On AI, honestly.

There's a question I've been asked in every prospect call so far, and I notice it's almost never asked directly. It's asked sideways. "How do you scale this?" Or: "What's your tooling stack?" Or sometimes, the bravest version: "I have to ask — are you using ChatGPT for this?"

The question underneath all three is the same. The prospect wants to know if she's about to pay a premium for something an AI tool is already producing for free somewhere else. She's right to ask. The honest answer is one almost no agency is giving, and I think the absence of the honest answer is doing more damage to this category than anything AI itself has done.

So here it is, said plainly.

I use AI. A lot. Every day. In specific places, for specific jobs, with a clear sense of where it stops being useful and where it would actively make the work worse.

I want to walk you through where it goes in and where it stays out, because the line is the whole answer. Anyone who tells you they don't use AI in this work in 2026 is either lying or producing a worse version of the work than they could be. Anyone who tells you they use AI to write the messages is producing exactly the kind of output you've been pattern-matching and archiving for the last two years. The truth, like most true things about this work, is in the middle and requires more nuance than either pole.

Where AI goes in.

The first place is research synthesis. A senior buyer's footprint, across a quarter, is enormous. Earnings call transcripts that run forty pages. A podcast appearance that runs forty-five minutes. Six months of LinkedIn posts. Maybe a conference talk on YouTube. Maybe a long-form interview on a Substack I've never heard of. If I read all of that linearly, the thirty-minute budget evaporates before I've gotten through the second source. So AI does the first pass. It pulls the transcript of the podcast in ninety seconds. It summarizes the last quarter's earnings commentary into the three themes that repeat. It tells me which of her LinkedIn posts in the last six months are about the topic I'm looking for, so I can skip the ones about her dog and her vacation.

That's not me being lazy. That's me being able to actually do the work in thirty minutes instead of three hours. Without AI, the thirty-minute floor wouldn't exist — it would be a two-hour floor, and the business would price out of every engagement.

The second place is draft compression. When I write the first pass of a message, it's almost always too long. 140 words. Sometimes 180. I know it has to come down to around seventy. AI is good at the first round of compression — at finding the four "I wanted to reach out" filler phrases, the two adjectives that aren't doing work, the one "I'd love to learn more" that could go. It's a fast second pair of eyes on the cuts that don't require taste, only ruthlessness.

The third place is pattern recognition. If I'm writing for a founder whose prospects are all VPs of Clinical Operations, I'll feed the public footprint of fifteen of them into a session and ask what the common preoccupations are this quarter. The output isn't the message. The output is the category map — the three things this kind of buyer is most likely to be circling right now, which helps me get to the right angle faster when I do the actual reading on any individual prospect.

That's where AI goes in. None of it is the message. None of it is the judgment. All of it is the volume work that used to take hours and now takes minutes, which is the only reason this version of the practice can exist.

Where AI stays out.

AI cannot hear how someone speaks.

I keep coming back to this because I haven't found a better way to say it. The podcast appearance is the most important source I have on a senior prospect, and the most important moment in the podcast is the one where her cadence breaks. Where she stops giving the prepared answer and starts thinking out loud. Where there's a half-laugh, a pause, an "honestly, I don't know yet." AI can transcribe the words. It cannot tell you that the pause before that sentence was a tell. It cannot read the room of a board update. It cannot tell you that her hiring post last week was defensive rather than ambitious, because that distinction lives in the gap between what she said and what she would have said if she'd had three more drafts.

The layer-four observation — the one specific thing nobody else would have noticed — almost always lives in something AI can't surface. A contradiction between her public preoccupation and her company's public output. A moment of unrehearsed honesty in an interview. A pattern across three posts that requires holding all three in your head at once and noticing what's missing from the third one.

I have tried, more than once, to get AI to surface a layer-four observation. It cannot. It surfaces what I'd call a layer-two-and-a-half observation — something better than what an SDR with three minutes would find, but worse than what I'd find with thirty minutes of careful reading. The gap between those two outputs is small in word count and enormous in reply rate. The prospect can tell the difference. She is doing nothing all day, in her inbox, except telling the difference.

The other place AI stays out is the writing itself. I do not let AI write the messages, and I will not. Not because AI writes badly — it often writes fine. Because the prospect can pattern-match AI-written prose in under two seconds, and the pattern-match is not based on quality. It's based on rhythm. There is a specific rhythm to AI-generated sentences — a kind of evenness, a balanced clause structure, a tendency toward symmetry — that any senior reader can spot at a glance. Once she's spotted it, the message is over. The thirty minutes of research are erased by the rhythm of the prose.

The way I write a sentence has a different rhythm than the way an AI writes one. Mine has rougher edges, more starts and stops, the occasional clause that runs longer than it should because the thought ran longer than it should. AI smooths those edges out by default. The smoothing is the tell.

Where the line lives.

The line, then, is roughly this: AI handles the volume work that doesn't require judgment. Humans handle the judgment work. The two combine in the middle, where the human looks at what the AI produced and decides which 12 percent of it is actually useful.

That 12 percent is the work. The 88 percent the AI did to produce it is the cost of the work, the same way the 22 minutes of reading per message that don't show up in the final draft are the cost of the message. Neither is the product. The product is what survives.

The reason I'm telling you any of this is that the alternative is the lie almost every agency is currently telling, which is some version of "we don't use AI, all our work is human." That lie is unsustainable, and I think it's actively damaging the trust senior buyers are willing to extend to anyone who claims to do real outbound. The buyer reads the messages. The buyer can tell which ones an AI wrote and which ones a human wrote. The buyer cannot tell — and does not care to tell — whether the human who wrote the message used AI to summarize a podcast in advance. The line is in the right place if she can't tell, and in the wrong place if she can.

What you should pay me for, if you pay me at all, is the judgment. The decision about which 12 percent. The thirty minutes of reading that lets me get to the right angle instead of the obvious one. The taste that knows which AI-generated summary line is useful and which one is the kind of synthetic insight that would land flat in a message. If you're paying me for the message and not the judgment, you're overpaying, because the message is the last ten minutes of a thirty-minute process and the message is the part the AI helped with most.

You're paying for the part the AI couldn't do. That's the part the prospect is also, silently, paying for when she replies.