AI SDRs work for part of the job. The strongest evidence available shows they save sales teams time on prospecting work. I found no independent study with a published method showing that they book more qualified meetings, or bring in more revenue, than people do.
The failure numbers are no better. "Only 2% of AI SDR implementations stick long-term", "50-70% churn within a year" and "human SDRs generated 2.6x more revenue" are repeated on page after page. On September 28, 2026 I followed each one to the page it is cited with. None of those pages contains the number.
So the honest answer to "do AI SDRs work" is this: they save time, the rest is unproven in both directions, and a measured pilot on your own list is the only data that settles it for your company. Below is every figure I checked, graded, with a link you can open yourself.
What counts as an AI SDR?
An AI SDR is software that does the work of a sales development rep. It builds prospect lists, looks each contact up, drafts the emails and sends them on a schedule. The more autonomous products also answer replies and put meetings on the calendar with nobody approving each step.
Some tools only draft and queue the work for a person. Others send on their own. Almost none of the statistics below say which kind they measured, and that is the first reason to read them carefully.
How I graded each number
I used four grades, one rule each:
- High: official documentation, or a study with a published method from someone who does not sell the product, and the link works.
- Medium: a named, dated source with a stated sample size, where the publisher sells in this category or the evidence covers a single company.
- Low: a vendor's own survey with no method, no selection details and no date.
- Untraceable: no source at all, or the cited page does not contain the number.
One figure can carry two grades. The same 45% is Medium when you cite the survey it comes from and Untraceable through the link I found attached to it.
Every AI SDR statistic I checked, graded
Checked September 28, 2026. Pages get edited, so a number may have sat on one of these pages once. A reader can only check what is there today.
| The claim, in short | Usually credited to | What I found | Grade |
|---|---|---|---|
| Only 2% of AI SDR implementations stick long-term | "Topo.io research", linked to a Topo blog post | Not in the linked post, dated October 23, 2025. No Topo report, sample or method found | Untraceable |
| 50-70% of AI SDR customers churn within a year | The same Topo link, or unnamed "2026 industry surveys" | Not in the linked post. No survey named anywhere I looked | Untraceable |
| Human SDRs bring in 2.6x more revenue than AI SDRs | A Dashly blog post | The post, dated June 27, 2025, contains neither the revenue figures nor the show rates attached to the claim | Untraceable |
| 45% of sales teams run a hybrid AI and human model | A UserGems blog post | No 45% on that page. The figure does appear in Outreach's survey of 500 sales professionals | Untraceable via UserGems, Medium via Outreach |
| 22% of teams now run SDR work on AI alone | Often no source, sometimes an AI vendor's home page | Appears in the same Outreach survey | Medium |
| Every AI user saves over an hour a week | Outreach's Prospecting 2025 report | In the report, self-reported by respondents | Medium |
| Most AI SDR pilots, 88%, stall before production | AiSDR's industry report | On the page, from "75+ teams" running AI SDRs; no method, selection or date | Low |
| Teams that churn do not move to another AI SDR | AiSDR's industry report | On the page, same undisclosed sample | Low |
| One AI SDR startup lost 70-80% of its customers | TechCrunch, March 24, 2025 | A former employee's claim about one company, which put its current retention at 79% | Medium, one company |
| Gmail bulk senders: authentication, one-click unsubscribe, spam rate under 0.3% | Google's sender guidelines | Stated on Google's own help page | High |
Ten rows and one High grade, and that row is about Gmail, not about AI SDR performance. Every row about AI SDRs themselves is Medium at best, and the Medium rows come from companies that sell sales software or from reporting on a single startup.
What the evidence covers, task by task
Put the same grades against the parts of the job an AI SDR is sold to do, and a pattern shows up quickly.
| Part of the job | What the evidence shows | Best grade available |
|---|---|---|
| Prospecting work: research, lists, first drafts | Users report saving hours every week | Medium, vendor survey, self-reported |
| Sending at volume | Google sets hard limits for bulk senders | High, platform rules |
| Getting from pilot to daily use | One vendor reports that most pilots stall | Low |
| Keeping customers | One startup's disputed churn, no number for the category | Medium for one company, nothing for the category |
| Replies, objections and qualification | No study found | None |
| Meetings held and revenue against human SDRs | The famous 2.6x figure is untraceable, no study found | None |
So the best evidence sits on the cheap part of the work. Nobody has published checkable data on the part of the job that decides whether a reply turns into a sales call.
Four famous numbers whose links lead nowhere
These figures travel on 2026 comparison pages about AI and human SDRs. The one I traced link by link is Salesmotion's comparison page, updated June 25, 2026, published by an account intelligence software company.
It links four headline figures to three sources. I opened all three.
"Only 2% stick" and "50-70% churn"
Both figures are credited to "Topo.io research" and linked to one Topo blog post that ranks AI SDR tools. Topo sells an AI SDR itself.
That post contains neither number. Its only statistic says 22% of sales teams now let AI do all of their SDR work, credited to another AI vendor and linked to that vendor's home page, not to a study.
Other pages repeat the pair under vaguer labels, like "aggregated 2026 industry surveys". I found no survey, no sample size and no date behind either figure. When the exact same "2%" shows up on unrelated sites with no study behind it, the likeliest explanation is that it is being copied from page to page.
"Human SDRs generated 2.6x more revenue"
This one is linked to Dashly's post comparing AI and human SDRs. The post does not contain the 2.6x figure, the revenue amounts behind it, or the meeting show rates quoted next to it. I am not repeating those amounts here, because repeating an untraceable number with a warning label still spreads it.
"45% run a hybrid model"
Linked to a UserGems post asking whether AI SDRs are worth it. There is no 45% on that page. Its numbers are about something else entirely: market size, cost per SDR and cost per lead, credited to other sources.
The 45% itself is real, though. It comes from Outreach's Prospecting 2025 report, published May 20, 2025. So the number survives and the citation does not, which is exactly why a figure and its link have to be checked separately.
What the evidence does support
Teams say AI saves them time
Outreach surveyed 500 sales professionals. Every respondent reported saving more than 1 hour a week with AI, and 38% of reps said it saves them 4-7 hours a week.
In the same survey, 45% of teams run a hybrid model, 22% have handed SDR work entirely to AI, and 23% use no AI.
Two caveats. Outreach sells sales engagement software, so this is a vendor-run survey of an audience that uses such tools. And those three groups add up to 90%, so read the report itself before quoting them as the whole market. The report also says every respondent saved more than an hour a week with AI, while 23% say they use no AI at all, so the survey wording is loose.
Also notice what was measured: hours saved, as reported by the people saving them. Nothing in that data tells you how many qualified meetings or how much revenue the saved hours produced.
Pilots stall, by a vendor's own account
AiSDR, which sells an AI SDR, published an industry report built on conversations with "75+ teams" running AI SDRs in production. It says 88% of pilots stall before production, and that teams who churned did not move to another AI SDR.
A vendor publishing failure rates about its own category is unusual, and that counts for something. But the page gives no method, no selection criteria and no date, so there is no way to check who those teams were.
One company's churn, reported by journalists
TechCrunch reported in March 2025 that a well-funded AI SDR startup had shown logos of companies that were not its customers. A former employee said the company was losing 70-80% of the customers who came through the door.
The company told TechCrunch its highest churn came from its first cohorts in late 2023 and that its retention was currently 79%. That is one company, with the two sides disputing the numbers. It tells you to ask every vendor for its own retention, and nothing about the category as a whole.
The rules no survey can change
Google's sender guidelines, checked September 28, 2026, class you as a bulk sender once your domain sends Gmail accounts over 5,000 messages in a day. From then on you need SPF, DKIM and DMARC, plus one-click unsubscribe on marketing mail.
The spam rate Postmaster Tools reports for you has to stay under 0.3%, and Google's preferred ceiling is 0.1%.
That is the one hard limit in this whole topic. An autonomous tool sending to a loose list at volume is how a team crosses it, and the damage lands on your own domain. The domain and mailbox setup that keeps you under it is in how I set up sending infrastructure for cold email.
How to grade any AI SDR number you see
You can run this in a few minutes on any stat in a vendor deck, a LinkedIn post or a proposal:
- Is there a named source and a date? "Industry surveys" and "recent research" are not sources.
- Does the link contain the number? Open it and search the page for the exact figure. All three links I traced failed this test.
- Is a sample size or method stated? "75+ teams" with no selection method is a sample size without a method.
- What was measured? Hours saved, meetings booked, meetings held and revenue are four different things. A time-saving stat cannot answer a revenue question.
- One company or the whole category? A startup's churn tells you about that startup.
- Who sells what? AI SDR vendors, sales software vendors and agencies that sell human-run outbound each have numbers they would like to be true.
Apply number 6 to me as well. Smirnov Consulting Group is a Prague-based B2B outbound lead generation agency that runs cold email and LinkedIn campaigns for founder-led B2B companies and books qualified sales calls. I have a side in this argument, which is why every number above links to the page it came from and none of them is mine.
What this means if you are deciding this quarter
Your prospects never see your CRM, your automation or the AI tool that wrote the email, and they do not care whose churn figure wins the argument online. They see one email from your domain, and either it is worth answering or it is not.
So I would skip the statistics war and use the evidence for what it can carry:
- Budget for the time savings. Time saved is the only benefit with a sample behind it. Treat any meeting or revenue lift a vendor quotes as a hypothesis for your pilot.
- Run the pilot as a test with two arms. Split one list from one target market in half. One half goes through the tool, the other through whatever you do today, and every contact gets the full sequence before you judge. If both arms sit far below published cold email reply rates, fix the list or the offer before you blame the tool.
- Count qualified, held meetings. Sends, opens and even booked meetings flatter a tool. Published prices for tools, hires and agencies, and a formula for the cost of each meeting, are in my cost-per-meeting comparison of all three.
- Set a stop rule before you start. If the spam rate in Postmaster Tools climbs toward 0.1%, pause the sending, whatever the reply numbers say.
- Keep a person on replies. A reply is where a lead becomes a conversation, and where I would not let a machine improvise. In practice that is a hybrid setup: software on the prospecting work, a person on everything the prospect answers.
- Ask the vendor two questions in writing: is its evidence from drafting mode or autonomous sending, and what is its own customer retention after 12 months?
If a vendor cannot answer the second question, the churn debate is already settled for that vendor.
Questions founders ask about AI SDR results
What is the real churn rate for AI SDR tools?
Nobody has published a checkable number for the category. The widely quoted "50-70% within a year" does not appear on the page it is usually linked to. The best documented case is one startup: a former employee alleged 70-80% customer loss, and the company said its retention was 79%.
Is there a reply rate benchmark for AI SDRs?
I found no AI-specific reply rate benchmark with a stated method. Judge the tool the way you would judge any cold email: replies per person contacted, compared with your current process on the same list. A tool that only matches your current rate has saved you time, not won you meetings.
Is there an independent study comparing AI and human SDRs?
None that I found has both a published method and no product to sell. The largest named survey, Outreach's 500 respondents, measures adoption and hours saved. It does not measure meetings or revenue. The rest are vendor reports, broken citations or single-company stories.
Are AI SDR vendors' own case studies reliable?
Treat them as sales material. Ask for the customer's name, the time period, the number of contacts and how many meetings actually took place. Then ask to speak to that customer.