If your pipeline depends on referrals and you have decided outbound is the fix, the next question arrives fast: do you hire a sales development rep, or do you set up an AI SDR? Most of what you will read on this is written by people selling one side. This guide breaks the SDR job into its parts, shows which parts AI does well and which still need a person, and gives you a worksheet to run the numbers for your own business.
An AI SDR is a software agent that does the research, list building, personalised first touches and follow-ups a sales development rep would do. Use it for volume and consistency; use a human SDR for live conversations, judgement and calls. For most B2B service businesses the best answer is a hybrid: AI prepares and sends, a person owns every reply.
Disclosure: some tool links in this guide are referral links from our AI Stack page. If you sign up through them we may earn a commission at no extra cost to you. We list tools we use or have vetted, and you can run every step here with other tools.
Below you will find a task-by-task comparison, a cost worksheet you fill in with your own figures, a decision guide, the hybrid operating model we use, the rules for the AI side, a worked example, a two-week pilot to test it, and the metrics that tell you whether it is working. For the full outbound system around this decision, see AI agents for lead generation and signal-based outbound.
What an SDR actually does, task by task
"AI SDR vs human SDR" is the wrong framing until you split the role into tasks. An SDR's week is mostly research, writing, sending, chasing and sorting, with a smaller amount of live conversation. AI is strong at the first group and weak at the second.
Asset 1: SDR task split| Task | AI SDR | Human SDR | Our recommendation |
|---|---|---|---|
| Building the account and contact list against an ICP | Strong: fast, consistent filtering and enrichment | Slow and inconsistent at volume | AI, with a human spot-check of 10 rows per batch |
| Verifying emails and data | Strong with verification tools | Manual and error prone | AI plus tools |
| Researching a company before outreach | Strong at summarising public sources | Deeper on a few accounts | AI for most, human for top-tier accounts |
| Writing personalised first lines and first touches | Good when given verified facts and strict rules | Good but slow | AI drafts, human approves until quality is proven |
| Sending sequences and follow-ups on time | Strong: never forgets | Inconsistent when busy | AI |
| Classifying replies | Good at sorting into clear buckets | Good | AI classifies, human reads every positive or unclear reply |
| Answering interested prospects and handling objections | Risky: tone and judgement errors cost deals | Strong | Human |
| Qualifying on a live call | Limited | Strong | Human |
| Booking the meeting | Good once the prospect has said yes | Good | Human sends times; booking link can be automated |
| CRM updates and notes | Strong | Often skipped | AI |
| Feeding market insight back to sales | Can summarise patterns | Strong on nuance | Both: AI summary, human interpretation |
Read down the last column and a pattern appears: AI does the work before the first reply and the admin after it; a person owns the conversation in the middle. That is the hybrid model this guide recommends for most teams.
AI SDR vs human SDR: the comparison that matters
| Factor | AI SDR | Human SDR |
|---|---|---|
| Cost structure | Tool subscriptions, data and usage costs, plus setup time | Salary, benefits, tools, management time, recruitment |
| Time to first output | Days to weeks once the ICP and messaging are clear | Recruitment plus a ramp period before full output |
| Volume | Scales with data and sending infrastructure, not hours | Limited by hours in the day |
| Consistency | Does the same thing every time, including the same mistakes | Varies with energy, workload and skill |
| Personalisation depth | Good on public facts; can sound generic without strict rules | Can go deep on a handful of accounts |
| Judgement and tone in conversation | The weak spot | The strong spot |
| Risk | Sending at scale can damage domains and reputation fast if quality slips | Lower blast radius, slower to make mistakes at scale |
| Management | Needs an owner who reviews output, data and deliverability weekly | Needs coaching, targets and one-to-ones |
| What you keep if it ends | The prompts, workflows and data, in your own accounts | Whatever they documented before leaving |
Two things this table does not say. An AI SDR is not "set and forget": someone has to own it. Our AI outbound agent builds are designed for an in-house GTM engineer, SDR or marketing specialist to run in as little as one to two hours a week once live, but that person has to exist. And a human SDR is not automatically better at writing: without a clear ICP and message, they will send generic emails too.
Cost: run the numbers with your own figures
We are not going to quote salaries or tool prices. They vary by country, seniority, market and plan, and any number we printed would be wrong for most readers. Instead, fill in this worksheet with real quotes and your own payroll costs. It takes 20 minutes and gives you a defensible answer.
Asset 2: SDR cost worksheetPERIOD: first 6 months (use the same period for every option)
OPTION A: HUMAN SDR
A1 Monthly salary or contractor fee ..................... <AMOUNT>
A2 Benefits, visa or payroll costs per month ........... <AMOUNT>
A3 Tools per seat per month (data, sending, LinkedIn) .. <AMOUNT>
A4 Recruitment cost (one-off) .......................... <AMOUNT>
A5 Your management time per week (hours) x your hourly
value x 4.3 ......................................... <AMOUNT>/month
A6 Months before full output (ramp) .................... <MONTHS>
TOTAL A = (A1 + A2 + A3 + A5) x 6 + A4
OPTION B: AI SDR
B1 Data and enrichment per month (at your volume) ...... <AMOUNT>
B2 Sending infrastructure and tools per month .......... <AMOUNT>
B3 AI model or platform usage per month ................ <AMOUNT>
B4 Setup (your time or a build fee, one-off) ........... <AMOUNT>
B5 Owner's time per week (hours) x hourly value x 4.3 .. <AMOUNT>/month
B6 Person who handles replies (hours/week x rate x 4.3) <AMOUNT>/month
TOTAL B = (B1 + B2 + B3 + B5 + B6) x 6 + B4
OPTION C: HYBRID (AI does prep and sending, a part-time human owns replies)
Use B1 to B5, and replace B6 with the part-time setter's monthly cost.
THEN, FOR EACH OPTION
Expected qualified meetings per month (from your pilot, not a guess) .. <N>
Cost per qualified meeting = monthly run cost / meetings ............... <AMOUNT>
Break-even meetings per month = monthly run cost / (deal value x close
rate) .................................................................. <N>Two rules for reading it. First, never forget line B6: someone has to answer the replies an AI SDR creates, and their time is a real cost. Second, do not fill "expected meetings" from a vendor's case study. Run the two-week pilot below and use your own number.
The decision guide: when to use each
Asset 3: Decision rules- If you do not yet have a clear ICP and an offer that has closed deals, do neither at scale. Write the ICP and message first; send 50 manual, carefully researched emails yourself and see what replies come back.
- If your deals are large, your buyers are few and every account is known by name, use a human SDR (or the founder) and give them AI for research and admin.
- If your market is broad, your offer is clear and the bottleneck is volume of good first touches, start with an AI SDR plus a person who owns replies.
- If your prospects expect a phone call or a long relationship before a meeting, keep a human SDR for calls and use AI to prepare their call list and notes.
- If nobody in the business can own a system for one to two hours a week, do not buy an AI SDR yet; it will drift. Hire or assign the owner first.
- If you already have a human SDR who is drowning in research and admin, keep them and add AI underneath them. This is often the fastest win.
- If your domain reputation or brand would suffer badly from one bad batch, keep human approval on every first touch until four weeks of output pass your QA checklist.
The hybrid model: who does what
This is the operating model we use in our own outreach and recommend to clients. The AI handles the steps before a reply and the admin after it; a named person owns every conversation.
Asset 4: Hybrid operating model and handoff rules| Stage | Owner | Rule |
|---|---|---|
| List building and enrichment | AI | Only confirmed decision makers at ICP-fit companies; pilot 10 to 20 leads before scaling and estimate cost from the pilot |
| Research and first line | AI | One real, verifiable fact about their business; no guessed numbers |
| First touch approval | Human | Approve every message in weeks 1 to 4, then sample 10 per batch |
| Sending and follow-ups | AI | Short sequence, then stop; any reply stops the sequence |
| Reply classification | AI | Sorts into interested, question, not now, not interested, wrong person, out of office |
| Interested reply | Human | Reply within a few working hours with two time options or a booking link; no long explanation first |
| Question | Human | Answer in two or three lines, then offer the call |
| Not now or not interested | Human | Thank them, offer a soft next step only if they asked, suppress if they opted out |
| Wrong person | Human | Add to the blocklist so they are never contacted again |
| CRM update | AI | Every outcome logged the same day |
Tools that can run these steps
You can test everything in this guide with a spreadsheet and an AI assistant. When you build the AI side, these tools from our stack fit each step. Descriptions follow our AI Stack page; check each vendor's documentation for current features, limits and pricing.
| Step | Tool | How it fits |
|---|---|---|
| Sourcing and enrichment | Deepline with Claude Code | GTM data infrastructure built for AI agents (waterfall enrichment, lead and account discovery, signal detection) run through a Claude Code skill |
| Finding verified emails | Findymail | Finds and verifies decision-maker emails so lists convert instead of bouncing |
| Second-layer validation | ZeroBounce | A second layer of email validation before any list goes live |
| Sending inboxes | Zapmail | Done-for-you inbox infrastructure so sending domains stay warm and out of spam |
| Cold email sending | Instantly | Cold email sending infrastructure with built-in deliverability |
| LinkedIn outreach | HeyReach | Multi-account LinkedIn outreach built for running outbound at scale |
| Workflow glue | Make or n8n | Connects sourcing, sending, reply classification and the CRM |
| CRM | Pipedrive or Breakcold | A simple visual pipeline, or a social-selling CRM built around LinkedIn engagement |
| Booking | Cal.com | Our preferred booking tool for the link the human sends to interested prospects |
If you go the human route, the same tools make one SDR far more productive. If you need the person rather than the system, we also place trained appointment setters.
The rules for the AI side
Whatever tool or assistant you use to draft and sort, these are the rules we hold it to. Write them into your assistant's instructions once and check them in every batch.
Asset 5: First-touch and reply rules| Area | Rule |
|---|---|
| Facts | Use only verified facts: one recent, public, professional fact about the prospect's business, with its source link. If there is no usable fact, skip the prospect. |
| First line | Under 15 words, positive and specific, and never mention how you found it. |
| Message shape | First line, then one sentence linking the fact to the problem you solve, phrased as a question, then one low-effort ask. |
| Length | Email: subject under 7 words and under 90 words in total, no links or images. LinkedIn: under 80 words, conversational. |
| Claims | No statistics, results, client names or promises you cannot prove. No flattery words or emojis. |
| Reply categories | Interested, question, not now, not interested, wrong person, out of office, unsubscribe, unclear. |
| Reply handling | Any reply stops the sequence. Interested, question and unclear always go to a named person. Unsubscribe and not interested are suppressed after a person confirms. The AI never answers an interested prospect on its own. |
| Suggested replies | For interested prospects, two specific time options or your booking link and nothing else. Never invent prices or results. |
A worked example (illustrative)
This example is fictional. It shows the method, not a real client or a real result.
The business: "Harbourline Advisory", a six-person operations consultancy selling process improvement projects to logistics firms. The founder closes well but only has a handful of new conversations a month, all from referrals. They are weighing a full-time SDR against an AI SDR.
Task split (abridged): of a planned 40-hour SDR week, list building, research, first drafts, follow-up scheduling and CRM updates (about 26 hours) were marked "automate fully" or "AI drafts, human approves". Replying to interested prospects, discovery calls and objection handling (about 14 hours, though far less at the start) were marked "keep human". Riskiest automation: sending first touches without review, because the logistics niche is small and one bad batch would be noticed.
Asset 2 result: using their own quotes, the founder found the full-time hire cost more per month than the hybrid option once the ramp period and management time were included. The hybrid option only became cheaper when line B6 (reply handling) was filled with a part-time setter rather than the founder's own hours.
Asset 3 decision: clear ICP and proven offer, broad enough market, bottleneck is volume of good first touches, and the operations manager can own the system for two hours a week. Decision: hybrid, with human approval of every first touch for four weeks.
A first touch for one prospect:
Subject: New Jebel Ali warehouse
Hi Dana, congrats on opening the new Jebel Ali warehouse last month.
Adding a second site often exposes handoffs that worked fine with one.
Is cross-site order routing something your team is working on right now?
Worth a quick chat?
SamA reply ("Maybe, what does this usually involve?"): classified as a question, sent to a person with a suggested answer: "Usually a two-week review of how orders move between sites, then a short plan. Happy to walk you through it: does Tuesday 10:00 or Wednesday 15:00 work?" The part-time setter edits one word and sends it within the hour.
Prove it with a two-week pilot
- Days 1 to 2: write the ICP in one paragraph and the offer in one sentence. Split the tasks with Asset 1. Fill in Asset 2 with real quotes.
- Days 3 to 4: build a pilot list of 10 to 20 accounts, check every row by hand, and estimate the cost of a full batch from the pilot before scaling.
- Days 5 to 6: draft first touches with your AI assistant following the rules in Asset 5. Approve or edit each one. Count how many you had to rewrite.
- Days 7 to 10: send to a small batch from an authenticated secondary domain. Classify every reply into the categories in Asset 5 and have the human respond.
- Days 11 to 14: fill in the KPI table, compare cost per qualified conversation with the human-only option, and decide: scale, adjust or stop.
Deliverability and data rules either way
Whether an AI or a person sends the email, the inbox providers judge the same signals. Google's sender guidelines require all senders to Gmail accounts to set up SPF or DKIM, keep the spam rate reported in Postmaster Tools below 0.3%, and format messages to the standard; senders of more than 5,000 messages a day to Gmail accounts must also set up SPF, DKIM and DMARC, align the From domain, and support one-click unsubscribe for marketing messages. Volume is where an AI SDR can hurt you fastest, so keep batches small until the data says otherwise.
On personal data, rules differ by country. The UAE government portal, for example, describes the Personal Data Protection Law as prohibiting the processing of personal data without consent except in specific cases. Check the rules that apply to you and your prospects, contact professional addresses with relevant business messages, and honour every opt-out immediately.
Common mistakes and how to fix them
| Mistake | What happens | Fix |
|---|---|---|
| Buying an AI SDR before the ICP and offer are clear | Faster delivery of the wrong message | Prove the message with 50 manual sends first |
| Letting the AI answer interested prospects | Awkward replies lose warm leads | Human owns every interested or unclear reply |
| No owner for the system | Data and copy quality drift for weeks | Named owner, weekly review, written in their job description |
| Hiring a human SDR and giving them no system | Long ramp, generic emails, early churn | Give them the ICP, prompts, verified data and the reply rules on day one |
| Scaling volume before deliverability is set up | Domains burned, emails in spam | Authenticated secondary domains, small batches, monitor spam rate |
| Comparing costs without management and reply time | The "cheap" option is not cheap | Fill every line of Asset 2, including A5, B5 and B6 |
| Judging on opens or sends | Optimising activity, not meetings | Judge on positive replies and qualified meetings |
What to measure and how to read it
Asset 6: SDR KPI table| Metric | How to calculate | How to read it | If it looks wrong |
|---|---|---|---|
| Rewrite rate | AI drafts you had to rewrite divided by drafts reviewed | Falling means the prompt and data are improving | Tighten the Asset 5 rules; improve the facts you feed it |
| Bounce rate | Bounced divided by sent | Data quality check | Add or tighten verification before sending |
| Positive reply rate | Interested replies divided by delivered first touches | The core quality signal; compare against your own baseline | Review ICP fit and the first line |
| Speed to human reply | Hours from interested reply to human response | Warm leads cool quickly | Alerts to the owner; set a response-time target |
| Qualified meetings per month | Meetings that met your criteria | The outcome | Check handoff speed before blaming the copy |
| Cost per qualified meeting | Monthly run cost divided by qualified meetings | Compare AI, human and hybrid on the same basis | Revisit Asset 2 with real numbers |
| Opt-out and complaint rate | Unsubscribes and complaints divided by delivered | Early warning for brand and deliverability damage | Cut volume, retarget, review copy |
Pre-launch QA checklist
- The ICP is one paragraph and the offer is one sentence, and both have produced at least some positive replies manually.
- Asset 2 is filled with real quotes, including management and reply-handling time.
- A named owner reviews the system weekly, and a named person owns replies with a response-time target.
- The pilot list was checked by hand and the cost of a full batch was estimated from it.
- Every first touch uses only verified facts; no statistics, results or client names that you cannot prove.
- Sending domains are authenticated, separate from your main domain, and every email has an opt-out.
- Any reply stops the sequence, and opt-outs are suppressed across every channel.
- Interested, question and unclear replies always reach a human.
- Wrong-person replies are added to a blocklist.
- Data protection rules for your location and your prospects' have been checked.
Your first week, day by day
- Day 1: Write the ICP paragraph and offer sentence. Use Asset 1 to split the SDR work you do or plan.
- Day 2: Fill in Asset 2 with real quotes for all three options. Apply Asset 3 and pick human, AI or hybrid.
- Day 3: Name the system owner and the reply owner. Write the handoff rules from Asset 4 into a one-page doc.
- Day 4: Build and hand-check a pilot list of 10 to 20 accounts. Estimate the cost of a full batch.
- Day 5: Generate first touches following Asset 5 and approve each one. Set up and authenticate the sending domain.
- Day 6: Send the pilot batch. Set up reply classification and routing using the categories in Asset 5.
- Day 7: Start the KPI table and book the day 14 review in your calendar.
Frequently asked questions
What is an AI SDR?
An AI SDR is a software agent that does sales development work: building and enriching prospect lists, researching accounts, writing personalised first messages, sending follow-ups on schedule, sorting replies and updating the CRM. It works best with a clear ideal customer profile, verified data and a person who owns the conversations it creates.
Should I use an AI SDR or hire a human SDR?
Use an AI SDR when your offer and ICP are proven and the bottleneck is volume of good first touches. Hire a human when deals are large, accounts are few or buyers expect calls. For most B2B service businesses a hybrid works best: AI handles research, sending and admin, and a person owns every reply.
Can an AI SDR fully replace a sales development rep?
It can replace much of the research, writing, sending and admin, but not the judgement in live conversations. Interested prospects, questions and objections should go to a person. Someone also has to own the system weekly, checking data quality, copy and deliverability.
Is an AI SDR cheaper than a human SDR?
Often, but not always, and it depends on your figures. Compare like with like over the same period, including data, tools, usage, setup, the owner's time and the cost of whoever handles replies. Then divide by qualified meetings from your own pilot rather than a vendor's case study.
How do I automate SDR work without hurting deliverability?
Use authenticated secondary domains, verify every email before sending, keep batches small, include an opt-out, stop the sequence on any reply and monitor your spam rate. Google's sender guidelines set requirements such as SPF or DKIM authentication and a reported spam rate below 0.3% for senders to Gmail accounts.
What does a hybrid AI and human SDR model look like?
AI builds and enriches the list, researches accounts, drafts first touches, sends follow-ups, classifies replies and updates the CRM. A person approves early messages, answers every interested or unclear reply within hours, handles objections and books meetings. Each side does what it is best at.
Sources and further reading
- Anthropic, "Building effective agents" (start simple, human checkpoints, guardrails, testing before production). Accessed 13 November 2026.
- Google, "Email sender guidelines" (authentication, spam rate below 0.3%, one-click unsubscribe for senders of more than 5,000 messages a day). Accessed 13 November 2026.
- Google, "Email sender guidelines FAQ". Accessed 13 November 2026.
- UAE Government portal, "Data protection laws" (Personal Data Protection Law and consent). Accessed 13 November 2026.
- SFC AI Stack page (tool descriptions; some links are referral links).
- SFC AI Outbound Agents service page (in-house owner model and weekly time commitment).
- SFC internal prompt library, enrichment SOP and outreach reply-handling guide (methods adapted and anonymised).
What to do next
Do Day 1 today: run the task audit and fill in the cost worksheet with real quotes. If the answer is AI or hybrid and you would rather have it built in your own accounts, see how we build AI outbound agents. If the answer is a person, we can place a trained appointment setter. Either way, you can book a short call to talk it through. For the booking side once replies arrive, read our AI appointment setter guide.