2021 to 2023: mailbox rotation and AI drafting make campaigns cheap, and the filters learn faster

Warmup, mailbox rotation and AI drafting work at what they were built for: between 2021 and 2023 they made a new cold campaign nearly free to launch, from a spare domain and a dozen mailboxes to a fluent first draft. They did not make the email worth reading, and in the same years mailbox providers got faster at classifying senders by behavior, infrastructure and feedback, while polished copy stopped telling the reader anything about the sender.

How did a cold campaign get so cheap?

Start with the volume problem. A single mailbox that suddenly sends hundreds of cold emails a day looks like what it is. So senders spread the load: extra domains that resemble the company's own, a few mailboxes on each, a modest cap per mailbox per day, and software that rotates the sends across all of them. The per-mailbox cap has some support in data. In Hunter's analysis of 31 million emails sent by its users in 2025, mailboxes sending 20 to 49 emails a day got a 5.7% reply rate against a 4.5% average.

New mailboxes have no history, so a second tool filled that gap. Instantly's help center describes how its warmup works: your mailbox joins a pool of other users' mailboxes, sends warmup emails to them and receives theirs, every warmup email is automatically opened, and a high percentage get positive replies. The goal is a record of human-looking interaction before the first real prospect sees a message.

Then the writing got cheap. OpenAI released ChatGPT on November 30, 2022. In March 2023 Google said Gmail users would be able to draft, reply and summarize with generative AI, starting with trusted testers in the United States. By spring 2023 the same capability was arriving on both sides of the inbox: the sender drafted with it, and the recipient's own provider was testing it for the recipient.

What did the filters learn in the same years?

Authentication got its first hard floor. Google said in October 2023 that the year before it had started requiring mail sent to Gmail addresses to carry some form of authentication, and that the number of unauthenticated messages Gmail users received fell by 75%.

The text classifiers got harder to fool. In November 2023 Google described RETVec, a text vectorizer built to resist the tricks used against classifiers: homoglyphs, invisible characters, keyword stuffing. Putting it into the Gmail spam classifier improved the spam detection rate over the baseline by 38% and cut the false positive rate by 19.4%.

The anti-abuse industry named the workaround. No mailbox provider has published a rule that names warmup pools. M3AAWG, the industry's anti-abuse working group, has published its view in its current sender best practices.

Artificial warming practices are not recommended and may be illegal.

The same document describes real warming as raising volume slowly from a new domain, about six weeks on average, and watching deferrals, bounces, complaints and unsubscribes along the way.

Why did competent copy stop being a signal?

Before 2022 a well-written, specific-sounding email was a weak sign that someone had spent time on it. Generative drafting removed the cost, so it removed the signal. Dana Reyes, head of sales at Northwind, now gets several emails a week that open with a compliment about a post she wrote, in fluent paragraphs, each one generated from the same public profile. They read well. They read the same.

Readers noticed. Hunter's 2026 State of Email Outreach report, which pairs those 31 million emails with recipient surveys, found that 69% of decision makers say it bothers them when AI was used, unless the output feels genuinely human. In the same data, manually edited emails got a 5.2% reply rate against 4.4% for fully automated ones. It is vendor data and a correlation, not an experiment, but it points the same way as the survey.

The point is not that AI drafting is bad. It is that a draft is no longer evidence of effort. Evidence has to come from something the reader can check: a current fact about their company, stated plainly, with nothing invented around it.

What era four means for a head of sales in 2026

• More mailboxes is a capacity decision, not a strategy. Each one is infrastructure to authenticate, monitor and eventually retire. Part 5 covers what every one of them must now pass.
• Warmup is not reputation. Reputation is what real recipients do with your mail. Engagement manufactured by a pool of mailboxes opening each other's email is the practice M3AAWG tells senders to avoid.
• Use AI where it is cheap and accurate, and let a person decide what sends. Research and first drafts are good uses. Choosing which opening, value line and ask actually earn replies is a measurement problem, and part 10 shows how to score them separately.

Who this does not apply to: if your outbound already runs from one well-kept domain at low volume, with every email read before it sends, this era's shortcuts were never yours to unwind. The rest of the series still applies.

Next in the series

Part 5 covers February 2024, when Gmail and Yahoo turned authentication, one-click unsubscribe and a spam-rate ceiling from best practice into enforceable rules, and Microsoft followed in 2025. For what came just before this era, see part 3. The series page lists every part in order.