An AI SDR reply rate is the easiest number in outbound to quote and the hardest to compare, because almost nobody publishing one says what sits underneath the division line. Two prominent accounts from the last few weeks make the problem concrete, and one of them contradicts itself.
Start with the bullish case, because it is the better documented of the two. Jason Lemkin, writing at SaaStr on six months of running AI sales development agents, reports an outbound agent that sent 19,847 messages over the period at a 6.7 percent overall response rate and a 4 percent positive response rate, against the roughly 3 percent the same write-up cites as an industry average. Monthly volume ran near 3,000 messages, where the human reps it replaced had managed 75 to 285 each.
That is a real result from a named operator with the figures attached, and it deserves to be taken seriously rather than waved off. It is also not the claim most people repeat from it. The repeated claim is the multiple, the ten times more email. Lemkin's own assessment of the rate is considerably more modest: he writes that response rates have not changed materially in six months, and that the agents are not materially better than the humans on that measure, just basically the same.
Read those two statements together and the actual finding appears. The agents did not improve the rate. They multiplied the volume while holding the rate roughly flat, which is a volume win, not a quality win. Whether a volume win is a win at all depends entirely on what the rate is a rate of, and that is where the published numbers stop cooperating.
What The AI SDR Reply Rate Numbers Actually Say
SaaStr publishes three campaign breakdowns with leads, messages, responses and a stated rate for each, and in all three the stated rate sits between replies per message sent and replies per lead contacted without equalling either one. The gap is not consistent across the three.
| REPORTED CAMPAIGN | LEADS | MESSAGES | RESPONSES | STATED RATE | REPLIES PER MESSAGE | REPLIES PER LEAD |
|---|---|---|---|---|---|---|
| Europa prior-year attendees | 800 | 2,175 | 137 | 12.13% | 6.3% | 17.1% |
| London known website visitors | 265 | 735 | 33 | 11.70% | 4.5% | 12.5% |
| Europa earlier attendees, batches 1 and 2 | about 880 | about 2,490 | 78 | 5.54% to 5.66% | 3.1% | 8.9% |
The two right-hand columns are arithmetic on the figures as published, not a competing measurement. Work the first row and the issue is plain. One hundred and thirty-seven responses from 2,175 messages is 6.3 percent. The same responses against 800 leads is 17.1 percent. The stated 12.13 percent is neither, and no definition in the write-up reconciles it. The London row lands close to its lead denominator, and the third row sits in the middle of its two. Three campaigns, three different relationships between the published rate and the published inputs.
None of that makes the program a failure or the author careless. Sequencers report several reply metrics side by side, deduplication across a multi-message sequence is genuinely fiddly, and an operator writing up his own results is not running an audit. It does mean the headline rate cannot be compared to your rate, or to a vendor benchmark, or to the roughly 3 percent industry figure quoted next to it, because those are very unlikely to share a denominator. We made the wider version of this argument when we traced how outbound benchmarks get laundered through citation, and this is the same failure at the level of a single honest write-up.
SaaStr's reported Europa campaign (800 leads, 2,175 messages, 137 responses, stated rate 12.13 percent), with each candidate denominator shown against the highest reading. Bar length is indexed to that highest reading, so it is not itself a percentage.
The spread between the top and bottom bar is the whole commercial question. At 17.1 percent an outbound program is excellent and should be funded harder. At 6.3 percent it is ordinary and the sequence length is doing a lot of the work. Same campaign, same responses, same week, and the number a buyer would act on moves by a factor of nearly three depending on a definition nobody wrote down.
Why The Denominator Decides The Whole Argument
Replies per message sent and replies per account reached move in opposite directions when volume rises on a fixed list, so a program can hold one steady while the other degrades, and only one of the two is connected to pipeline. Choosing between them is a strategy decision, not a reporting preference.
Here is the mechanism. Your total addressable account list is finite and it does not grow because you bought a sending tool. Adding volume on top of a fixed list means more messages per account: longer sequences, more personas per company, more follow-ups before you give up. Each extra touch is worth less than the one before it, so replies per message sent falls by construction. Replies per account reached, meanwhile, keeps climbing, because you are extracting more response from the same universe.
Both of those are real and neither is a trick. The problem is that they justify opposite decisions. A falling rate per message reads as deteriorating creative and invites a rewrite. A rising rate per account reads as a working channel and invites more budget. A team watching only the first metric cuts a program that is growing pipeline. A team watching only the second adds volume right up until the sending infrastructure gives out, which is a separate ceiling with no warning light on this dashboard at all.
Lemkin is specific about what the agents are and are not, and the sentence that travels furthest is the one about amplification.
“AI SDRs scale what's already working. They can't create something from nothing.”
Take that literally and it is a constraint, not a slogan. Scaling what works also scales what does not, and the thing most reliably scaled by a tenfold volume increase is your complaint rate. We have argued before that sending volume is the deliverability variable teams mismanage first, and an agent that writes 3,000 messages a month on infrastructure sized for 300 is the fastest route to finding out. The reply rate denominator and the reputation ceiling are the same conversation, which is why the volume case has to be argued on both at once.
Lemkin's other framing is worth keeping too, because it sets the honest expectation for autonomy. He writes that the agents do not fail catastrophically without him, they just perform at a B-plus level instead of an A-plus level, on 15 to 20 hours a week of operator attention across five of them. That is a managed system with a person in it, not a replacement for one, and it is roughly what the stop rules an outbound agent needs are designed to formalise.
What Atlassian's Own Reps Did Instead
Atlassian's Teamwork Lab interviewed eight sales development representatives across three regions in August and September 2026 and found a hard internal boundary: no AI-written communication or artifact reaches a customer or an account executive without human review first. The AI is used on preparation, not on sending.
Beth Klawitter, Atlassian's global head of enterprise development representatives, published the findings on November 9, 2026, with contributions from regional manager Toni Cadden and Lora Kassanchuk, who leads its mid-market reps. The detail that matters most for anyone reading reply rate claims is a single rep's experiment: they stopped using AI for outbound email altogether, wrote in a casual register closer to texting a friend, and booked ten meetings in a week.
One rep for one week is an anecdote and should be labelled as one. It is a useful anecdote because it is the opposite experiment from SaaStr's, run inside a large enterprise, and reported by the people who manage the function rather than by anyone selling software. Where AI clearly did help was upstream of the message: one rep went from five or six calls a week to twenty once preparation was automated. Patrick Konadu, a rep quoted in the write-up, put the limit at the thing a tool cannot do, the ability to make the customer laugh.
Notice that Atlassian publishes no reply rate at all, and that this is a defensible choice rather than an omission. A function that reviews every outbound message by hand has a volume ceiling set by reviewer hours, so its binding constraint is review capacity, and replies per message is the number it should optimise. SaaStr's constraint is list exhaustion and deliverability, so replies per account is its number. The two organisations are not disagreeing about AI. They are operating under different binding constraints and reporting the metric that matches.
How To Read An AI SDR Reply Rate Before You Buy One
Any reply rate quoted to you by a vendor, a case study or a peer is uninterpretable until you know four things: the denominator, the reply definition, the deduplication rule, and the list quality underneath it. Ask for all four before the number enters a business case.
| DENOMINATOR | WHAT IT MEASURES | WHAT ADDED VOLUME DOES TO IT | WHEN IT IS THE RIGHT NUMBER |
|---|---|---|---|
| Replies per message sent | Efficiency of each individual send | Falls, by construction, on a fixed list | Judging creative, targeting and sequence length |
| Replies per account reached | Coverage of your addressable list | Rises until the list is exhausted | Judging whether the channel is still growing pipeline |
| Positive replies per account reached | Quality of the demand created | Roughly flat, it tracks fit rather than effort | The only one tied to revenue, and the one to commit to |
| Replies per sending domain per day | Load on your infrastructure | Rises toward the reputation ceiling | Judging whether more volume is safe to add at all |
| Negative and unqualified replies per 1,000 sends | Cost of the volume you added | Rises faster than positive replies do | Judging when to stop adding volume |
The third row is the one to commit to, and it is deliberately the least flattering. Positive replies per account reached cannot be improved by sending more, which is exactly why it is the honest measure of an outbound program. A tool that triples your sends will move every other row in this table and leave that one roughly where it was, and a vendor conversation that avoids it is telling you something.
Pick The Denominator Before You Pick The Volume
The decision in front of most outbound teams right now is not whether to use AI agents, which is close to settled, but which number they are going to be held to once they do, and that choice has to be made before volume rises rather than after.
Make it afterwards and the data no longer supports the argument. Once sends have tripled, replies per message has fallen for reasons that have nothing to do with the quality of the work, and nobody in the room can separate the mechanical dilution from a genuine decline in performance. The baseline you needed was the one from before the volume change, computed on a denominator you had already agreed on. That is a fifteen-minute job this week and an unanswerable question next quarter.
The two accounts in front of us are not actually in conflict, and reading them as a debate is the mistake. SaaStr ran a volume play on a list it was exhausting, measured it against messages, and got a defensible result it slightly over-claimed in the headline. Atlassian ran a quality play against a reviewer-hours ceiling, measured nothing publicly, and got a defensible result too. The failure mode is importing one organisation's metric into the other's constraint, which is what happens every time a reply rate crosses a slide boundary without its denominator attached.
For B2B programs the practical consequence is narrower than the discourse suggests. Volume is worth buying when your addressable list is genuinely larger than your current coverage, your infrastructure has headroom measured per domain rather than in aggregate, and your positive reply definition is strict enough that the extra sends cannot flatter it. Fail any of those three and added volume is a cost dressed as growth, and it will show up first in your deliverability and only later in your pipeline.
See where you are cited today
A free snapshot audit of your rankings and AI citations before we ever talk.
Tyler leads work at the intersection of SEO and generative engines at Something Inc., helping B2B brands get ranked and cited across every major AI engine.