If you Google "AI SDR" right now, you'll find about 40 vendors selling you a product that promises to "book meetings while you sleep." Most of them ship a chatbot you embed in your website and a templated email sequence you can already build in Smartlead.
That's not an AI SDR. That's a chat widget with extra steps.
A real AI SDR is an autonomous agent that handles the workflow of an SDR — sourcing, enriching, scoring, personalizing, sending, replying, escalating — across multiple channels, with judgment, evals, and human-in-the-loop fallbacks. We've shipped 11 of these for paying clients in the last year. The good ones book 18+ meetings/month. The bad ones get domain-blocklisted in week three.
Here's the difference.
What an AI SDR actually is
The job of an SDR has four phases:
- Source — find the right prospects.
- Engage — get them to reply.
- Qualify — figure out if they're a real fit.
- Schedule — get a meeting on the calendar.
An AI SDR is software that does all four phases autonomously, with a human reviewer in the loop on high-risk decisions, and a hand-off to a human (an AE or yourself) when the conversation gets meaty.
Critically: it's not a chatbot. The output of an AI SDR is a booked meeting in your AE's calendar, with the prospect briefed on what to expect. Not "a chat session." Not "a lead in HubSpot." A meeting.
The architecture
SOURCING TIER
├─ Apollo / LinkedIn Sales Nav / scraped intent signals
└─ Output: a ranked list of prospects, daily
ENGAGEMENT TIER
├─ Multi-channel sequencer (email primary, LinkedIn secondary, phone as escalation)
├─ Per-prospect personalization via LLM
└─ Output: emails / DMs sent, opens / replies tracked
REPLY HANDLING TIER
├─ Classifier: positive / neutral / negative / out-of-office / unsubscribe
├─ Drafter: Claude composes a reply per classification, with context from the original message + prospect data
├─ Approver: human reviews drafts in Slack (week 1–3), agent flips autonomous (week 4+)
└─ Output: replies sent, hot replies escalated to AE
SCHEDULING TIER
├─ Calendly / Chili Piper integration
├─ Agent proposes 3 specific times in time-zone (not "let me know what works")
└─ Output: meetings booked on AE calendar, prep notes attached
OBSERVABILITY TIER
├─ Every decision logged
├─ Slack weekly digest of stats + edge cases
└─ Output: trust, learning loop, eval feedback
If any one of those tiers is missing, you don't have an AI SDR — you have a partial automation.
The system prompt that works
The single biggest determinant of AI SDR quality is the system prompt. Here's an excerpt of the system prompt template we use, redacted:
You are an SDR Agent for [COMPANY]. Your role is to handle outbound and reply
work the way a thoughtful, direct human SDR would.
PRINCIPLES
- You always sound like a real human. You never reveal you are an AI unless directly asked, in which case you say "I'm an AI assistant working with [Name], happy to introduce you when we set up a call."
- You are honest. You never embellish capabilities, never lie about availability, never promise outcomes.
- You write in [COMPANY]'s voice: [SHORT VOICE GUIDE — e.g., direct, lower-case, no exclamation points, never use "synergy" or "leverage" as a verb].
ICP
[CONCRETE ICP DEFINITION — firmographic + role + intent signals]
DECISION RULES
- Score the prospect 0.0–1.0 on ICP fit before any outreach.
- Score >= 0.8: full personalization, book directly with [AE NAME].
- Score 0.5–0.79: lighter personalization, offer a relevant case study, book optional.
- Score < 0.5: do not pursue. Add to a long-term nurture list.
REPLY HANDLING
- Positive replies: book a meeting. Propose 3 specific times in their timezone. Never "let me know what works for you" — that's friction.
- Neutral replies: respond with one specific piece of value (case study, data point, peer reference). Re-offer the meeting.
- Negative replies: respect the no. Reply "Got it — appreciate the candor. If anything changes, the door's open." Add to suppression.
- Out-of-office replies: hold for [their return date + 1 day].
- Unsubscribe / spam-flag: process immediately, remove from all sequences.
ESCALATION
Escalate to a human reviewer in Slack any reply that:
- Asks a question outside the standard ICP qualification flow
- Contains a legal/contract reference
- Mentions a competitor
- Indicates a security/compliance concern
- Comes from a prospect with score >= 0.9
That's about 30% of the actual system prompt. The other 70% is voice training (5–10 example exchanges), edge cases, and tooling specifications.
The eval suite
This is where most teams die. They build the agent, ship it, and don't have a way to measure whether prompt changes make things better or worse.
A minimum eval suite for an AI SDR has 40–60 cases:
- 10 obvious-yes prospects (perfect ICP fit)
- 10 obvious-no prospects (spam, free email, students, competitors)
- 8 edge cases in sourcing (right industry / wrong size, right size / wrong role, right role / wrong title-spelling)
- 5 reply scenarios: positive, neutral, negative, OOO, unsubscribe
- 5 adversarial cases: jailbreak attempts, "ignore previous instructions," social engineering
- 5 voice tests: confirming the agent doesn't drift into corporate-speak
- 5 localization cases: non-US names, non-English names, time zones across continents
You run this suite against every prompt change, every model swap, every tool change. You track pass rate over time. Anything below 92% gets a fix before deployment.
The first time I helped a client build an eval suite they pushed back: "isn't this overkill for a sales agent?" Three weeks later they ran the suite against a new model release and found their pass rate had dropped 18 points. They would have shipped that to production without the eval. Imagine.
The rollout
The single biggest mistake I see teams make is to flip the switch too fast.
Week 1 — Build, don't send
Build the agent, build the eval suite, hit ≥92% pass rate before any production sending. This is also where you set up your sending domain warmup — usually a separate domain like mail.yourcompany.com, warmed for 14 days before any real volume.
Week 2 — Shadow mode
Agent drafts every email. Every draft lands in a Slack channel for human approval. Approve / edit / reject. The point is to catch the cases your eval suite missed.
By end of week 2, you should be approving 95%+ of drafts without edits. If you're still editing most drafts, the agent isn't ready and you need to fix the prompt.
Week 3 — Tiered autonomy
Start sending autonomously on the medium-fit prospects (score 0.5–0.79). Keep human-approval on high-fit (≥0.8) and edge cases. Watch reply rates daily.
Week 4+ — Full autonomy on most paths
Flip full autonomy for everything except escalation triggers (competitor mentions, legal questions, etc.). The agent is now operating. You're watching, not running.
Three of the AI SDRs we've shipped reached "full autonomy on most paths" by week 4. The other eight took 6–8 weeks. The difference is almost always how well-defined the ICP was at the start — fuzzy ICPs require more human-in-the-loop work.
What goes wrong
Here are the actual failure modes we've seen in production:
Domain reputation collapse. Someone bypassed warmup and started sending 500/day from a cold domain on day 3. Domain went on a major blocklist within 36 hours. Recovery took 4 weeks. Lesson: warmup is non-negotiable.
Prompt drift after model update. A model provider shipped a new version of their flagship model on a Tuesday. By Friday the agent's reply tone had drifted noticeably toward corporate-speak. We caught it because the weekly eval failed. Lesson: weekly eval cron.
The "we wrote your name wrong" failure. Enrichment returned an incorrect first name for a high-value prospect ("Mike" instead of "Michael"). Agent used it. Prospect noticed. Lost the deal. Lesson: defensive prompt — if first-name confidence is below a threshold, drop it from the salutation.
Autonomous unsubscribe failure. Our agent processed an "unsubscribe" reply as a soft no and sent a re-engagement email a week later. The prospect filed a complaint. We added a hard-classifier with no LLM judgment on unsubscribe signals. Lesson: deterministic rules for compliance-critical paths.
Champion blow-up. Agent had a great back-and-forth with a champion at a target account. Champion forwarded to their CEO. Agent's next reply assumed it was still in the original thread context and addressed the CEO by the champion's name. We were lucky — they laughed. Lesson: every reply re-loads the current recipient identity from the email headers.
The numbers
For a well-built AI SDR on a tier-1 ICP, here's what we expect at 90 days:
- Sourced volume: 5,000–8,000 cold contacts/month
- Reply rate (cold): 25–35%
- Positive reply rate: 8–14%
- Meetings booked: 12–20/month
- Pipeline coverage: 2–3× sourced pipeline target
Cost to operate (post-build): typically $400–$1,200/month in API + sending + tools, depending on volume. Build cost: $25K–$60K depending on integration complexity.
The math: at $50K AOV and even a 15% close rate on AI-booked meetings, you're looking at ~$1M of incremental pipeline per quarter and ~$150K of incremental closed-won. Even on the low end, the system pays for itself in the first three months.
When this doesn't work
AI SDRs don't work for every business. They don't work if:
- Your ICP is fuzzy or your sales motion is product-led (PLG companies often have better growth ROI elsewhere).
- Your target market is small (under 5,000 total addressable accounts), because you'll burn through them in a quarter.
- Your sales motion requires deep relationship-building (an AI SDR can open a door; it can't carry a 9-month enterprise sale).
- You don't have AEs or yourself ready to take handoffs — booked meetings that nobody attends are worse than no meetings.
If any of those describe you, look at marketing or ops automation first; come back to outbound when the foundations are right.
What to do next
If you want to build this yourself, start with our AI Agent Starter Kit. It includes the working code, the eval suite template, and the SDR system prompt as a starting point.
If you want to ship it without becoming an AI agent engineer, book a call. Most clients have their AI SDR live and booking meetings within 30 days.
— Drew Chimento