TL;DR
- For classification inside a sending platform, compare Smartlead, Instantly, Woodpecker and lemlist. Their labels and automation actions differ; “AI reply” may mean a category, a draft or an automatically sent answer.
- Reply.io exposes inbox categories across email and LinkedIn. HubSpot is useful when classification should create a CRM task or update ownership, but it is a workflow layer rather than a cold-email sequencer.
- Start with separate outcomes for interest, objection or question, not now, out of office, opt-out, wrong person, referral and uncertainty. Keep opt-outs and uncertain cases out of automated reply generation.
- Evaluate every product on the same labelled reply sample. A false positive that sends an eager answer to an opt-out is more costly than a neutral reply sent to a human queue.
What reply classification does
A classifier assigns an inbound message to a useful category: interest, objection, referral, not now, out-of-office, wrong person, unsubscribe, or uncertain. A sentiment score is broader: positive or negative tone does not tell a sales team what action to take. A reply generator writes text; an auto-responder sends it. These functions should not be treated as synonyms.
The safest design separates classification from action. A clear opt-out updates suppression immediately. An interested reply becomes a human task. A question or objection goes to a trained representative. An out-of-office response gets a return-date follow-up if the product supports it. Low-confidence or ambiguous messages wait for review.
Quick comparison
| Product | Documented classification | Action and gate | Main limitation |
|---|---|---|---|
| Smartlead | AI lead categories on Pro and Custom; Pro supports up to five categories | Category can trigger a subsequence; not available on Base | Exact classification performance is not independently measured here |
| Instantly | AI applies built-in or custom labels; default labels include Interested, Not interested and OOO | Custom label feature available on Hyper Growth and Light Speed; AI Reply Agent is a separate, credit-using function | Do not confuse automatic labels with an agent that drafts or sends |
| Woodpecker | First reply tagged Interested, Maybe later or Not interested | Smart Automations can qualify interest and handle autoresponder follow-up | Narrower public categories than a bespoke taxonomy |
| lemlist | AI can mark a lead Interested or Not interested | Marking is irreversible; scope of pausing other campaigns is selected at confirmation | Two-way classification is too coarse for detailed routing alone |
| Reply.io | Inbox categories include Interested, Not interested, Do not contact, Not now and Forwarded; Meeting intent is a subcategory | AI can categorize email, LinkedIn and voice-message replies; Jason AI can draft/send subject to settings | Verify category rules, channel coverage and autopilot boundary on your plan |
Vendor documentation confirms advertised behaviour, not classification accuracy on your prospect population. No product was tested for this article.
Tools
Smartlead: flexible campaign categories
Smartlead’s help article says AI categorization is available on Pro and Custom, not Base. Pro can apply up to five lead categories; Custom lists up to ten and may require the customer’s GPT-4 key for more than five. Categories can also be managed in the Master Inbox and used as triggers for subsequences. This is useful when an agency wants “send the requested case study” or “route referral” to start a defined next action.
Keep suppression separate from interest. The classifier’s label must not be the only control that prevents another campaign from contacting an opted-out person. Maintain a durable blocklist and test whether category changes update all relevant campaigns. Smartlead advertises AI reply drafting separately; the help guide says the Reply Agent drafts for human review and auto-send is listed as coming soon. Don’t buy an auto-responder on the basis of a classification feature.
Instantly: labels and reply agent are separate
Instantly’s automatic reply tagging guide documents AI-applied built-in and custom labels. Default labels include Interested, Not interested and Out of Office; custom label selection is available on Hyper Growth and Light Speed. A label has a description and a positive, neutral or negative status. The user can let AI update existing labels or leave updates manual.
The separate AI Reply Agent can read messages, draft or send replies depending on setup, handle objections and configure label-based inclusion/exclusion. Each generated reply costs five Instantly credits whether it is used, edited or discarded. That is a different cost from classifying a lead. A team that wants triage only should evaluate labels and not automatically budget the reply agent as required.
Woodpecker: simple interest triage
Woodpecker’s Smart Automations use AI to assess the first reply and mark a prospect Interested, Maybe later or Not interested. Its broader product materials describe an autoresponder folder and the option to set a follow-up after the prospect returns. This fits a small team that wants a short queue rather than a detailed taxonomy.
The cost is expressive range. “Wrong person”, “referral”, “question” and “remove me” should not all collapse into one negative bucket if they need different owners or legal treatment. Keep the original message visible, and send any opt-out or ambiguous phrase to a suppression/human-review path until you establish how the status maps to campaigns.
lemlist: automatic interest labels with an irreversible choice
lemlist’s help centre says AI can analyse replies and mark a lead Interested or Not interested. Confirming a label is irreversible. The UI can optionally pause the contact in other campaigns or pause other contacts at the company. For opt-outs, the help article recommends an appropriate unsubscribe action, because a two-label interest state is not the same thing as a global suppression record.
Use lemlist if the team wants simple prioritisation in its campaign inbox. Do not let a two-class model decide that a polite “not this quarter” is a permanent rejection, or that “remove me” is merely not interested. Define who reviews classifications and how to correct mistakes before enabling automatic marking on a large list.
Reply.io: email and LinkedIn categories
Reply’s documentation describes categories such as Interested, Not interested, Do not contact, Not now and Forwarded, with Meeting intent as a subcategory. Its Inbox covers email and LinkedIn threads; its capability reference says classification applies to email, LinkedIn and LinkedIn voice-message replies. Jason AI can draft an answer from the thread and context, with review or automatic handling depending on configuration.
This is a stronger fit when a sales team wants a shared conversation queue across channels. Verify which events stop a sequence, how categories are assigned, and whether auto-generated replies wait for approval. Do not infer that every plan includes the same API, AI or automation allowance from the inbox feature alone.
CRM routing layer: HubSpot is downstream
A CRM can own the contact, company, suppression state, owner and follow-up task after a classifier returns a category. A workflow can branch on a property, assign a task, set a lifecycle status or stop a sequence through an integration. But workflow branching alone does not understand the message. You still need an inbox platform, a supported AI step or a classifier you operate.
This split can be useful if several sending tools feed the same sales team. It also creates a failure point: if the webhook is late, duplicated or matched to the wrong contact, the CRM may route the reply incorrectly. Preserve the source thread ID and client ID; make event processing idempotent and keep a manual review queue.
Subscription and action economics
| Native option | Public subscription basis for an eligible evaluation | Distinct cost or gate |
|---|---|---|
| Smartlead | Pro starts at $94/month; Base at $39 lacks AI categorisation | Pro has up to five AI categories; Custom expands taxonomy and may require a separate GPT key |
| Instantly | Hyper Growth $97/month for custom labels; Growth $47 for core outreach | Classification and AI Reply Agent are separate; a generated agent answer consumes five credits |
| Woodpecker | Starts at $35/month for up to 500 contacted prospects | Verify selected contact tier and Smart Automations availability at checkout |
| lemlist | Email plan $55/month equivalent on annual term at 50,000 email volume | Confirm automatic marking entitlement and actual sending tier for your workload |
| Reply.io | Multichannel starts at $99/user/month on the current US page | Confirm AI category and agent entitlement on the quoted plan |
These are entry or feature-gate figures, not five like-for-like totals: seats, contacted prospects, contacts, sends and annual commitments differ. For 200 inbound replies a month, classification alone does not justify buying a larger sending tier without the outbound workload. If a team asks Instantly's agent to generate a response to all 200, that would use 1,000 credits at five per generated reply. Its separate Growth Credits plan currently starts at $47 for 1,500 credits; the illustrative $97 Hyper Growth plus $47 credits is $144/month before mailboxes and taxes, assuming the credits plan is otherwise unused and all 200 generations are eligible. This is drafting cost, not classification cost. Auto-sending is a separate decision and should stay off during evaluation.
A 200-reply evaluation
Create a labelled set of 200 de-identified replies drawn from the kinds of campaigns you actually run. Have two people independently label each message, reconcile disagreements, then compare each product’s output with the agreed label. Use exactly 30 interested, 30 objection/question, 25 not now, 25 out-of-office, 25 opt-out, 20 wrong-person, 20 referral and 25 uncertain messages. That is 200. Keep the eight ground-truth labels separate even if a vendor exposes fewer classes; map its output to the intended action for scoring. Adjust a later sample to match your actual mix, but retain a challenge set for rarer, consequential cases. This is a proposed test, not a product result.
Include constructed ambiguous cases such as “Please ask our finance lead instead” (referral, with a human check before new contact), “I am away until 12 October” (OOO, defer and check the date), and “Thanks, please remove me” (opt-out despite polite tone). These examples set expected handling; they are not observed vendor results.
Report a confusion matrix by category, not one overall “accuracy” number. Track false negatives for interest, false positives for interest, missed opt-outs, OOO messages treated as live replies and referrals sent to the wrong owner. Then replay the same dataset through the complete action path: sequence stop, suppression, CRM update and human assignment.
To value errors without claiming a product result, assume ten ordinary misroutes take 15 minutes each to investigate at $40/hour: $100 of rework. Assume two missed opt-outs take 45 minutes each for incident review at the same rate: $60 more, excluding any legal or reputational effect. A hypothetical $160 monthly error burden could outweigh a lower subscription, but the actual count must come from the labelled test.
A mistaken positive can trigger an inappropriate automated answer; a missed opt-out can continue unwanted outreach; a false OOO can delay a real conversation. Set different thresholds by consequence. Auto-label low-risk categories if useful, but route opt-outs, legal language and low-confidence messages to a deterministic block or human queue. Store the original reply alongside the output so an operator can audit the decision.
Which should you choose?
Choose Smartlead when campaign categories should start subsequences, Instantly when configurable labels and a separate reply agent suit your operation, Woodpecker for a compact interest triage, lemlist for simple two-way priority marking, and Reply when the inbox must cover email and LinkedIn. Use HubSpot as the routing and ownership layer when it is already the CRM, not as proof that classification is included.
Run the same labelled sample and measure the downstream mistakes that matter to the business. A category model that saves ten minutes but misses suppression is a poor trade. Keep a human owner for edge cases and an independent opt-out control.
The signal-based outbound playbook can help define the handoff from a classified reply to an owned sales action.
FAQ
Is reply classification the same as sentiment analysis?
No. Sentiment measures tone; classification maps a message to an action. “Thanks, but remove me” can sound polite yet must become an opt-out, not a positive or neutral sentiment label.
Should an out-of-office message stop the sequence?
Usually it should pause or reschedule a follow-up to the stated return date, not mark the prospect interested or continue the normal cadence. Check whether your platform detects auto-replies separately and what happens when no return date is given.
Can an AI send the reply automatically?
Some tools can draft and some can send under configured agent modes. Keep human approval for objections, legal questions, pricing commitments and uncertain intent until a labelled evaluation supports broader automation. Classification quality does not establish reply quality.
Where should the original message live?
Keep it in the mailbox or sequencer thread and link it to the CRM record. Store the category, model/version, timestamp and owner separately. This preserves an audit trail when a prospect disputes the classification or an operator needs to correct it.
What is the most useful metric?
Measure false opt-out misses and wrong automatic replies alongside time saved. Break results out by category and language. A single average hides the errors with the highest operational cost.
Do I need a custom classifier?
Only when built-in categories cannot represent your routing rules, or your team needs a common taxonomy across several providers. First test native labels on real de-identified replies. A custom model adds upkeep, privacy review and monitoring for drift.
