ai-sdr-setup

Files (5)

• SKILL.md
• framework-notes.md · references/
• justin-michael-sales-borg.md · references/
• check-output.py · scripts/
• output-template.md · templates/

SKILL.md (first lines)

# AI SDR Setup

## Overview

Most AI SDR deployments fail because the team treats the agent like a fully autonomous rep instead of a constrained workflow engine. The mistake is giving the agent a list, a writing prompt, and send access before defining quality gates, source-of-truth data, suppression logic, and human handoff rules.

This skill builds an AI SDR launch program that starts narrow, validates message quality, routes the right signals to the right action, and keeps humans in control of anything that can affect brand, compliance, or customer relationships.

## When to Use

Use this skill when the user asks to:

- "set up an AI SDR"
- "automate outbound prospecting"
- "deploy 11x / Artisan / AiSDR / Jason AI"
- "build an AI sales agent"
- "replace manual SDR research"
- "automate prospect research and email drafting"
- "create AI-powered sales development"
- "route buying signals into outbound"
- "design human handoff for AI replies"

Do not use this for generic outbound strategy. Use `cold-email-strategy` first if the user has no ICP, segment, offer, or sequence design.

## Authoritative Foundations

### Anthropic — Tool Use and Human-in-the-Loop Patterns
AI agents should be given constrained tools, clear success criteria, and explicit approval gates for external actions. For SDR work, drafts can be automated, but sending, CRM mutation, and sequence enrollment need controls.

### ColdIQ — Signal-to-Action Routing
ColdIQ's automation patterns map observed signals to deterministic plays: hiring signal → hiring-specific sequence, funding signal → expansion or scale message, job change → champion-tracking play. AI SDRs work best when the routing is deterministic and the writing is assisted.

### Winning by Design — Bowtie Revenue Model
Outbound is only one stage of the revenue bowtie. AI SDRs should optimize for qualified pipeline and customer fit, not raw activity volume. The handoff to AE and CS must preserve context.

### 11x / Artisan / AiSDR / Jason AI — AI SDR Category Patterns
AI SDR platforms converge around the same operating model: data input, account/contact research, sequence generation, reply classification, and human handoff. Vendor choice matters less than workflow control.

### Justin Michael — Sales Borg (*Tech-Powered Sales*)
Michael's first book frames outbound automation as **semi-automation**, not full replacement: machines handle data, triggers, cadence execution, and A/B tests; humans handle empathy, referrals, strategic accounts, and all high-stakes replies. RevOps **orchestrates bots** — CRM, enrichment, and SEP must sync with suppression governance. AI SDR pilots should mirror this split: automate Tier 1–2 workflow steps, keep humans on send approval (week 1), positive replies, pricing, and legal threads. Part III caveat: bots can replace humans for commodity buying, not complex B2B value selling. Playbook → `cold-email-strategy/references/justin-michael-sales-borg.md`.

## Prerequisites

- ICP definition with exclusions
- Suppression list and existing customer list
- Approved messaging examples
- CRM ownership rules
- Sequencer account with sending limits
- Human reviewer assigned for the pilot
- Metrics baseline from human SDR or previous outbound motion

## Step-by-Step Process

### Phase 1: Scope the Pilot (Day 0)

| Decision | Recommended Starting Point | Why |
|---|---|---|
| Contact count | 100-200 contacts | Enough signal without large brand risk |
| Segment | One ICP segment only | Prevents mixed messaging |
| Channels | Email only at first | Easier QA and compliance |
| Send access | Draft-only for week 1 | Prevents bad autonomous sends |
| Human review | 100% of messages | Builds quality baseline |

Create a pilot charter: target segment, excluded segments, message variants, approval workflow, stop conditions, and success metrics.

### Phase 2: Configure Data Inputs (Day 1-2)

The AI SDR should receive clean inputs, not infer everything itself. Required fields: account name, domain, segment/tier, persona, trigger signal, source URL or evidence, relationship context, and suppression status.

Never let the agent create contacts from vague search prompts without source attribution. Every generated contact needs a source and a verification status before sequencing.

### Phase 3: Build Signal-to-Action Rules (Day 2-3)

| Signal | AI Action | Human Action |
|---|---|---|
| Funding announced | Draft funding-specific opener | Approve target account list |
| Hiring for GTM roles | Draft hiring-pain sequence | Confirm persona relevance |
| Job change | Draft champion/new-role message | Confirm relationship context |
| Competitor mention | Draft displacement email | AE reviews before send |
| Pricing question reply | Classify and escalate | AE responds manually |

The key: the agent does not decide the play from scratch. It selects from approved plays based on observed signal.

### Phase 4: Set Guardrails (Day 3-4)

Hard rules the pilot must enforce before anyone gets send access:

| Guardrail | What It Means in Practice |
|---|---|
| Draft-only pilot | Week 1: agent writes drafts only. A human approves every message before it leaves the sequencer. |
| Send caps | Cap daily and weekly sends per inbox and per segment. Start low (e.g., 20-30/day per mailbox) and raise only after quality holds. |
| Suppression | Block customers, competitors, opted-out contacts, and do-not-contact domains before any enrollment. Re-check suppression on every batch. |
| Kill switch | One owner can pause all agent sends instantly — disable the sequencer connection, revoke API keys, or flip the vendor "pause" toggle. Document who holds the switch and how to trigger it. |
| Budget caps | Set monthly credit/API spend limits on enrichment and AI SDR vendor accounts. Alert at 80%; hard stop at 100%. |
| No invented facts | Every personalized claim needs a source URL or verified data field. If the agent cannot cite it, cut the line. |
| No sensitive attributes | Do not reference health, religion, politics, family status, or other protected personal details — even if public. |
| No commitments | Agent cannot promise pricing, discounts, security certifications, legal terms, or procurement timelines. |
| Human handoff | Positive replies, pricing questions, security reviews, legal requests, and procurement threads go to a named human within one business day. |
| CAN-SPAM basics | Include a physical mailing address, working unsubscribe link, and accurate "From" identity. Honor opt-outs within 10 business days. No deceptive subject lines. |

These are operational controls, not optional nice-to-haves. If any guardrail cannot be enforced with the current tooling, fix the tooling before scaling.

### Phase 5: Pilot and QA (Week 1-2)

| Criterion | Pass Standard |
|---|---|
| Relevance | Message ties to account/persona/signal |
| Evidence | Personalization has a source |
| Brevity | Under 90 words for first email |
| CTA | One clear, low-friction ask |
| Compliance | Unsubscribe and suppression respected |
| Voice | Sounds like your company, not generic SaaS |

### Phase 6: Controlled Scale (Week 3-4)

Only scale after quality holds for two consecutive review cycles. Move from 100% human review → 50% sampled review → exception review for high-risk accounts → weekly random QA.

## Output Format

From LeadMagic/gtm-skills, MIT.

More from LeadMagic/gtm-skills