── ── AI & agents

Agent Outreach & Lead Nurturing

AI agents can research every prospect before writing to them, at close to the cost of a mass blast. The same agents make it just as cheap to send spam at scale, which damages domain reputation and the brand. Optimizing for send volume brings low reply rates and rising complaints. Then the domain lands on blocklists, and even normal email…

How it works

Output: Outreach Spec (ICP / signals / sequence / gates / compliance / metrics)

1. ICP + exclusion list. Industry, size, role, tech stack. List who is off-limits: current customers, competitors, restricted regions, anyone who has opted out on any channel. Gate: the ICP can be written as an executable filter → otherwise stop. 2. Trigger signals. Define public "contact now" signals: hiring for a role, a funding round, a product launch, a technology change, an active site visit. Gate: each signal's source can be verified by the agent → otherwise drop it. 3. Sourced personalization. The agent researches public information. The first sentence links the signal to the prospect's likely pain. Every personalized claim carries a source link. Gate: delete any personalization without a source; never invent facts or relationships. 4. Sequence design. 3–5 touches across email, LinkedIn and content. Each touch adds new value (a case, a checklist, a diagnostic), never another "just following up." Set a maximum total duration. Gate: every touch has standalone value → otherwise merge or cut it. 5. Human approval gate. A human reviews new sequences, new segments, the first batch, and high-value accounts. Lower the spot-check rate only as quality stays stable. Gate: an unapproved template is never sent. 6. Compliance and deliverability. U.S. CAN-SPAM: accurate header and sender identity, non-deceptive subject line, valid physical postal address, a clear opt-out honored promptly, and responsibility that stays with you even when a vendor sends. EU/UK: a confirmed lawful basis (legitimate interests for B2B requires an assessment) and the UK PECR rules (sole traders and some partnerships count as individuals; named work addresses at companies still carry data-protection duties). Mailbox providers: SPF, DKIM and DMARC authentication; one-click unsubscribe (RFC 8058) for bulk sends; complaint-rate and volume limits; warm up new domains. Pull current thresholds and deadlines from the live regulator and provider pages; never state them from memory. Gate: any item unmet → stop sending. 7. Reply handling. The agent classifies replies: interested / later / no / unsubscribe / referral / auto-reply. "No" and "unsubscribe" stop all channels immediately. "Interested" goes to a human within an agreed time. 8. Metrics and iteration. Track positive-reply rate, meetings, complaint rate, bounce rate and unsubscribe rate, split by signal and segment. Stop-rule: complaints or bounces cross the pause threshold → pause all sequences and return to steps 1–3.

Output template: Segment · Trigger signal · Personalization source (link) · Touches 1–5 (channel/value) · Approver · Compliance check · Positive-reply % · Complaint % · Next step

When to use it

• user says 'use an AI agent for outreach / cold email', 'automate lead follow-up', 'multi-touch nurture sequence', 'personalized prospecting emails', asks 'how do we let the agent send email without it looking like spam?', or is designing an agent-driven outbound and lead-nurturing process

Worked example

United States v. Verkada — 30 Million Prospecting Emails and the Largest CAN-SPAM Penalty…

This is not an AI-agent case. The conduct dates from 2019–2021. It is included because it is the clearest public record of what happens when B2B outbound scales faster than its compliance plumbing, and that is the risk an agent multiplies. Facts come from the FTC's August 30, 2024 press release and Verkada's own statement the same day. The claims are allegations: Verkada settled without agreeing with them, and its statement says, "We do not agree with the FTC's allegations."

Related mental models