AI SDR
Do AI SDRs actually work? A 30 day way to find out
By Viktorio Shopov, co-founder of Sellinger. Reviewed by Asen Levov. Published . Last updated . 9 minute read.
AI SDRs can work, but you cannot tell from a demo or a week. Test one audience and one offer for 30 days. Count people messaged, replies, positive replies, meetings booked and meetings held as separate numbers, and decide by cost per held meeting. A reply rate alone proves very little.
What do we mean by "work"?
An AI SDR works if it produces conversations that a person wants to have, at a cost you accept, with messages you would sign. That is four conditions, and a reply rate checks none of them. On our own LinkedIn account, 40% of people messaged replied, and that rate held between 35% and 45% in each of twelve monthly cohorts. Meetings per person messaged peaked at 2.4% in November 2025 and drifted below 1% from June 2026 (Sellinger case study). Replies stayed flat while meetings fell. So a test has to look past replies.
This is the plan we would run on a new account. The thresholds are our suggestions, not industry benchmarks, and we say so where it matters.
What are the honest definitions?
Agree these before day one and write them down.
| Term | Definition | Common way it is faked |
|---|---|---|
| People messaged | Unique people who received a first message | Counting messages instead of people |
| Reply | Any human reply | Counting automatic replies, out of office notes and helpdesk tickets |
| Positive reply | A reply that moves toward a next step: interest, a question about fit or price, a request for a time | Counting polite refusals |
| Meeting booked | A time was agreed or picked through a link | Counting it as a meeting |
| Meeting held | A prospect and a host were both there, with evidence: a recording, a transcript or a host mark | Trusting a calendar's completed status |
| Cost per held meeting | Total cost in the period divided by held meetings | Dividing by bookings |
A calendar marks a meeting completed when its slot ends, whether or not anyone joined, so "completed" is not "held". See Count meetings honestly. In our own data, 351 showcase meetings were booked through the agents and 33 are confirmed held. The rest are unconfirmed, and the honest way to report that is to say so.
How big must the sample be?
Small samples fool people in both directions. Two rules of thumb:
- Below about 100 people messaged, a result says almost nothing. One extra meeting changes the rate by a full point.
- Zero is not proof. A standard statistical rule of three says that if you see zero events in n tries, the true rate could still be as high as about 3 divided by n. Message 300 people, book nothing, and the true rate could still be close to 1%. Message 1,000 and book nothing, and it is likely under 0.3%.
For a meeting rate near 1%, which is in the range of our own data, you need roughly 300 to 500 people messaged in one audience before the number means something. That is why the plan below starts with one audience. If you split 300 people across five audiences, you have 60 per audience and no signal.
Also give each audience time. Our median wait from first message to the prospect's first reply was 10 hours, 80% of conversions came from a reply to the first message, and in one study a solo founder's median from first touch to booking was 22 days, with 11 of 19 bookings more than two weeks after the first touch (ShortVibe case study). Some bookings from week 4 will land in week 6.
What do you measure week by week?
Week 0: set up before you send
- Pick one audience and write who they are in one paragraph.
- Write one offer: one claim, one proof point, one next step.
- Spot check 30 rows of data by hand.
- Set the definitions from the table above and a place to log them.
- Name the person who takes over warm conversations, and agree a response time. One practitioner (SaaStr) puts the sweet spot for human follow up at within 2 hours of the agent's engagement, which is his view, not a study. Our hotel study showed staff taking a median of almost 4 hours on group requests while the agent replied in minutes.
- Start in approval mode so every message waits for a person. A campaign can run in approval mode or send automatically.
Week 1: is the quality right?
This week is about messages, not results. You will have too little data for rates.
- Read every message the agent drafts, 50 to 100.
- Check claims. Does it say anything you cannot back up?
- Check fit. Are the replies from people who are in your audience?
- Check opt outs. When someone says stop, does the agent stop?
- Fix the instructions after each batch, and log what you changed.
Pass: you would send these messages under your own name. If not, do not scale. Fix the offer and the rules.
Week 2: is it reaching people?
Now volume ramps within the account's limits. On LinkedIn, acceptance is the early health signal: across 42 customer teams in the 90 days to 19 September 2026, about 4 in 10 invitations were accepted (39%), and on our own account 41% of 68,818 requests were accepted. Treat a result far below that as a sign to look at the audience, not to push volume. Check:
- People messaged so far (target a few hundred in the audience by the end of the week, within the pacing limits).
- Acceptance rate on LinkedIn, or delivery and bounce rate on email.
- Reply rate, and the share of replies that are real, not automatic.
- Time from first message to first reply. A quiet first day says little.
Week 3: is it converting?
- Positive replies as a share of people messaged.
- Meetings booked as a share of people messaged, and of positive replies.
- Hand over speed: how long from "interested" to a human response.
- Where threads die: after the first reply, after the booking link, after a question the agent could not answer. On our own account the booking link typically went out after three prospect messages, and more than half of the bookings confirmed in the chat needed one more nudge after the link.
Week 4: the decision week
- Meetings held versus booked. Confirm each with evidence.
- Cost per held meeting. Total cost (software, the person's hours, any data) divided by held meetings. The method is in the ROI post.
- Quality of meetings: did the people fit, and did the call lead to a next step?
- A written verdict using the table below.
How do you decide at day 30?
| What you see | Likely cause | What to do |
|---|---|---|
| Few acceptances or replies, enough people messaged | Audience or opening message | Change the audience first, then the opener |
| Good replies, few positive replies | Offer is thin or off target | Rewrite the offer and the proof point |
| Positive replies, few bookings | Booking step or hand over | Check how fast a person answers and how the link is offered |
| Bookings, few held | Show rate or the hand off | Add a reminder and a named host. Count held, not booked |
| Held meetings, no next steps | Wrong people, or the sales call | Tighten the ICP. Review the call, not the agent |
| Fewer than 100 people messaged | No signal | Extend the test. Do not decide |
| 300 or more people messaged, no conversations worth a call, after one audience or offer change | The channel or the idea | Stop, or try a different audience, with an honest write up of why |
Our own change from replies to meetings shows why you change one thing at a time. The same agent booked 2.6% to 3.7% meetings per person messaged for one audience and under 0.4% for another. If you change the audience, the offer and the copy at once, you will not know which one mattered.
What should you stop doing?
- Stop reading the reply rate alone. On our account it stayed near 40% for twelve months.
- Stop comparing with a headline number from a vendor without a denominator. Our case studies show how we publish ours, with counts and definitions.
- Stop adding audiences before one has 300 people in it.
- Stop judging at day 7. Judge quality at day 7 and results at day 30.
One practitioner, writing about AI agents in general, argues for a 90 day discipline before judging results (SaaStr). That is one operator's view with his own results, but the instinct is sound: 30 days tests the setup, and a full verdict on pipeline needs longer.
What does a good 30 day report look like?
One page, same layout every week:
- Audience and offer in one sentence each.
- People messaged, replies, positive replies, booked, held, with the counts and the percentages of people messaged.
- Cost for the period and cost per held meeting.
- What the agent got wrong, and what you changed.
- Verdict: continue, change one thing, or stop, and why.
If you can fill this in honestly, you know whether an AI SDR works for you. If you cannot, the problem is measurement, and the next step is the checklist of questions to ask of the vendor.
Common questions
Do AI SDRs actually work?
How many people do I need to message to know?
What counts as a positive reply?
Is 30 days long enough?
When should I stop?
How Sellinger helps
Read how it works on See how Sellinger books meetings, or see a real campaign in Read the Sellinger case study.
