1. Look for a referral path before you consider writing. Check current customers, former colleagues, investors for a route to this person. A warm intro doesn't win on copy — it isn't queueing in the same place. As noise rises, the return on relationship stock only goes up.
2. Point AI at trigger events, not at the letter. Hiring, funding, a system migration, a compliance deadline, a new person in the seat — monitoring and summarising is what it's genuinely good at. Judging which trigger actually hurts right now stays with you.
3. One line per message that a model couldn't write. The specific thing a peer of theirs told you last week, the step you got stuck on inside their product, a number from something they said last month. The test: if the sentence would still be true for another company, it doesn't count.
4. Cut volume, raise preparation time per message. The structure from the piece on cold outreach still holds — what changed is that volume no longer buys an edge, because it got cheap for everyone.
Mechanism: an arms race under a congestion externality. When marginal cost approaches zero and total attention is fixed, "send a few more" is always individually optimal, and in aggregate it pushes everyone's reply rate down until it barely covers the residual cost. Personalised copy is the same story — it worked because it was scarce, not because it carried information, and once anyone can generate it the rent disappears.
Boundary: outreach isn't dead — the "bulk and generic" box is. More noise actually raises the value of the other box: few, precise, a real trigger event, a verifiable sender. In a genuinely narrow market, fifty accurate letters still build a pipeline.
1. An interactive bot says it in its first line. "I'm the automated assistant — I can pull pricing and availability; for anything about the proposal I'll hand you straight to the person who owns your account." Put the disclosure and the exit in the same sentence — "I'm a bot" alone loses points; "I'm a bot and here's who takes over" buys certainty.
2. The more subjective and consequential the stage, the more a person shows up. Pricing, bad news, objections, negotiation, the first call after something breaks — all human. What buyers evaluate there isn't whether the answer is right, it's whether this person will be on their side when it goes wrong: precisely the inference the label switches off.
3. Don't put AI behind a human mask. Caught once, everything true you ever said gets repriced — and Article 50 of the EU AI Act requires people be told they are interacting with an AI system (applying from August 2026). The test from the piece on sales ethics still does the work: would this survive being said out loud?
4. The more unusual they think their situation is, the more a human takes it. The more people believe their own case is unique, the more they resist handing it to an algorithm. When you hear "we're not like the others", the answer is a person plus a question about the specific difference — not a longer automated report.
Mechanism: most trust in selling rests on an inference — why did he say that, is he looking out for me. That inference has a fairly specific home in the brain (the mentalizing network: medial prefrontal cortex and right temporo-parietal junction), and how strongly it engages depends on how human the counterpart is taken to be. Label the counterpart a machine and the buyer stops inferring intent and simply inspects the output. Where correctness is the whole job that costs nothing; where the question is "is he on my side", what you lose is the only currency there is.
Failure mode: don't read this as "conceal where you can". Beyond the law there's a purely economic reason — selling is a repeated game, and being caught is priced against the whole relationship. Also, the disclosure penalty isn't a constant: it's far smaller for buyers already used to AI assistants.
1. Lay last week out in three columns. Most people find column one is a third to 40% of it — list building, notes, follow-up reminders, answering the same question again. Hand that over this week; it is nobody's advantage any more.
2. Column two gets one hard rule: verification must cost less than the time saved. Drafts, competitor comparisons and business cases genuinely halve; but if you spend forty minutes checking three numbers it invented, that item goes back to column three and you write it. The illusion of saved time is the expensive trap here.
3. Write column three as actions, not adjectives. Not "build trust" — get three buyers' real approval chains this month; say out loud which option they shouldn't buy; commit in writing to showing up when it breaks. The prescribe-don't-pitch logic from the consultative selling piece is the moat: a model can generate advice, it cannot carry the consequences of the advice being wrong.
4. If you're new: this is your catch-up window, and the fastest-closing door. AI can pull your scripts, structure and follow-up discipline up to average — but average itself is depreciating. Spend every saved hour on real customer conversations; that's the input a model can't give you.
Mechanism: these tools turn "how the best people do it" from tacit knowledge into a copyable public good. Once copying gets cheap, the rent on that skill disappears — the beneficiaries are the people who were furthest from best practice, and whoever used to earn a premium on that tacit knowledge gains nothing. What premium remains has to come from things the model can't reach: a buyer's private internal constraints, judgement someone has to answer for, and relationship stock that can't be generated.
Boundary: a higher floor is good news if you're new, and it also means "slightly better than average" no longer pays. And if you hand over column one without moving the time into column three, the net result is just doing the same undifferentiated work faster. The measurement discipline from the data-driven selling piece applies: compare output per hour before and after, not how it feels.
1. Numbers trace to systems; customer facts trace to records. Any percentage, amount or timeline you send must point back to the CRM, the finance system or a real call record; model-generated numbers never leave the building. This isn't fastidiousness — one invented case study, caught once, costs you every remaining deal in that relationship.
2. Every outbound piece has a named person who signs it. Not process approval — the person the customer calls when it's wrong. A signature makes someone actually read it, which is the cheapest gate there is.
3. The bigger the time saving, the more the check has to be written as an action. "Review carefully" is not a check. "Verify three numbers, open one cited source, confirm the name and title" is — and it fits in two minutes, which is why it gets done.
4. Keep an editing seat; don't go fully automatic. This is also about whether you'll still be using it in three months: people abandon an algorithm outright after seeing it err once, while being allowed to adjust the output even slightly restores use substantially. Draft-then-human-approve outlives full automation.
Mechanism: people treat fluency, confidence and clean formatting as cues to correctness, so generative output naturally survives one quick look; and whenever an automated cue is present, independent checking drops (that's automation bias). Stacked together, the error rate isn't high but the miss rate is — and what gets missed is exactly the expensive category: things outside the model's competence where you have no independent information to compare against. Cheap checks written as actions are the only thing that holds both down.
Failure mode: don't turn verification into a second full rewrite — that saves nothing and the team quietly skips it within three weeks. Checks should be short, specific and doable in two minutes. The other common trap is pasting sensitive customer data into an uncontrolled tool; settle the compliance question before legal comes to ask you.
One hour to turn "we should use AI" into four things you can check and change.