The reply desk

The reply is where cold email becomes money, and it is the part of every sequencer that has been built last. Instantly charges 5 credits per AI-generated reply and puts Autopilot on Hyper Growth and above. Smartlead's SmartAgents are workflow automations, not a reply product. EmailBison, the agency favourite, has "no native AI features documented". An agency's real reply work runs across a unibox, Slack, Zapier, a client CRM and a spreadsheet. The job is classifying, answering within minutes, booking onto the client's calendar, proving the meeting and billing for it. See Replies, booking and handoff.

#The opening

Five jobs follow every send: unified inbox, classification, drafting, booking and handoff (Replies, booking and handoff). The volume sequencers cover the first two and meter the third. Nobody sells the last two for agencies.

The evidence:

• Volume of work. Instantly's 2026 benchmark puts the average reply rate at 3.43%. Belkins measures 0.45% once auto-replies are excluded. That spread is itself the problem: a large share of what lands is out-of-office mail, referrals, objections and unsubscribes that someone has to sort. At 3.43%, an agency contacting 10,000 prospects a month gets roughly 340 replies across dozens of mailboxes (my arithmetic).
• Metering confusion. Instantly's help centre says the 5 credits are charged "regardless of whether you send, edit, or skip it". Its own pricing blog says "Rejected drafts in HITL mode incur no credit charges". Reviewers also complain that replying from Unibox sits behind a higher plan (Trustpilot, Sep 2026).
• Split inboxes. HeyReach has its own Unibox for LinkedIn. Smartlead added LinkedIn steps through a HeyReach integration in January 2026. The lead's email reply and LinkedIn reply still live in different places. See Multichannel: LinkedIn, calls, SMS and video.
• Routing is priced for enterprise. Chili Piper starts at $15,000 a year. Calendly Teams is $16 a seat. Nothing in between is built for cold-email replies.
• Opt-outs carry real penalties. Verkada paid $2.95M, partly for failing to honour unsubscribes. ACMA's notice against Tabcorp was for A$1,254,000, over messages sent after consent was withdrawn. In Germany the dispute value rises to €3,000 per email after a warning letter. An "unsubscribe me" buried in a reply is a legal event, not a label.

#Who pays and how much

Agencies pay first, because their contracts promise meetings and they cannot prove them without spreadsheets (Lead-gen agencies). A 110-client shop like Danish Lead Co. reports 10,000 meetings booked on Smartlead. That is the unit an agency bills on, and the unit a reply desk should charge for.

Price anchors on either side:

The cost side is tiny. Classifying a reply on Claude Haiku 4.5 costs roughly a tenth of a cent at $1/$5 per million tokens (my estimate, ~800 tokens in). Per-meeting pricing therefore carries software margin even at a few euros a meeting.

#Why incumbents have not closed it

• Their margin is in metering. Sequencers price sending flat and replies per use. Giving the reply layer away flat, or pricing it on meetings, would cut into the line they are expanding (Replies, booking and handoff).
• Their tenancy model is wrong for it. Instantly's white-label portal "can't restrict" a client to their own campaigns, and multiple portals need separate workspaces and subscriptions. A client-facing approval queue is impossible on that base. See White-label and agency portals.
• They chase autonomy. Instantly markets an AI SDR that "can replace a human SDR". The buyer who got burned by autonomy wants a fast approval queue instead. See Fully autonomous AI SDR.
• EmailBison has no AI layer at all, and its customers are exactly the large agencies that need one (EmailBison).

#What you would build first

1. Connectors, not a sequencer. Ingest replies from Smartlead (EMAIL_REPLY, LEAD_CATEGORY_UPDATED webhooks), Instantly (reply_received, lead_meeting_booked webhooks), EmailBison (reply and interested events) and HeyReach. Reading through the sequencer's API avoids Google's restricted-scope CASA review at launch (OAuth verification and the end of basic auth).
2. Classification with a hard opt-out path. Interested, not now, referral, wrong person, OOO with return date, objection, unsubscribe. Any opt-out phrasing suppresses the person across every campaign and client workspace within minutes and writes an audit record.
3. Approval queue per client. Drafts wait for an agency operator or the client's own approver, with SLA timers on "interested" replies. Every AI send is logged with who approved it.
4. One thread per lead. Email and LinkedIn replies merged by lead identity.
5. Handoff and proof. Book onto the client's calendar, push to the client's CRM, notify, then mark booked → held → opportunity. The "held" event is the billable one.

#How the leaders would respond

Instantly would bundle more credits into Scale and add client-scoped approvals, probably within two or three quarters. Smartlead would turn SmartAgents into a reply agent and lean on its white-label base. Neither is likely to switch to per-meeting pricing, because that exposes their credit revenue to their weakest-performing customers. The overlay model also cuts the other way: Smartlead or Instantly can throttle the webhooks or APIs you depend on. Both publish them today (API, webhooks and MCP), but terms can change.

#Scores, argued

Pain: 4. Reply handling is daily manual labour for every agency, and the legal stakes of a missed opt-out are documented in three jurisdictions. It is not a 5, because founders sending low volume handle replies by hand without much complaint.

Gap: 3. Unified inboxes and AI drafting exist (Instantly, Smartlead, HeyReach). What is missing is multi-client approval, cross-channel threads, meeting attribution and outcome pricing. That is a real gap, but not open ground.

Size: 3. The buyer is the agency segment plus reply-heavy founders. It is a layer on top of sequencer spend, not a replacement for it, so it caps below the sequencer pool unless it becomes the front end of a full platform.

Moat: 3. Per-client playbooks, meeting-outcome data and agency client logins create switching costs. Per-meeting pricing is awkward for credit-metered incumbents to copy. The classification itself is commodity LLM work.

Speed: 4. An overlay needs no sending infrastructure, no warmup and no CASA review. Agencies can trial it on one client in a day.

Safety: 4. Post-send work carries little provider risk. The residual risk is autopilot replies that embarrass a client, which is why approval is the default.

#What would kill it

#What this means for an entrant

• Start here, then grow down the stack. A reply desk that works on Smartlead and EmailBison gets paying agencies before your own engine exists. Once you own their replies, migrating their sending to you is the natural second sale. See The wedge and MVP spec.
• Stack it with Agency operating system. The desk is the client portal: approvals, branded reports and meeting proof are what an agency resells. Together they are one product.
• Stack it with EU-native compliant outbound. Hard opt-out handling, global objection lists and Art. 21 suppression are features EU legal buyers ask for by name (GDPR and ePrivacy).
• Price against the meter. "Unlimited AI drafts, you pay per held meeting" is a line Instantly cannot match without repricing. See Pricing strategy.
• Publish classification accuracy. No vendor has independent numbers (Replies, booking and handoff). A published confusion matrix on a labelled sample would be the first one in the category.