AI SDR
It drafts every reply. You approve what sends
Outreach creates replies, and replies are where most campaigns quietly die, because they arrive while you are delivering for a client and get answered three days later or not at all. This reads every one the moment it lands, sorts it, and hands you a written response. You approve, edit or skip. Nothing goes out in your name that you have not seen.
What does an AI SDR actually do here?
It does the part of the job that is reading, sorting and typing. Every reply that comes back gets read straight away, given one label, and attached to a drafted response written in your voice and against your offer.
What it does not do is decide on its own that a human being is ready to be pushed, told a price, or given a commitment. That judgment stays with a person, and it is deliberately not an efficiency we go looking for.
Why is a human still in this?
Because the fully autonomous version has been tried, at scale, with far more money behind it than this, and it did not hold. The companies selling an AI that runs sales conversations end to end have been walking that promise back rather than extending it.
The reason is not that the models are bad. It is that they are excellent at research, volume and a first draft, and unreliable at the exact moment someone says something that was not anticipated. That moment is not a rare edge case. It is where deals are won and lost, and it is most of what an interested buyer actually says.
What can it send without asking you?
At the beginning, nothing. Everything is drafted and waits. Categories graduate to automatic one at a time, starting with the safest one, and only once it has been right about that category on your own replies rather than on a demo.
There is one exception in the other direction. A request to be removed is suppressed immediately, with no reply sent and no approval step, because that is a legal obligation rather than a judgment call and a delay there is the one mistake with a real cost attached.
How is this different from an agent deployment?
Scope. Agent deployment gives you a general purpose agent on its own machine that holds the repeatable parts of your whole week, from morning reports to research to admin. This is one job, done properly: the inbound half of outreach.
If outreach is the only thing you want handled, this is the smaller and cheaper answer. If the reply layer is one of six things quietly not getting done, the agent is the better shape. We will tell you which one your week actually points at.
How is this different from hiring an SDR?
A sales development rep is a salary, a ramp period, a manager, and somebody who works their own hours in one timezone. Replies that land at ten at night or on a Saturday wait until Monday.
This costs a fraction of the hire, reads everything the moment it arrives, and never has a slow week. What it does not have is the ability to sit on a call and read a room, which is exactly why a person stays in the loop instead of being removed from it.
Every reply gets exactly one label, and the label decides what happens
- 1Interested. A warm response with the next step attached, drafted and waiting for your approval.
- 2Asking a question. A drafted answer plus a light nudge toward a call, for you to approve.
- 3An objection. Always a human. It gets a drafted reframe and you decide whether to send it, because this is the reply most likely to be answered badly at speed.
- 4A referral to somebody else. A drafted thank you and a request for the introduction, which is the highest value reply in the whole list and the one most often left sitting.
- 5A clear no. A gracious close, and they come off the list rather than getting caught by a follow-up two weeks later.
- 6Remove me. Suppressed immediately, no reply, no approval step. That one is not optional.
- 7An out of office or an autoresponder. Ignored, not counted as a reply, and the follow-up gets moved past the date they gave.
- 8Anything it is not confident about. Routed to you no matter which label it chose. This rule catches more real problems than the labels do.
Three versions of this exist, and two of them do not work
| The version | What goes wrong | What we run |
|---|---|---|
| Fully autonomous | It sends confidently wrong things to real buyers in your name, and you find out from the buyer. A message sent in your name cannot be edited afterwards. | AI drafts and sorts. A human approves anything that carries a judgment. |
| All manual, by you | Replies sit for hours or days, because the person meant to answer them is the same person delivering the work. The interested ones go cold while you are busy being busy. | Every reply is read and drafted within minutes, so your job becomes approving rather than composing. |
| A generic AI inbox assistant | It writes in nobody's voice, knows nothing specific about your offer, and treats an objection exactly the same as a simple question. | Built on your offer and your voice, with objections and referrals routed to you by rule rather than by luck. |
| Nothing, and hope the good ones follow up | The best prospects are the busiest, so they are the least likely to chase you twice. Silence reads as disinterest and they go elsewhere. | Nothing that replied goes unanswered, and nothing that said not yet gets forgotten. |
Questions people ask
Does it send emails as me without me seeing them?
No. At the start it sends nothing at all. Every response is drafted and waits for you, and approving one takes a couple of seconds on your phone. Categories graduate to sending automatically one at a time, starting with the safest, and only after we have both watched it get that category right on your actual replies.
How does it know what to say about my offer?
It is built on what you already have: your offer, your pricing rules, what you will and will not promise. When a reply needs a number or a commitment that is not written down anywhere, it flags it for you instead of inventing one. That rule exists because inventing one is the default failure of every model ever built.
What happens to the replies that are not really replies?
Out of office messages, autoresponders, receipts and bounces get labeled as what they are and are not counted as replies. That matters more than it sounds. Counting them is how a campaign looks like it is working for a month while nothing is actually happening.
Is this replacing a salesperson?
No, and it is not trying to. It removes the reading, the sorting, the typing and the chasing. It does not take the call and it does not close. If you want the calls taken as well, that is a real operator's time and it is a different conversation.
Can it work with the CRM I already use?
Usually. Replies route into wherever you already track deals, so you are not keeping a second list by hand. If your setup will not connect cleanly, we tell you that before you sign anything rather than after.
What happens when it gets one wrong?
You see it before it sends, which is the whole purpose of the approval step. On top of that, anything the classifier is not confident about routes to you regardless of which label it picked. The failure we design against is not a bad draft. It is a confident bad draft that nobody looked at.
Stop losing the replies you already earned.
Bring the last twenty replies you got from any campaign. We will show you how each one would have been labeled, and what the draft would have said.
20 minutes. Straight answer, no pitch.