AI sdr setup skill

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

# AI SDR Deployment Plan

## Pilot Scope
- Segment:
- Personas:
- Contact count:
- Channels:
- Send permissions:

## Signal-to-Action Matrix
| Signal | Evidence Required | AI Draft | Human Review | Handoff Rule |
|---|---|---|---|---|

## Guardrails
- Hard stop rules:
- Suppression rules:
- Approval gates:

## 4-Week Rollout
Week 1:
Week 2:
Week 3:
Week 4:

## Metrics
- Positive reply rate:
- Meeting conversion:
- Bounce rate:
- Bad-draft rate:
- Human override rate:

Quality Check

Before delivering, verify:

• Pilot is limited to one ICP segment
• Human review gates are explicit
• Suppression and compliance rules are included
• Every trigger requires source evidence
• Handoff rules cover positive replies, objections, pricing, security, and legal topics
• Metrics include quality and safety, not just volume
• No autonomous send access is recommended before QA passes

Common Pitfalls

1. Starting with auto-send. AI can produce plausible but wrong copy. Fix: draft-only pilot first.
2. Mixed ICP pilot. SMB, mid-market, and enterprise require different messages. Fix: one segment per pilot.
3. No suppression logic. Agents can accidentally contact customers, competitors, or opted-out contacts. Fix: suppression check before every sequence enrollment.
4. Measuring only activity. Emails sent is not success. Fix: track positive replies, meetings, bad-draft rate, and human override rate.
5. No handoff map. Agents keep replying when a human should intervene. Fix: hard handoff on positive intent, pricing, security, procurement, or legal.
6. Unverified personalization. AI invents reasons to reach out. Fix: require source URL or evidence for every personalized claim.

Execution Artifacts

• references/framework-notes.md — named frameworks, citation anchors, and operating assumptions
• templates/output-template.md — copy-paste deliverable structure for the user
• scripts/check-output.py — local checklist validator for required sections This skill includes lightweight artifacts the agent can load on demand:
• references/justin-michael-sales-borg.md — Sales Borg human/bot division, TQ, guardrails alignment (canonical) Use the artifacts when the user asks for an implementation-ready deliverable, a repeatable workflow, or a quality check rather than generic advice.

Related Skills

• cold-email-strategy — sequence and message architecture
• signal-scoring — which triggers should activate outreach
• reply-handling — classification and escalation rules
• mcp-setup — tool permissions and audit logging for agent workflows