AI SDR stack

The default stack for a small RevOps team running outbound with an AI SDR agent doing the work and a human reviewing the drafts before they hit prospect inboxes. Six tools, one motion: signal-aware list, agent-written first touch, human-checked send, deliverability-safe infrastructure, CRM sync. The piece most teams skip — the QA layer — is the difference between “we rented an SDR” and “we torched our domain.”

How the pieces fit

• 11x or Artisan is the agent. One of the two does the SDR work: ICP research per account, multi-step email and LinkedIn sequencing, reply handling, meeting booking. 11x (Alice + Julian) is the pick when you also want a phone agent for inbound qualification — median Vendr contract $40K/year. Artisan (Ava) is email-and-LinkedIn only, with built-in contact data and warmup, at a lower median ($25-30K/year). Pick one; running both is overlap, not coverage.
• Clay is the signal and list layer feeding the agent. ICP list-build from seeds (job titles, technographics, intent), waterfall enrichment across 100+ providers, AI columns that score fit and pre-research each row before the agent ever sees it. The agent’s output quality is mostly a function of what enters it — Clay is where you control that input. From $349/mo (Explorer); a 5-seat team typically runs Pro at $800+/mo.
• The AI SDR draft QA skill is the human-review gate. Before the agent sends, a Claude Skill flags claim accuracy (does the agent’s “I saw you just launched X” hallucinate?), personalization quality, and compliance (CAN-SPAM, GDPR, one-click unsubscribe). A revops or SDR-manager spends 15-30 minutes a day reviewing flagged drafts. This is the only reason the agent doesn’t burn your domain on month two.
• Smartlead is the sending infrastructure. Both 11x and Artisan offer their own sending, but on volume above ~3,000 sends/day, or when you want a secondary brand sending separately from your primary domain, Smartlead’s unlimited-mailbox rotation, warmup pool, and SmartDelivery placement tests are the upgrade. From $174/mo plan + $50-200/mo mailboxes and SmartDelivery — budget the all-in number, not the plan price.
• Salesforce or HubSpot is the CRM source of truth. Every account, every activity, every reply lands in CRM — not in the agent’s parallel database. Salesforce for >$10M ARR with complex pipeline; HubSpot for SMB-to-mid-market who’d rather not staff a Salesforce admin.

The handoffs

1. Clay → agent. Clay builds the daily target list, enriches each row, and writes scored accounts to a CRM custom object or directly to the agent’s worklist. The agent runs only against ICP-qualified rows; ICP rejects never reach it.
2. Agent → QA skill. Drafts queue in a holding state (11x’s “Review Mode,” Artisan’s draft queue, or a dedicated Slack channel for the skill to read). The QA skill flags drafts that fail the rubric; approved drafts release for send.
3. QA → Smartlead (or native send). Approved drafts dispatch through the configured sending path. The email-deliverability monitor watches DMARC failures and complaint rate; volume throttles down automatically when complaint rate crosses 0.08%.
4. Send → CRM. Every send, open, reply, and booked meeting logs back to Salesforce or HubSpot through the agent’s native sync. The CRM is where the AE picks up after the meeting books.

Cost baseline

A 5-seat outbound team running this stack end-to-end:

• AI SDR agent: $25-50K/year (Artisan low, 11x high)
• Clay Pro: ~$10K/year
• Smartlead all-in (plan + mailboxes + SmartDelivery): ~$4-6K/year
• CRM (HubSpot Sales Hub Pro 5 seats: ~$6K/year; or Salesforce Sales Cloud Enterprise 5 seats: ~$10-15K/year)
• QA skill: free (open-source Claude Skill), but ~10-20 hours/week of a manager’s time to run

Annual range: $45-80K/year depending on AI SDR choice and CRM. That sits at roughly half a fully-loaded mid-market SDR salary — the buyer math the AI SDR vendors quote. The honest version: the agent replaces SDR labor at the prospecting end; you still need an AE to take the meeting and a manager to run the QA loop.

Common variations

• DIY agent path. Replace 11x/Artisan with Clay plus Claude (the lead enrichment Skill is the reference) plus Outreach or Salesloft for sequencing. Trades a ready-made agent for control over signal, copy, and cost — typically cuts spend in half but adds 4-8 weeks of GTM-engineer build time. Pick this when you have an in-house GTM engineer and want to own the loop.
• Signal-driven, human-in-the-loop instead of fully autonomous. Swap 11x/Artisan for Unify: warm signals trigger plays, but humans write and send. Reply rates run higher; volume runs lower. Pick when ACV is high enough that fewer-but-better touches beat agent volume.
• Hybrid AI-SDR + human SDR. Keep one human SDR on enterprise/strategic outbound (where the agent’s reply quality demonstrably trails a trained rep) and run the AI SDR on SMB and mid-market only. ICP routing in Clay decides which list each account lands on.
• Inbound-led with AI phone agent. If inbound is the bulk of the motion, pair 11x’s Julian (the phone agent) with RevenueHero or Default for routing; drop Artisan as an option (no voice). The inbound conversion stack is the better reference once it ships.

What this stack does NOT replace

• An ICP rubric and a calibrated definition of “qualified meeting.” Without those upstream, the agent generates volume and the QA layer drowns. See ICP and lead scoring with an ICP rubric.
• A conversation intelligence layer for the meetings the agent books. The AE motion after the meeting needs Gong or Chorus — the AI SDR doesn’t.
• Demand generation. Outbound feeds on signals — intent data, product usage, content engagement, ecosystem events. None of those come from inside the stack; they come from paid, content, partnerships, or a signal orchestration program upstream.
• The AE. The AI SDR books meetings; closing is still a human job, and on B2B ACVs above $25K it’s the gating skill, not prospecting volume.
• A meeting scheduler that survives reschedules at scale. Calendly, Chili Piper, or RevenueHero sits between the agent and the AE’s calendar.

Match rules — when this stack is the right pick

• Right pick: B2B SaaS at $2-20M ARR, North-American TAM, one to three SDR seats you would otherwise hire, a manager who can spend 10-20 hours/week on QA and signal calibration, and a willingness to pilot for 90 days before judging reply rates. CAC math that works at $40-60K agent cost producing 8-15 qualified meetings/month.
• Wrong pick — go human SDR + Apollo instead: Enterprise outbound above $100K ACV, regulated industry where every touch needs compliance review, EMEA/APAC as primary geography (both 11x and Artisan have thin non-NA data), or a team without anyone to own the QA layer. A fully-loaded human SDR on Apollo runs $80-120K/year and is still the right answer when complexity wins over volume.
• Wrong pick — go DIY (Clay + Claude + Smartlead) instead: You have an in-house GTM engineer, you sell to a niche ICP where prebuilt agent personalization is generic, and you’d rather spend $15-25K/year on infrastructure than $40-50K on an agent. The trade is 4-8 weeks of build time and an ongoing maintenance load.
• If reply rates haven’t crossed your bar by day 90, kill the agent contract at the next opt-out window. The stack assumes the pilot gate is real and enforced; treat the first year as committed only if pilot data justifies it.