No send button, and never will be

Open the outreach panel in Open Pulse and you will find a first message written from what the person actually said: their post, their words, the moment that made them worth contacting. Read it, edit it, copy it out. What you will not find is a send button. That is not a missing feature. It is a deliberate line, and the last two years of sales software explain why.

The experiment that burned the channel

Two years ago the hottest pitch in B2B software was the AI SDR: a digital worker that would prospect, write, call and book meetings while you slept. Hundreds of millions of dollars went into it. Then the post-mortems started. Dan Worman, writing in July 2026, put the obituary plainly: the autonomous SDR failed because outbound is not a volume problem, it is a judgment problem. The tools scaled the sending and skipped the knowing, producing "spam at machine speed" plus hallucinated personalisation, his example being a "congrats on your Series B!" sent to a 40-year-old family firm. (Opinion piece; the author notes he has a vested interest, which is worth knowing before you trust it.)

The numbers underneath that obituary are measured, though the loudest ones come from vendors with something to sell, so read them with that filter on. SmartReach.io's State of Cold Email 2026, drawn from more than 40 million emails across 5,000-plus campaigns in the first half of 2026, reports reply rates falling from 8.5 percent in 2019 to between 3 and 5 percent in 2026, with disciplined agency-run campaigns holding near 10 percent (EIN Presswire, 23 September 2026; SmartReach sells cold email software). Instantly's 2026 benchmarks put the average at 3.43 percent (vendor-authored). A UK sales-training firm, Maverrik, described an "AI deadlock" in September 2026: sellers using AI to create outreach while buyers use AI to remove it, with large-scale studies recording human reply rates below 1.5 percent (The European; Maverrik sells sales training too).

Then the channel itself started punishing volume mechanically. In February 2024 Google and Yahoo made three things mandatory for anyone sending 5,000 or more messages a day to their addresses: authenticate with SPF, DKIM and DMARC, provide one-click unsubscribe, and keep the spam complaint rate below 0.3 percent, aiming for under 0.1 percent. Microsoft applied the same bar to consumer Outlook addresses in May 2025, and by late 2025 non-compliant bulk mail was getting permanent 550 rejections rather than quietly landing in spam (Tech Times, 22 September 2026). Three complaints per thousand delivered emails and the domain starts getting throttled or blocked, sometimes for weeks. An autonomous sender firing thousands of machine-written messages a morning was always going to cross that line, and "eventually" turned out to be fast.

The failure mode is not being wrong

Here is the part the AI SDR post-mortems keep circling: the thing that kills a first message is not a wrong fact. It is the smell of being generated. Buyers now spot AI outreach in seconds and delete it in one. In September 2026 a founder received a cold email from an AI SDR company that had raised around $46 million; it opened by quoting one of his own posts back at him, then pitched the AI BDR. It landed in spam. Two weeks later the company reversed its whole position (Sanket Goyal, Medium, 17 September 2026; a single anecdote, but a pointed one).

This is why the failure compounds. A wrong sentence costs you one reply. A recognisably generated sentence costs you the contact, because the reader has filed you under "machine" and every future message starts in that folder. Worse, it costs you the domain: complaint rates are measured per sender, and the February 2024 rules mean a burned domain stays burned for weeks regardless of the business behind it. Our rulebook states it directly: the failure mode of a generated first message is not being wrong, it is being recognisably generated, which burns the contact permanently. That sentence is what the whole outreach design hangs on.

Judgment is the part you cannot outsource

Worman's line is worth repeating because it is the entire argument in one sentence: outbound is not a volume problem, it is a judgment problem. An AI that sends a thousand emails is not a thousand times better than a rep who sends one good one. The thing that makes outreach work, knowing this company, this person, this moment, is exactly the part the autonomous tools skipped.

The market has, by most accounts, figured this out. Industry analyses from early 2026 report annual churn of 50 to 70 percent for autonomous SDR tools, and the vendors that sold "AI replaces your reps" have repositioned as human copilots (LinkedIn, 2026; vendor-adjacent analysis, treat the band as directional). The model that survived is human in the loop: AI does the research, the drafting and the data plumbing, and a human reviews before anything reaches a buyer.

There is a reason that model wins, and it is not sentimentality. The press-send moment is the only point in the pipeline where judgment is cheap and its absence is expensive. Research and drafting get better with scale; sending does not. Every message carries your name, your domain's reputation, and one irreversible first impression. A human at exactly that gate, and nowhere else, is not a compromise between automation and craft. It is the correct division of labour.

What Open Pulse drafts, and what it refuses to

So here is what the outreach panel actually does. From a signal, it drafts a first message written from what the person actually said, not from their LinkedIn headline and not from a template with their company name mail-merged in.

Eleven channels, because a Reddit comment, a Hacker News reply and a public response to a one-star Trustpilot review are not one format with three word counts: email, LinkedIn message, LinkedIn comment, LinkedIn DM, a generic forum reply, Reddit, Hacker News, X, Facebook, YouTube, and a review response for G2, Trustpilot or Google. Four tones: helpful, which is the default because it is the register that actually works on a public thread, plus direct, warm and consultative as deliberate departures from it. A failed generation does not spend your allowance, because a model's bad afternoon should not eat your month. Go includes 50 drafts a month, Growth 200, Pro 500.

And here is what it refuses to do, because each refusal is load-bearing:

• No placeholders. A [Name] in a draft is a [Name] in a sent email. If the model does not know something, it does not get a token to hide behind.
• No invented facts about the recipient. The draft may only use what the signal actually contains. It may not upgrade "they mentioned evaluating tools" into "they are the decision-maker for a Q4 purchase", because that is how "congrats on your Series B" emails to family firms get written.
• No claimed case studies. There are no testimonials, no customer logos and no case studies on this site, because there are none to show yet. The drafts are not allowed to invent what the marketing will not claim.
• None of the openers that announce themselves. The sentences that exist only to prove the message was personalised, the "I noticed you..." scaffolding, the flattery that reads as mail merge with extra steps. If the reader can feel the machinery, the draft has failed.

Copy is the only exit. There is no send button, and there will not be one. The product's claim is that it finds people worth talking to, not that it talks to them.

Why the line is drawn where it is

Every product has a sentence that explains what it will not do, and that sentence is usually more honest than the feature list. Ours is the missing send button. It says: we will do the finding, the reading, the judging of intent, the scoring and the drafting, because those get better with machinery. We will not do the sending, because that is the one step where a human's judgment is the entire value, and automating it is what burned the channel in the first place.

The draft caps say the same thing in numbers. Fifty a month on Go, 500 on Pro: enough to work every real signal your listeners surface, not enough to become the machine-speed sender the February 2024 rules were written to stop. The cap is not stinginess. It is the product refusing to let you become the thing it was built to replace.

If you want to see what a first message looks like when it starts from evidence instead of a list, run a free report: a capped, real scan of what is being said about the problem you solve. Take the first signal, draft the first message from what the person actually said, and then press send yourself. That last step is yours, and it always will be.

References

• Dan Worman, "The AI SDR is dead. Good riddance. Here's what we're building instead", LinkedIn, 2 July 2026. Opinion piece; the author discloses a vested interest. Source for the volume-versus-judgment framing and the "spam at machine speed" line.
• SmartReach.io, reported via EIN Presswire, "The Entry-Level Sales Job Is Vanishing. Data From 40 Million Cold Emails Explains Why", 23 September 2026. Vendor-authored benchmark: 40M+ emails, 5,000+ campaigns, January to June 2026. Source for the 8.5% to 3-5% reply rate decline.
• Maverrik, reported in The European, "AI Arms Race Could Mean Fewer Cold Emails Reaching Your Inbox", September 2026. Vendor-authored (a sales-training firm). Source for the "AI deadlock" and reply rates below 1.5%.
• Tech Times, "Why Google and Yahoo's Bulk Sender Rules Made Email Verification Non-Negotiable", 22 September 2026. Source for the Microsoft extension and the 550 rejections.
• Sender.net, "New Email Sending Requirements for Gmail and Yahoo", updated 11 April 2025. Source for the February 2024 authentication, unsubscribe and 0.3% complaint-rate requirements.
• Sanket Goyal, "Here is why AI SDRs are bullshit", Medium, 17 September 2026. A single anecdote and an opinion, cited as such.
• Coevera CRM, "The AI SDR Experiment Is Over. The Human Advantage Is Back", LinkedIn, 2026. Vendor-adjacent analysis; the 50-70% churn band is treated as directional.
• "Cold Outreach Is Losing Impact. Signal-Based Prospecting Is Changing the Game", LinkedIn, 2026. Cites Instantly's 2026 benchmarks (3.43% average reply rate) and Autobound's 2026 State of AI Sales Prospecting (15-25% for signal-referencing emails). Both vendor-authored.