10-percent error band concentrates at the boundary between soft-pass and objection-with-signal, and between auto-responder sub-categories.

Tier 3 — human triage

A human reviewer reads and routes. Slowest and most expensive per-reply, and the only tier that reliably captures the boundary cases where the next operational step requires a judgment the upstream classifier cannot make. Accuracy ceiling ~98 percent. The remaining 2 percent is ambiguity no classifier would resolve without additional context.

The production architecture is a confidence-thresholded combination: keyword heuristics for the high-volume unambiguous categories (auto-responders, clear positive, clear negative); LLM classification for the productive middle; human triage gated on a confidence threshold — typically 0.75 to 0.80 — below which the reply is escalated regardless of which tier produced the candidate label.

The negative-with-signal subcategory

A no is operationally distinct from a no with a reason. “Not interested” routes to suppression. “Not interested, we just signed a 24-month contract with another vendor” routes to suppression for the contract duration and a CRM stage with a 22-month re-engagement trigger. “Not the right person — I left product to run sales ops” routes to a role-change handler that updates the contact record and surfaces the new title for upstream segmentation. “Not now, we’re mid-fundraise” routes to a trigger-based nurture keyed to the next funding announcement.

The subcategory is, in our observation, 30 to 50 percent of total negative-intent volume. The operator who routes all negative replies to undifferentiated suppression discards a population with a measurably non-zero future-conversion rate. The implementation cost is a single additional LLM classification step.

The soft-pass trap

Soft-pass becomes operationally dangerous when it becomes the operator’s default-conversion-bucket. The pattern is recognizable: reply rate healthy, meetings-booked below target, and the gap filled with soft-pass contacts re-classified as warm-pipeline. The downstream consequence is a CRM stage whose per-stage conversion rate is 4 to 9 percent, and a pipeline forecast that mistakes soft-pass volume for qualified opportunity volume.

The operational ceiling on soft-pass conversion is the per-soft-pass conversion rate itself. Increasing soft-pass volume does not move the rate. The only structurally effective intervention is the trigger-based re-engagement layer (Chapter 6) that re-classifies soft-pass contacts back into positive-intent or objection-with-signal at a future moment when the timing constraint dissolves.

The “send to other inbox” routing

A high-value pattern that surfaces inside auto-responder and soft-pass: the recipient explicitly redirects to a different stakeholder with a named contact and, often, a working email address. “Please reach out to David, our VP Engineering, at david@…” is, operationally, a warm introduction with a verifiable upstream source.

The per-pattern follow-through rate — meeting booked with the redirected stakeholder within 14 days, when the operator executes correctly — is between 30 and 45 percent, three to five times the cold first-touch rate to the same stakeholder. The follow-through requires (a) referencing the original recipient by name, (b) attaching the original thread, (c) executing within 48 hours while the redirect is still fresh. Most teams do not.

The auto-responder sub-classification

The category collapses three operationally distinct patterns into one undifferentiated stream. The sub-classifier separates them.

Return-date OOO

A standard out-of-office with an explicit return date. The operational response is a re-send queued for one to two business days after the stated return — not the day-of, which lands at the top of the contact’s post-vacation inbox triage and is statistically the worst moment to re-engage. The thread is suppressed from the upstream sequence for the OOO window to prevent stacking touches against an absent recipient.

Left-the-company

An auto-response explicitly stating the contact has left, often with a redirect to a replacement. The operational response is (a) update the contact record to closed-lost / role-change, (b) if a redirect is named, treat as “send to other inbox” with the 30-45% follow-through rate, (c) trigger an upstream re-segmentation pass to find the new role-holder at the same company. The contact themselves becomes a future-touch target at the new company after a 90-day onboarding window.

Unmonitored-inbox

An auto-response indicating the inbox is not monitored, typically with a redirect to a contact form, support address, or different team. Suppress the address from sequence, log the reason, and — if a verified alternate route is named — treat as a routing handoff. The pattern is a strong leading indicator of an upstream data-quality issue on the address.

The classification-confidence threshold

The threshold gating autonomous routing from human review is, in our observation, the most consequential parameter in the classification layer. Set too high (e.g., 0.95), the human queue saturates and the four-hour SLA collapses. Set too low (e.g., 0.50), low-confidence misclassifications produce the most expensive errors — a soft-pass auto-booked as a meeting, an objection routed to suppression, an OOO treated as negative intent.

The production-tested threshold sits between 0.75 and 0.82, calibrated by (a) the cost-of-error on each category, (b) human-triage capacity, and (c) the through-day volume profile. Mature operators adjust the threshold per-category — more conservative on positive intent (where misclassification means a missed meeting), more permissive on auto-responder (where misclassification means a delayed re-send).

Common operator failures observed in production

• Binary classification. Every reply routed as positive or negative; the productive middle — 60 to 70 percent of reply volume — treated as undifferentiated. Objection-with-signal routes to suppression, soft-pass routes to AE handoff, both downstream workflows degrade.
• Soft-pass-as-positive. Soft-pass routes to the AE calendar workflow. The meetings-booked metric inflates; the meetings-held metric does not. Show-rate, no-show, and meeting-to-opportunity rates degrade in proportion to the soft-pass admission, and the gap is initially attributed to the AE rather than the classifier.
• No auto-responder sub-classification. The entire auto-responder stream routes to a single “noise” bucket. Role-change, redirect, and return-date timing are all discarded. The upstream sequence continues stacking touches against absent recipients, degrading sending-domain deliverability in the process.
• Missing the “send to other inbox” follow-through. The redirected stakeholder is added to a generic prospecting list rather than receiving an immediate, in-thread, referenced handoff. The 30-45% follow-through rate degrades to the cold-touch baseline.
• Classification without a confidence score. The classifier returns a label with no confidence. The human-triage queue receives no prioritization signal and either reviews everything (SLA collapses) or reviews nothing (low-confidence misclassifications propagate).
• Static classification rules. The keyword classifier is deployed and never re-tuned against LLM- or human-tier ground truth. Category boundaries drift as upstream campaign content changes, and accuracy drops below the deployment-day baseline within a quarter.

Pre-deployment classification checklist

• The five-category taxonomy is documented, with explicit per-category routing targets and SLAs
• The negative-intent category is split into with-signal and without-signal subcategories
• The auto-responder category is split into return-date, left-the-company, and unmonitored-inbox
• A classification confidence score is returned with every label, and a threshold gates autonomous routing from human review
• The per-category accuracy is measured against a human-labeled ground-truth sample of at least 200 replies before production deployment
• The four-hour SLA on objection-with-signal has an owner and a measurable adherence rate
• The “send to other inbox” pattern has an explicit handoff workflow with the original-thread reference and the 48-hour execution window
• A weekly audit re-samples 50 to 100 classified replies and compares against a human re-label, with drift triggering a classifier re-tune

Where this fits

Classification is the first operation on the reply, and every subsequent step in the stack inherits its output. A correctly routed positive intent is the input to meeting-booking mechanics (Chapter 5). A correctly classified objection-with-signal is the input to the per-objection response library (Chapter 4). A correctly sub-classified auto-responder is the input to the upstream sequence’s OOO-aware re-send logic. A misclassification at this step is a misallocation of every downstream resource — AE calendar time, response-library effort, CRM stage discipline, upstream sequence inventory — propagated through every subsequent operation in the stack.

Chapter 2 addresses the routing architecture connecting classifier output to the downstream operational targets — Slack, CRM, calendar, suppression — and the per-target handoff mechanics. Chapter 4 addresses the per-objection response library that the objection-with-signal routing feeds into, including the per-objection conversion-rate-on-response and the failure mode of the operator who treats objections as final rejections.