It is the first question in most conversations I have, and it is the right one to ask. The category is full of software that finds prospects, writes to them, and books meetings while you sleep.
What I install does not send. It drafts, and then it waits for you.
That sounds like a small distinction. It is the whole design, and everything else follows from it.
What an AI SDR does
An AI SDR pulls contacts from a database, generates a message for each one, and sends on a schedule. You configure it once and review the results. The unit of work is messages delivered, and the arithmetic only closes at volume, because reply rates on generated mail are low enough that the model needs thousands of sends to produce a handful of conversations.
For a company with a large addressable market, a deliverability team, and domains it can afford to lose, that math works. A ten-person software company has none of those three, and it is the third one that ends badly.
Sending is the part worth keeping in human hands
Three reasons, in the order they will cost you.
The bill goes to your domain and your name
Software does not absorb the consequences of a bad message. Mailbox providers score complaints against the sending domain, which is yours, and the recipient forms an opinion about the company, which is also yours. We budget three to four weeks to warm a new sending domain to normal volume. Repairing one you have damaged takes longer and is less certain.
The people who run the inbox have written down what abuse looks like
Google requires senders above 5,000 messages a day to Gmail to authenticate with SPF, DKIM, and DMARC, to support one-click unsubscribe, and to keep the spam rate reported in Postmaster Tools below 0.30 percent. Microsoft applied its own version to Outlook.com, Hotmail, and Live addresses on May 5, 2025: mail from domains above 5,000 a day that fails SPF, DKIM, or DMARC is filtered to Junk, and Microsoft has said outright rejection follows.
Google, Email sender guidelines
You may never send five thousand messages in a day. The thresholds still tell you what the companies that run the inbox count as abuse, and complaint rate follows your domain at every volume. The rules point the same direction: authenticate, make it easy to opt out, keep the list clean, and do not mail people who did not want to hear from you.
The market remembers
A buyer who gets an obviously generated message learns something about your company, and a better message six months later has to undo it first. In a market of a few hundred companies that matter to you, that is not a rounding error.
What "drafts and waits" means in practice
Five rules, all of them enforced before anything reaches you.
Nothing sends without you. Every message waits in a morning queue. There is no setting that changes this.
A draft may use only what the prospect published themselves. Their post, their job listing, their release notes, their funding announcement. Never a third party's name, never a guess about what is broken inside their company. A check enforces it before the draft reaches the queue, and you are the second check.
Under ninety words, one open question, no meeting ask on the first message. The first message earns a reply, not a calendar invite.
Follow-ups are written for day 3, day 7, and day 14, and every one of them waits for approval too.
Volume is small on purpose. Messages are triggered by something that happened at a company this week, not by a daily quota, so the number rises and falls with the market instead of with a target.
Ten minutes in the morning to approve or kill what is in the queue, and one email on Sunday with what moved. That is the whole time cost. Skip the ten minutes and nothing goes out, which is the trade you are making.
Two designs, side by side
| Sends on a schedule | Drafts and waits | |
|---|---|---|
| Unit of work | Messages delivered | Conversations opened |
| A message is built from | A database record and a template | Something the prospect published |
| Who sees it first | The prospect | You |
| If you disappear for a week | It keeps sending | The queue waits |
| Cost of a bad message | Your domain, at scale | One message you approved |
| Where it breaks | Deliverability and reputation | The ten minutes you did not spend |
What this will not do
It does not replace a salesperson. It builds and runs the layer underneath one, so the first hire starts with live conversations instead of a blank page.
It does not scale to thousands of messages a week. If your model needs that, buy the other thing.
And it will not rescue an offer nobody wants. We have our own evidence on that, and it does not flatter the research. In February a four-touch LinkedIn sequence, built on templates, carrying one concrete and cheap offer, replied at roughly 25 percent. In August a batch of five deeply researched messages, each written from real detail about the company, replied at zero. Same operator, same market, better writing, no offer worth answering.
Relevance gets a message read. A concrete reason to act is what gets it answered, and no amount of generation supplies that. The approval step is where you catch a message that is relevant and still pointless.
Four questions to ask any outbound vendor
Including me.
1. Who presses send, and can I read every message before it goes? If the demo cannot show you that screen, it does not exist.
2. What is a message built from? Published facts, an inference about their pain, or a scraped profile. Ask to see the source of a claim in a sample draft.
3. What happens on a day nobody approves anything? If the answer is that it keeps sending, you have bought volume, whatever the page says.
4. Whose domain and whose name are on it, and who repairs the reputation? The answer is almost always you, which is the reason for question one.
Common questions
Is this an AI SDR?
No. AI SDR software sends on a schedule. This drafts and waits. Every message sits in a queue until you approve it, and each draft is built from something the prospect published themselves rather than from a scraped profile.
Does approving every message scale?
Not to thousands of messages a week, and it is not meant to. It scales to the number of companies worth writing to in a given week, which for a founder-led company is a smaller number than most tools assume.
What happens if I miss a morning?
Nothing goes out and nothing breaks. The queue waits. Missing a week costs you a week of first messages. It does not cost you a domain, which is what a week of unattended sending can cost.
Why does it matter who presses send?
Because the cost of a bad message lands on your domain and your name, not on the software. Mailbox providers score complaints against the sending domain, and a damaged domain takes weeks of careful sending to repair.
The honest summary is that this is the slower design. It produces fewer messages, it needs ten minutes of your attention a day, and it will not fill a quarter by itself. What it protects is the two assets a small company cannot buy back: a domain that reaches the inbox, and a name that a few hundred buyers still open.
If you want to see what your own list and your own first messages look like under those rules, that is the two-week GTM Audit: $750, your list cleaned and ranked with the reason each of the top twenty fits, and a written plan, credited to month one if you continue. If the audit says the timing is wrong, it says so in writing.