AI Is Writing Your Cold Emails, and Spam Filters Can Tell

Somewhere in the last two years, cold email quietly changed shape. Industry estimates now put AI-generated messages at over 40 percent of all cold email traffic, and the number keeps climbing as more tools promise to write, personalize, and send at scale with almost no manual effort. On paper this sounds like a win. In practice, it created a new problem nobody fully priced in: when everyone uses the same handful of AI tools, trained on similar patterns, a huge share of cold email starts to sound identical. Inboxes did not just get fuller, they got more repetitive, and both humans and spam filters noticed the pattern faster than most senders expected.

Every AI writing tool has a voice, and that voice repeats across millions of emails. Certain phrases became so common that they turned into red flags on their own: formal greetings, a compliment about the recipient's company growth, a soft transition into the pitch, a closing line asking for fifteen minutes. None of these phrases are individually suspicious. The problem is volume. Spam filters are trained on enormous datasets of real email traffic, and once a sentence structure shows up often enough attached to low-engagement messages, the filter learns to treat that structure as a warning sign, regardless of whether the specific email behind it is legitimate. Recipients learned the same pattern independently. Recent research suggests roughly seventy percent of professionals now delete an email the moment it reads as templated, often before finishing the first sentence.

One recent side-by-side test sent identical outreach to over twelve thousand verified prospects, split between fully AI-generated copy and human-written copy using the same targeting and infrastructure. The results were not subtle. AI-only emails landed a 4.1 percent reply rate. Human-written emails on the same list landed 10.4 percent, more than double. A separate test measuring spam placement specifically found AI-only emails got flagged as spam at roughly 7.8 percent, compared to 2.9 percent for human-written emails sent under the same conditions. The gap was not about grammar or clarity, since AI-written copy was often technically cleaner. It was about pattern recognition. Inbox providers are not reading your email the way a person does. They are comparing its structure against billions of other messages, and generic AI phrasing has become one of the most recognizable structures in that dataset.

The instinct many teams have is to disguise AI usage, adding typos on purpose or forcing casual slang to sound more human. This misses the actual issue. Recipients correctly identify AI-written emails less than half the time when asked directly, so detection by a human reader is not really the main risk. The real risk is structural: an email that was generated from a generic prompt, without real research behind it, tends to read as generic regardless of how it is phrased, because it has nothing specific to say. Trying to fake casualness on top of a hollow message does not fix that. What actually closes the gap is giving the AI real, specific information to work with, so the output has substance instead of just a friendlier tone wrapped around emptiness.

The senders getting strong results from AI-assisted outreach are not the ones avoiding AI, they are the ones changing what they feed it. Instead of a prompt like "write a cold email to a marketing director," the input includes the prospect's actual website, a recent product launch, a hiring pattern, or a specific detail from their public activity. The AI is doing the writing, but the substance comes from real research, not a generic template pretending to be personal. This distinction matters more than people expect. Personalized emails built around real account research have been shown to significantly lift open, response, and click-through rates compared to generic sends, and the difference tends to widen further once you account for how spam filters treat repetitive, low-substance patterns.

This is the gap that a research-first approach to outreach is built to close. Instead of generating a message from a short prompt and hoping it lands, the process starts by pulling real signals about the prospect's business, their site, recent news, and public activity, and only then builds the message or page around what was actually found. The output still moves at AI speed, but it is grounded in something real rather than a generic pattern repeated across thousands of sends. This is part of why tools like Greve start with account research before generating anything, building a page around real findings instead of asking an AI to invent relevance from a one-line prompt. The writing is still fast. The substance behind it is not fabricated.

Feed your writing tool specific, current information about the prospect, not just their name and job title. Avoid the same opening structure across your entire sequence, since repetition across sends is one of the clearest patterns filters pick up on. Keep the message short and specific to one angle instead of covering several reasons to buy in one email. Read the output out loud before sending, since anything that sounds like it could apply to a hundred other companies probably will. Vary sentence length and structure across your sequence rather than relying on the same generation pattern for every message. None of this requires abandoning AI tools. It requires giving them better raw material to work with.

AI is not killing cold email, and it is not going away from outreach workflows either. What it did was raise the cost of generic messaging past the point where it quietly works. A templated blast that used to slip through five years ago now gets caught by pattern-matching filters and deleted on sight by recipients who have seen the same structure a thousand times. The senders who benefit from AI going forward are the ones treating it as a writing accelerator sitting on top of real research, not a replacement for research itself. The technology got faster. The bar for what counts as a message worth reading did not get lower, it got higher.