Nurture AI - Email Sequences that Convert - Vertical Agent Spec

One-line definition

An agent that writes, sends, monitors, and adjusts cold email sequences for B2B sales teams trying to book meetings without hiring a dedicated copywriter.

The workflow it owns end-to-end

Intake: User uploads a target list (company, name, title, LinkedIn URL) and describes their offer in plain language
Research: Agent pulls firmographic and signal data on each contact and selects a sequence template appropriate for the buyer's role and likely stage
Draft: Agent writes a 3-5 touch sequence with subject lines, body copy, and a call-to-action timed around inferred buying window; user approves before first send
Send and track: Agent dispatches via the user's connected mailbox, monitors open and reply rates, and adjusts send timing based on engagement patterns
Reply routing: Agent reads inbound replies, classifies them (interested, unsubscribe, objection, wrong person), drafts a suggested response for the rep, and halts the sequence on anything that needs human judgment

What it knows that a generic LLM doesn't

Deliverability mechanics: Gmail's bulk sender rules require one-click unsubscribe headers, sub-0.3% spam complaint rates, and authenticated sending domains; a generic LLM will write clean copy that still lands in spam
Sequence cadence norms: 4-7 touches over 14-21 days consistently outperforms both single-touch and high-frequency sprays; Tuesday-Thursday send windows outperform Monday and Friday across most B2B verticals
Reply classification patterns: knows the operational difference between "not interested" (stop the sequence), "not right now" (pause 30 days, resume), and "who are you" (pivot angle, not offer)
Spam trigger vocabulary: words like "free," "guaranteed," and all-caps phrases reliably suppress deliverability even when the surrounding content is legitimate
Offer-to-list fit diagnosis: knows that low reply rate with high open rate points to copy or offer, while low open rate points to list or subject line -- this is the misattribution problem the product's own risk register identifies, and it is real and chronic

What it explicitly declines

Sending to purchased or scraped lists without consent verification -- the agent will not touch a list it cannot confirm has a legitimate business relationship basis
Legal guidance on CAN-SPAM, GDPR, or CASL beyond flagging known patterns; users need outside counsel for jurisdiction-specific questions
Fabricating social proof, testimonials, or case study details; it will identify gaps in the copy and ask the user to supply real evidence
Committing to reply rates or meeting-booking outcomes; it will surface benchmarks but will not promise results it cannot control

Tools and integrations required

Gmail and Outlook via OAuth for authenticated sending from the user's own domain, not a shared IP pool
A dedicated email warm-up tool such as Instantly or Mailreach to build sending reputation on new domains before sequences start
Apollo.io or Clay for contact enrichment (title, company size, tech stack, recent funding or hiring signals)
HubSpot or Salesforce for CRM write-back so reply classifications and booked meetings surface in the rep's existing workflow rather than a parallel dashboard
Unsubscribe list management synced across all active sequences to prevent compliance gaps

Trust escalation: when it pings a human

A reply contains a legal threat, a cease-and-desist request, or any mention of reporting the sender -- agent halts the sequence immediately and does nothing further until a human reviews
Reply classification confidence falls below a defined threshold; agent shows its top two classifications and asks the user to pick before drafting a response
A sequence reaches 500 or more contacts without the user having reviewed at least one live send sample -- agent pauses and requests a spot-check
Any contact at a company the user has flagged as an active opportunity -- those threads belong to the rep, not the sequence

Pricing model

The structurally honest model here is per-attributed-meeting-booked at roughly $45 to $60 per confirmed calendar hold that traces back to an agent-run sequence, with a monthly floor of $150 to cover infrastructure. The problem is attribution: a prospect who receives three cold emails, then a LinkedIn message from the rep, then books a call is a partial-credit situation at best, and disputes will be frequent. A more operationally defensible structure is $99 flat per month plus $20 per attributed booking, which gives the buyer a predictable floor and gives the agent skin in outcomes without turning every booking into an argument. Whether a sales manager at a 30-person company pays that readily depends on whether they can see the sequence-to-meeting trace clearly. If the reporting is opaque, churn follows.

Differentiation from a generic LLM wrapper

The honest answer is that the differentiation window is narrow. By mid-2025, HubSpot, ActiveCampaign, and Klaviyo all have sequence suggestion features, and the gap between "vertical agent" and "incumbent add-on" is measured in months. Where a dedicated agent holds ground is in the operational layer: authenticated mailbox connections, warm-up coordination, real-time reply classification piped back into a CRM, and the institutional knowledge of what a 0.35% spam complaint rate costs a domain long-term. A user pasting their workflow into Claude gets words. They do not get a system that sends, tracks, classifies, and routes. The agent's value is running the full loop, not writing better copy. Whether that loop is worth $99 a month once HubSpot bundles the same loop into a product the buyer already pays for is the question this product needs a concrete answer to before month six.