AI extrapolates fluency. I extrapolate emotion. AI 外推流畅，我外推情绪

Two failure modes, opposite by nature

Watch either side long enough and the pattern is consistent. Both kinds of failure are extrapolation, filling a gap with something plausible, but they fill different gaps.

In practice both columns come up regularly, and both have caught real damage. I’ve caught drafts claiming work was done when it wasn’t, tripled time estimates offered on someone else’s behalf, and sentences that read as a dig at a colleague who would have been on the email. Going the other way, I’ve been talked out of treating a client’s silence as displeasure, and shown, by going back through the thread, that everything I’d read into it was contradicted by what was actually written.

One detail matters more than the rest: the AI had the full context and still missed things. The sentence that undercut a colleague was written with every relevant file in its context. So the review step isn’t there because one side lacks information. It’s there because every layer has its own blind spot, and the only reliable fix for a blind spot is a second layer that’s blind somewhere else.

The reading time is the product, not a tax

Spending two hours understanding a draft I could have forwarded in five minutes looks like waste until you ask what the client is paying for. They’re paying for a person who will put their name on the result. You can’t stand behind something you don’t understand. When they call tomorrow and ask what a particular margin figure means, answering clearly is part of the deliverable, and those two hours are why you can.

And the investment compounds. Once you understand a mechanism, it stays with you, and the next letter that touches it costs almost nothing. The cost for a given client falls on its own as templates and shared shorthand build up, so don’t judge the steady state by what a phase-boundary letter costs.

Here’s what I’d bet on: in two years the market will be full of people forwarding AI drafts. They’ll fail in exactly the ways in the AI column above: the invented completion, the borrowed estimate, the sentence that undercuts someone. Not because their model is worse, but because someone who just forwards has no second layer. The process is the advantage, and an unusually cheap one.

Tiered review: spend where the stakes are

Reviewing everything at full depth doesn’t scale and isn’t necessary. Three tiers, chosen by what the letter can cost you:

• S: acceptance, quotes, scope, conflict, first impressions. Understand every detail, however long it takes. With a new client the first several letters are all tier S, because the first impression sets the price of everything that follows.
• A: substantive technical exchanges. Thirty to sixty minutes on three things only: (1) every commitment, checking tense, numbers and dates; (2) a scan for anything that undercuts a person, speaks in absolutes, or estimates on someone else’s behalf; (3) can I restate the core mechanism in my own words? If I can’t, I didn’t understand it, and it doesn’t go out.
• B: confirmations, status updates, short replies. A two-minute scan, then send.

The restatement test in tier A does most of the work. It isn’t a stand-in for understanding; it is understanding, and it’s the only one of the three checks you can’t fake by skimming.

What the AI owes the review

The pair only works if the reviewer knows where to look. So every draft comes with a list of review points: the two or three places the AI is least sure about, the full list of commitments the draft makes, and every sentence that names a person or carries a tone. That one habit cut my review time by about a third without reducing what I caught, because the time now goes where the problems are most likely to be.

The rule I’d keep if I could keep only one

Every “X should happen” must be preceded by X actually happening once. Checklists, procedures, promised behaviors: all of them get written after the thing has been done at least once, never before. Written the other way round, they read exactly the same, and they’re fiction. It’s the same discipline as the rest of my work: don’t automate a flow you haven’t run by hand many times, and don’t claim a result you haven’t measured.

Written from how I actually handle client correspondence. This note was itself produced by the pairing it describes, and the first draft contained three commitments I had to go back and verify. No specific client, project or person is described.

Related: the same “make it structural, don’t rely on anyone being careful” approach applied to a machine, in AI-written code always has bugs nobody has found yet, and to money, in Governing AI agents that touch money. On keeping communication overhead low in the first place: One email per milestone (on boheastill.com).