How to Classify Email Replies Without Overstating Buyer Intent

Who this framework is for

This framework is for teams that receive direct responses to outbound or follow-up email and need consistent routing across human reviewers, CRM workflows, and reporting. It is useful when small reply volumes make every response visible and when several languages, business cultures, or channel relationships create ambiguous wording.

Contents

1 Define the reply object before interpreting it 2 Separate response class from commercial conclusions 3 Apply the SellersGPT REPLY Function Matrix 4 Route each class with explicit stop rules 5 Record confidence, contradictions, and review 6 Illustrative example 7 Common mistakes 8 What this does not mean 9 Common questions 10 Sources and methodology

1. Define the reply object before interpreting it

Email has technical layers. SMTP transports messages and reports delivery outcomes; the Internet Message Format defines fields such as From, Date, Message-ID, In-Reply-To, and References. IETF - RFC 5321, Simple Mail Transfer Protocol IETF - RFC 5322, Internet Message Format Those fields can help establish that a received message belongs to a thread. They do not certify the sender's commercial role or meaning. Create a reply record with:

• immutable message ID, received time, sending mailbox, and thread references;

• the original message and campaign context;

• exact reply text and attachments, subject to access and retention policy;

• detected language and any translated working copy;

• sender-to-contact and contact-to-account linkage, including confidence;

• classification, classifier or reviewer, rule version, and decision time;

• suppression or stop action, owner, due time, and later correction history. Do not overwrite exact words with a label. A label is a derived interpretation. The message is the evidence. A translation should be marked as a translation, not substituted silently for the source. Delivery-status notifications are also messages, but they are not buyer replies. Enhanced mail status codes describe delivery conditions such as address, mailbox, and network failures. IETF - RFC 3463, Enhanced Mail System Status Codes Store those events in a delivery lane and link them to the sent message without increasing human reply counts.

2. Separate response class from commercial conclusions

Use independent fields: Field Question Example values Message origin Who or what produced the message? Person / Mail server / Auto-responder / Unknown Reply function What operational function does the text perform? Action / Decline / Information / Referral / Stop / Administration Sentiment How is the wording expressed? Positive / Negative / Neutral / Mixed / Unknown Explicit next action What, if anything, did the sender agree to or request? Send specification / Call Tuesday / No action / Unknown Fit evidence What does the reply establish about account or person fit? Supports / Contradicts / No evidence / Unknown Field Question Example values Urgency Is timing stated by the sender? Dated / Relative / None / Unknown Authority What decision authority is evidenced? Known / Inferred / Unknown / Contradicted Treatment What communication rule now governs? Continue / Pause / Suppress person / Suppress company / Review “Send a brochure” is a positive action reply because the sender requested something. It does not prove a funded project. “We buy through Distributor X” may be neutral in sentiment but highly valuable routing information. “Not interested” is a negative decision for the present approach. “Remove me” is a stop instruction and must not be reduced to sentiment. Likewise, a reply can contradict person fit while supporting account fit: “I no longer handle pumps; Maria Chen in reliability owns this.” Preserve both conclusions.

3. Apply the SellersGPT REPLY Function Matrix

The REPLY Function Matrix is an original SellersGPT editorial tool. It classifies the observable function before inferring intent. Class Minimum accepted evidence What it supports Default route Positive action Sender explicitly requests or accepts a defined next step Willingness for that step Assign owner and fulfill or confirm Negative decision Sender explicitly declines the offer or present conversation No current progression on stated approach Stop or pause according to scope Neutral information Sender provides facts without accepting or declining Research and record correction Update evidence; ask only if warranted Referral Sender names or describes another relevant person/team Possible new route, not transferred intent Verify referral and contact policy Forwarded Sender says the message was forwarded, recipient may be unknown Internal circulation occurred Keep identity and outcome Unknown Administrative Out-of-office, role mailbox, acknowledgment, vacation, or ticket receipt Timing or channel administration Schedule, reroute, or wait Delivery failure Mail-system status or non-delivery report Deliverability outcome Suppress/replace as rules require Person unsubscribe Clear request not to contact the person Person-level stop Suppress promptly Company stop Authorized or credible instruction to stop the account/company Company-level stop pending scope verification Suppress and escalate scope if needed Ambiguous Meaning or origin cannot be resolved safely Uncertainty only Human review; no intent claim Positive, negative, and neutral describe the reply's function in context, not the sender's personality. If a message contains more than one function, preserve multiple labels or set a primary routing class plus secondary evidence. A referral with “please stop emailing me” requires both referral handling and person suppression. The referral does not cancel the stop.

4. Route each class with explicit stop rules

Classification has value only when it controls an appropriate action. Positive action Create a dated task matching the request. If the person asks for a specification, send the specification; do not silently convert the response into a meeting request. Verify qualification separately. Direct replies are SellersGPT lead triggers, but the reply still enters identity, profile, ICP, and treatment checks. Negative decision Record whether the decline concerns timing, product fit, the person, the account, or all future contact. Do not invent a reason. “No budget this quarter” supports a timing constraint. It does not prove future budget. Neutral information Correct records without awarding positive intent. A factual correction may be the most valuable reply in the thread even when it does not advance a sale. Referral or forward Verify the named person's role, address source, and permitted route. Do not claim the referred person is interested. A forward with no response is evidence of a forward statement only. Administrative or delivery Out-of-office replies can set a return date but do not prove interest. Hard-bounce handling belongs to deliverability controls. Do not report either as buyer replies. Unsubscribe or company stop Apply the stop state promptly. The FTC states that commercial email recipients must be provided a way to opt out and that opt-out requests must be honored within the applicable U.S. requirements. FTC - CAN-SPAM Act: A Compliance Guide for Business Canadian electronic-message requirements are separate; the CRTC maintains CASL guidance. CRTC - Canada's anti-spam legislation FAQ Legal applicability depends on context and should be reviewed appropriately, but an operational classifier must never treat a stop request as an objection to overcome.

5. Record confidence, contradictions, and review

Use Known, Inferred, Unknown, and Contradicted for each commercial conclusion. Classification confidence should describe clarity of the text and origin, not likelihood of purchase. A useful record includes:

• Literal evidence: the exact sentence supporting the class.

• Scope: person, thread, project, product, account, or company.

• Confidence: high, medium, low, with reason.

• Commercial unknowns: fit, urgency, budget, authority, decision stage.

• Required action: task, owner, deadline, suppression, or review.

• Expiry: when timing or availability information becomes stale. If a later reply changes the meaning, append a correction. For example, “Yes, please send it” followed by “I meant your compliance certificate, not a proposal” narrows the positive action. Preserve both messages and update the derived task.

6. Illustrative example

Illustrative example. The company, people, messages, and calculations below are fictional. They are not a SellersGPT customer result or a forecast. A fictional component supplier receives eight inbound messages tied to one campaign over a week. Message Literal function Class Human reply count? Immediate route “Send the IP69K test report.” Requests defined evidence Positive action 1 Assign evidence task “Not my area. Ask Lena in quality.” Corrects person and refers Neutral + Referral 1 Mark wrong contact; verify referral “I have forwarded this internally.” Reports a forward Forwarded 1 Keep recipient and intent Unknown “No current project.” Declines present timing Negative decision 1 Stop current sequence; retain dated fact “Remove me.” Stop instruction Person unsubscribe 1 Suppress person “Out until Sept. 8.” Automated availability Administrative 0 Pause until after return date “550 5.1.1 user unknown” Delivery status Delivery failure 0 Hard-bounce route “Thanks.” Ambiguous acknowledgment Ambiguous 1 Human review; no intent claim There are 6 human-origin replies / 8 received messages = 75% human-message share. Of those six, one has an explicit positive action: 1 / 6 = 16.7%. These denominators describe this fictional batch only. They are not conversion benchmarks. The forwarded reply remains Forwarded, not Positive, because no receiving stakeholder or desired next action is known.

7. Common mistakes

Reporting delivery notices as replies Mail-server messages describe delivery, not buyer behavior. Calling polite language positive intent “Thank you” can be courtesy, closure, acknowledgment, or ambiguity. Classify the requested action. Treating any question as qualification A question can come from a student, competitor, supplier, wrong contact, or evaluator. Run identity, fit, and treatment gates. Ignoring compound replies “Try Alex, and remove me” requires both referral processing and suppression. Guessing reasons for a decline “Not interested” does not reveal price, timing, incumbent loyalty, or application mismatch. Losing the original language Translated summaries and categories can omit condition, tone, or scope. Preserve the source message with controlled access.

8. What this does not mean

• It does not claim that reply classification can reveal private intent.

• It does not make positive replies qualified leads automatically.

• It does not authorize outreach to a referral or forwarded recipient.

• It does not require a universal sentiment model, score, or phrase dictionary.

• It does not replace legal advice, communication policy, or human review for ambiguous and high-impact cases.

9. Common questions

Is “maybe later” positive, negative, or neutral? Classify the current action first. If no next step is accepted, it is not a positive action. Record the stated timing condition and any explicit permission to revisit; otherwise keep future intent Unknown. Is an out-of-office reply a reply rate event? It is a received administrative message, not a human buyer reply. Report the two separately. Does a referral prove that the account is interested? No. It proves that the sender provided a route or suggestion. The referred person's role, awareness, fit, and interest remain unverified. How should sarcasm or ambiguous wording be handled? Set the commercial interpretation to Unknown, preserve the message, and route for human review. Do not let a low-confidence automated label trigger aggressive action.

10. Sources and methodology

Sources 1 IETF - RFC 5321, Simple Mail Transfer Protocol. Used to distinguish message transport and server outcomes from human commercial replies. Accessed August 25, 2026. 2 IETF - RFC 5322, Internet Message Format. Used for email message fields and thread-evidence context. Accessed August 25, 2026. 3 IETF - RFC 3463, Enhanced Mail System Status Codes. Used for delivery-status categories and the boundary between non-delivery notices and human replies. Accessed August 25, 2026. 4 Federal Trade Commission - CAN-SPAM Act: A Compliance Guide for Business. Used for U.S. commercial-email opt-out and sender-obligation context. Accessed August 25, 2026. 5 Canadian Radio-television and Telecommunications Commission - Canada's anti-spam legislation FAQ. Used for Canadian electronic-message and unsubscribe context. Accessed August 25, 2026. SellersGPT analysis and framework The REPLY Function Matrix, derived evidence fields, and example are original SellersGPT editorial tools. The categories do not originate from the cited standards and do not estimate purchase probability. Exact legal treatment requires jurisdiction- and context-specific review. Update policy Review this framework when messaging standards, applicable rules, routing categories, or SellersGPT lead-trigger definitions materially change. Test production classifiers on multilingual, compound, and ambiguous replies.