The pitch two years ago was that you would not need SDRs. An agent would find the accounts, write the messages, handle the replies and book the meetings, and the whole function would collapse into a subscription.
That is not what happened. What happened is that a lot of teams bought the promise, ran it for two quarters, watched reply rates fall and domain reputation with them, and quietly went back to humans sending fewer, better messages with software doing the research underneath.
The interesting question is no longer whether AI belongs in outbound. It obviously does. The question is where the line sits, and the answer that has emerged from two years of expensive experimentation is more specific and less exciting than the original pitch.
This is a practical piece for small teams deciding what to buy and how to work. It covers what the autonomous model got wrong, which parts of the job genuinely automate well, which parts do not, and how to evaluate a tool without running a six-month experiment on your own domain reputation.
What actually went wrong with fully autonomous outbound
The failure was not that the writing was bad. By 2025 the drafting was competent. Messages were grammatical, contextually plausible and often better constructed than what a junior rep produced under quota pressure.
The failure was that competent writing at high volume made the underlying problem worse rather than better.
Outbound had already been sliding towards a volume equilibrium where response rates were falling because everyone was sending more. Vendor benchmark reports across 2026 put average B2B cold email reply rates in the low single digits, with the same reports consistently finding that smaller, tightly targeted sends outperform large ones by a wide margin. These are self-published datasets and should be treated as directional rather than definitive, but the pattern appears in every provider's numbers.
Autonomous tools removed the last constraint on volume. When sending a well-researched message cost a rep twenty minutes, volume was naturally capped. When it cost nothing, the cap disappeared, and a large number of teams responded exactly as you would predict.
The consequences were mechanical. Sending reputation degraded because more mail went to more people who had less reason to want it. Reply rates fell for everyone including the teams that had not changed anything. And the specific advantage of a thoughtful message, which is that it stands out against generic ones, eroded as generic messages became more polished.
There is a second failure that gets less attention, and it is about accountability. An autonomous system that sends a message and receives a reply has to decide what to do next, and that decision requires knowing things the system does not have access to. Whether this prospect is already in a conversation with your co-founder. Whether they are a former customer who left unhappily. Whether the tone of their reply is polite dismissal or genuine curiosity. Getting those wrong is not a minor quality issue. It burns the account.
Industry reporting through 2026 has described high churn among fully autonomous deployments, with hybrid configurations, where AI handles research and drafting while people own judgement and replies, becoming the dominant pattern. The figures published on this vary a lot between sources and most come from vendors with a position, so the specific percentages are not worth quoting with confidence. The direction of travel is not seriously disputed.
The split that works
The useful way to think about this is to separate the outbound job into its actual components and ask which ones have a clear right answer.
Work with a clear right answer automates well. Work requiring judgement about a specific human does not. That single distinction resolves most of the decisions.
Finding accounts that match a definition has a right answer. You describe the profile and the system returns companies. This is search, and there is no reason for a person to do it.
Monitoring for signals has a right answer. Whether a company announced funding, posted four sales roles or added a technology to their stack are facts. A person checking these manually every morning is doing work that should not exist.
Enriching a company into contactable people has a right answer. Who holds which role, and how to reach them, are facts with a correct value, even if that value is sometimes hard to determine.
Verifying that data is current has a right answer, and it is the least glamorous and most valuable automation available. It is also where a lot of teams underinvest while buying more interesting things.
Producing a first draft is where it gets interesting. There is no single right answer, but there is a large gap between a blank page and a reasonable starting point, and software crosses that gap quickly. A draft that establishes the structure and references the signal correctly saves real time even when a person rewrites half of it.
Deciding whether this message should go to this person has no right answer and requires context the system does not have. This is judgement.
Reading a reply and deciding what it means is judgement, and it is the highest-stakes moment in the entire sequence. A misread reply costs the account.
Handling an objection is judgement, and it is also the part of the job that develops the person doing it.
The pattern is clean. Automate everything up to and including the draft. Keep everything from the send decision onwards with a person. That is not a compromise position, it is where the value actually is, because the automated portion is the part that consumes hours without requiring skill, and the human portion is the part where skill changes the outcome.
What this looks like in practice for a small team
Take a two-person sales function at a company of thirty. Before, both reps spent most of a working day each week on list building, research and data cleanup, and the rest sending messages that were shallower than they wanted because the research time was gone.
The version that works now removes the list building and research entirely. Signals are monitored automatically. Accounts matching the profile arrive already enriched. A draft exists for each one, referencing the specific trigger.
What the reps do is read the draft, decide whether this particular company and person warrant contact this week, rewrite the parts that only a person would know to rewrite, and own every reply.
The volume goes down. That is the point and it needs to be understood before starting, because week two will look like a productivity decline if you are measuring activity. Thirty considered messages a week per rep, sent to accounts where something is genuinely happening, will beat four hundred automated ones on meetings booked, and it will do so while keeping the domain healthy enough to still be working in six months.
Empiraa Signal is built to this shape, with Prospect Spark doing the finding and enrichment and ANI producing drafts inside the sequence, while the send decision and the reply stay with the rep. The specific tool is less important than the boundary. Any stack that respects the same line will outperform one that does not.
Evaluating a tool without wrecking your domain
The most expensive mistake in this category is running the evaluation on your primary sending domain at full volume. If it goes badly, the recovery takes longer than the trial did.
A few things make the evaluation safer and faster.
Run it on a subdomain, not your main one. Any serious outbound operation should be doing this regardless, but during an evaluation it is essential, because it contains the damage if the tool sends more or worse than expected.
Cap the volume manually rather than trusting a setting. Decide the number of sends for the trial before you start and enforce it, because the natural behaviour of these tools is to fill available capacity.
Judge output quality on the drafts, not the results. Read fifty drafts before any of them send. If you would not send forty of the fifty without substantial rewriting, the tool is not saving you the time it claims to, and no amount of tuning will close a gap that large.
Test the research separately from the writing. Many tools are strong at one and weak at the other. If the enrichment is accurate and the drafting is poor, that is a usable tool with a workflow around it. If the enrichment is wrong, nothing downstream can be trusted, because a beautifully written message built on a wrong fact is worse than no message.
Check what happens on a reply. Ask directly what the system does when someone responds, and be sceptical of anything that classifies and responds without a person. This is the boundary that matters most and it is the one most often blurred in demonstrations.
Measure on meetings per hundred sent rather than on meetings in total. Total meetings can be increased by sending more, which is exactly the behaviour you are trying to avoid. The efficiency ratio is the honest number.
Doing the cost comparison honestly
The commercial case for these tools is usually presented as a comparison against the fully loaded cost of an SDR, and that comparison is misleading in both directions.
It overstates the case because a tool does not replace an SDR. It replaces a portion of an SDR's week, specifically the research and list-building portion, which is large but is not the whole job. If you cancel the headcount and buy the tool, you have not swapped one for the other. You have removed the judgement layer and kept the preparation layer, which is precisely the wrong half to keep.
It understates the case in a different way, because the relevant comparison for most small teams is not against a hire they were going to make. It is against the hours their existing people are currently burning on preparation. A founder doing their own prospecting is spending time that has a much higher opportunity cost than an SDR salary implies, and for them the tool is often worth it even at fairly poor output quality, simply because it converts an unpleasant multi-hour task into a review task.
The honest way to run the numbers is to measure the preparation hours first. For two weeks, have whoever does outbound log how long they spend on finding accounts, researching them and assembling lists, separately from writing and sending. Most teams are surprised by the total. That figure, multiplied by what those hours are worth, is the actual value the tool is competing against.
Then discount it, because the review step is not free. Reading and editing fifty drafts takes real time, and any calculation assuming the drafts go out untouched is calculating the autonomous model that does not work.
What tends to fall out of this exercise is that the tools are worth buying for most teams, but for a smaller saving than the marketing claims and with the saving landing in a different place than expected. The gain is rarely more messages sent. It is the same number of messages sent with substantially more thought behind each one, because the research time got returned to the part that needed it.
Judging draft quality when everything sounds fine
Evaluating output is harder than it looks, because modern drafts are uniformly fluent. Fluency stopped being a signal some time ago, and teams that assess on how the message reads tend to approve things that will not work.
A few specific tests separate a usable draft from a plausible one.
Check whether the message contains a fact that could only be known about this company. Not a fact retrieved about this company, which most tools manage, but one that required combining two things. A message referencing a funding round is retrieval. A message referencing what a funding round of that size usually means for a team of that shape is closer to thinking, and it is far rarer.
Check whether the message would still make sense addressed to a competitor of the recipient. If it would, the personalisation is decorative. This is the fastest test available and it fails a surprising proportion of output.
Check whether the ask is proportionate to the relationship. A cold first message requesting a forty-five minute discovery call is asking a stranger for most of an hour. Tools default to this constantly because it is what the training data is full of.
Check the opening line against the obvious patterns. Anything starting with a compliment about the company's growth, or a rhetorical question about a problem the reader has not stated they have, reads as automated to anyone who receives outbound regularly, which is exactly the audience you are writing to.
Run these four tests across fifty drafts and you will have a clearer view of a tool than any trial dashboard will give you.
The part nobody wants to say about junior reps
There is an uncomfortable second-order effect worth naming.
The work that automates most cleanly, research, list building, first drafts, is also the work that junior salespeople traditionally learned on. Writing two hundred bad emails and finding out which ones failed is how people develop a feel for what lands. Removing that step removes the training.
Teams that have automated the top of the funnel and kept the judgement layer human sometimes find, a year later, that they have nobody coming through capable of exercising that judgement, because the apprenticeship was the part that got optimised away.
There is no clean answer. The partial one is to be deliberate about it: have junior reps write from scratch some of the time, even when a draft exists, and review those drafts against the automated ones. It is deliberately inefficient, and it is the cost of having people who can still do the part that has not been automated.
This matters more for small teams than large ones, because a small team cannot hire its way around a capability gap.
Where this is heading
The reasonable expectation is that the automated portion keeps expanding into work with clear right answers, and that the judgement boundary moves more slowly than vendors suggest.
Research and enrichment will keep getting better and cheaper, to the point of being an assumed feature rather than a product. Drafting will keep improving and will matter less, because when everyone's drafts are good, the draft stops being the differentiator and the judgement about who to contact becomes the whole advantage.
That is the useful thing to invest in now. Not the tooling, which will be commoditised, but the discipline of contacting fewer people for better reasons, and having someone capable of telling the difference.
Frequently asked questions
What is an AI SDR?
An AI SDR is software that performs some or all of the sales development role, typically finding target accounts, researching them, writing outbound messages and in some cases handling replies. The category launched around the promise of full autonomy, meaning no human in the loop. In practice most teams now run a partial version, using the software for research and drafting while keeping a person responsible for the send decision and every reply.
Do AI SDRs actually work?
The fully autonomous version has broadly underdelivered, with high churn reported across the industry through 2026. The version that works is narrower: automating account discovery, signal monitoring, data enrichment and first-draft writing, while people retain judgement over who gets contacted and how replies are handled. The distinction matters because the failures were not caused by poor writing but by removing the volume constraint and the accountability for replies.
What should a small B2B team automate in outbound?
Anything with a clear right answer. Finding companies that match a defined profile, monitoring for events such as funding or hiring, enriching companies into contactable people, verifying that contact data is still current, and producing a first draft. Keep the decision to send, the reading of replies and objection handling with a person, because those require context about the specific relationship that the software does not have.
Will AI replace sales development representatives?
It has substantially replaced the research and list-building portion of the role, which was often the majority of the hours. It has not replaced the judgement portion. The practical effect for most small teams has been fewer people spending far less time on preparation and more on conversations, rather than the function disappearing. The open risk is that removing the preparation work also removes how junior reps traditionally learned the judgement part.
How do I trial an AI SDR tool safely?
Use a separate sending subdomain so any reputation damage is contained, cap the send volume manually rather than relying on a setting, and read at least fifty drafts before allowing anything to send. Evaluate the enrichment accuracy separately from the writing quality, since a well-written message built on incorrect data is worse than sending nothing. Judge the trial on meetings booked per hundred messages sent rather than total meetings, because total volume can be inflated by simply sending more.
Does using AI in outbound hurt deliverability?
The tools themselves do not, but the way they are typically used often does. Removing the time cost of producing a message removes the natural cap on volume, and higher volume to less relevant recipients produces more spam complaints and more bounces, which degrades sending reputation. Teams that keep volume deliberately low and relevance high generally see deliverability improve rather than decline, because the underlying quality of the send list gets better.
