Vol 1 · Issue 90 · Wednesday, August 12, 2026

There are two guys at every business dinner right now and they are both wrong.

The first guy is going to tell you he runs his whole company on AI. He has agents. He has workflows. He says the word "stack" a lot. If you ask him what his revenue did last year he changes the subject to something he read about model releases.

The second guy folds his arms and says he does not use any of it. Says it is a fad. Says his clients want a real person. Says this while paying somebody eighteen dollars an hour to retype meeting notes into a CRM.

Both of them are leaving money on the table. The first one is going to embarrass himself in front of a client eventually. The second one already is, he just cannot see it because the bleeding is slow.

The useful position is boring and sits in the middle, and it is one sentence long.

AI is a first draft machine. It is not a judgment machine.

Everything downstream of that sentence is just knowing which of your tasks are drafts and which ones are judgment.

A first draft task has three properties. There is a lot of raw input. There is a predictable output shape. And a human is going to look at it before it matters.

That last one is the whole ballgame. If nobody checks it before it goes out the door, it is not a first draft task, no matter how much you want it to be.

So go look at your week. Not your strategy, your actual week. The Tuesday afternoon stuff. How much of it is you taking a pile of information that already exists and reshaping it into a different format for a different audience?

For most service businesses the honest answer is somewhere between a third and half of everything. Meeting happened, now it needs to become notes. Notes need to become a follow up email. Call happened, now it needs to become a proposal. Proposal got signed, now it needs to become an onboarding plan. Month ended, now it needs to become a client update.

None of that is thinking. All of it is transcription with extra steps. And all of it currently costs you or somebody on your payroll real hours.

I am going to skip the theory. Here are the six places I see this pay off consistently in service businesses under a few million in revenue, roughly in order of how fast you will feel it.

One. Call to follow up. Record the call, get the transcript, turn it into a follow up email that references the specific things the person said. This is the single highest return use in the entire category and almost nobody does it. Most follow up emails are generic because the person writing them is going from memory four hours later. A follow up that quotes the actual worry the prospect raised closes deals that generic follow up does not.

I use Fathom to capture the calls, mostly because it stays out of the way and the transcript quality is good enough to work from directly. Then the transcript becomes the raw material for everything else that call is supposed to produce.

Two. Proposal drafts. You already have a template and you already have a call transcript. The gap between those two things is about forty minutes of your life every single time, and it is a gap a machine closes well. Feed it your best previous proposal as a format example, feed it the transcript, get a draft, then spend fifteen minutes fixing the parts that are wrong instead of ninety minutes staring at a blank document.

Three. The client update nobody has time to write. You know the one. The monthly email where you tell the client what you did and why it mattered. Most owners skip it, and then wonder why clients who are getting good results still churn. If you have any kind of activity log, that log plus a prompt gets you a readable update in about four minutes. Clients renew for work they can see.

Four. One asset into ten. You wrote something good, or you recorded a twenty minute answer to a client question. That is not one piece of content, that is a week of it. The reshaping is mechanical. The original thinking already happened.

Five. Inbox triage. Not sending, triage. Sort what came in, flag what actually needs you, draft responses to the routine five that you will read before sending. Half your inbox is the same eleven questions and you have answered all of them before.

Six. Writing down the thing that only lives in your head. Talk through your process out loud for ten minutes while screen recording. Feed the transcript in. Get a written procedure back. It will be about seventy percent right, which is roughly seventy percent more than the zero percent that currently exists in writing.

The models themselves are close to interchangeable for this kind of work, and the differences matter less than people on the internet insist. I run most of mine through Claude.ai and ChatGPT.com depending on the job, and I keep them in one place through Galaxy.ai so I am not paying four separate subscriptions to compare outputs. Pick one, learn it properly, stop reading launch announcements.

I want to make this concrete because the numbers are more boring and more convincing than the theory.

A client of mine runs an IT services shop, fourteen people, mostly small business contracts. His problem was not sales and it was not delivery. It was the eleven hours a week his operations manager spent turning things into other things. Notes into tickets. Tickets into client summaries. Calls into scope documents.

We picked one loop. Calls to scope documents, because that was the one holding up money.

Before: client call, then somewhere between two and nine days later a scope document appeared, because the person writing it had fourteen other things and this one had no deadline attached to it.

After: the call ends, the transcript lands, a draft scope doc gets generated against their standard template, and it sits in a folder with the operations manager's name on it. She opens it, fixes what is wrong, sends it. Average time from call to document went from about four days to about six hours.

Nothing about that is impressive technologically. Here is the part that is: their close rate on scoped work went up nine points. Not because the documents got better. They got slightly worse, honestly. They got faster, and it turns out that in their market the person who sends the scope first usually wins, and the reason they were losing was never quality.

Eleven hours a week became roughly four. The nine points were worth more than the hours.

That is the shape of the return. It is not usually the time savings that pays. It is what becomes possible when a thing takes six hours instead of four days.

Now the other half, because this is the part that costs people real money and nobody writes about it honestly.

There are three ways this goes wrong and they are all reputation problems, not efficiency problems.

The first is the confident wrong fact. The machine does not know when it does not know. It will tell you a number, a date, a regulation, or a capability with exactly the same tone whether it is right or invented. If you put an unverified specific in front of a client and it is wrong, you do not lose an argument, you lose the assumption that you know what you are talking about. That assumption is most of what they are paying for.

Rule: anything with a number, a name, a date, or a promise in it gets verified by a human before it leaves the building. Every time. No exceptions for busy days.

The second is voice collapse. There is a flavor to unedited output and your readers can taste it even if they cannot name it. It is the sentence that says something without saying anything. It is the three word list where two words would do. It is enthusiasm with no specific attached. The moment your writing starts tasting like that, everything you send gets read with a little less attention, including the stuff you wrote yourself.

The fix is not clever prompting. The fix is that you rewrite the opening and the closing by hand, always, and you cut anything that could have appeared in somebody else's email. Your voice lives in your specifics. Nobody else has your Ray, your bad Tuesday, or your eighteen percent price increase.

The third is the one that actually scares me. Delegating judgment. Pricing. Whether to fire a client. How to respond to somebody who is upset. Whether a project is going sideways. These feel like they have right answers and they do not. They have consequences you own. A machine has no skin in the game and it will produce a perfectly reasonable sounding answer to a question that required you to know something it has no way to know.

You can ask it to lay out considerations. You can ask it to argue the other side. You cannot ask it to decide, and if you catch yourself doing that because deciding is uncomfortable, that is a you problem wearing a technology costume.

Practical, since this is where most people give up.

Feed it your own writing. Not a description of your writing, the actual writing. Paste in three things you wrote that sounded like you, tell it those are the reference, and ask it to match. This works dramatically better than any adjective you can supply. "Conversational and direct" means nothing. Three emails you actually sent mean everything.

Then ban the tells. Tell it up front: no em dashes, no phrases like "in today's fast paced world," no sentence that starts with "It's not just X, it's Y," no headers unless asked. Every writer has their own list. Build yours by noticing what makes you wince.

And keep the messy parts. The aside. The half joke. The sentence that runs a little long because you got going. Those are proof of life. Sanded off, everything reads like a press release.

Here is the thing that separates people who get value out of this from people who post about it.

If using it requires you to remember to use it, you will use it for eleven days.

The wins come from wiring. Call ends, transcript lands somewhere automatically, draft gets generated, draft shows up in your queue waiting for a human. You did not decide anything. It just happened, and now there is a draft sitting there and the only remaining task is the one that needed you. That is what Make.com is for, and it is why the automation piece and the AI piece are really the same conversation. The model is the engine. The wiring is the car.

Start with one. Genuinely one. The call to follow up loop is the one I would pick if somebody made me choose, because it touches revenue directly and you will know within two weeks whether it is working.

Pick the task you do most often that is clearly a first draft task. Do it the machine way for thirty days. Track two things: minutes saved, and how many times you had to fix something material.

If minutes saved is high and material fixes are low, wire it and move to the next task.

If material fixes are high, you either picked a judgment task and called it a draft task, or your inputs are garbage. Usually inputs. The quality of what comes out is almost entirely a function of what you put in, which is an uncomfortable amount like managing people.

That is the whole thing. Not agents. Not stacks. One task, thirty days, honest numbers.

The guy at dinner running his whole company on AI is not going to do that, because it is boring and it does not sound good out loud. Which is exactly why it works.

Talk Soon,

Dan
Dan Kaufman
Founder, Dead Simple Growth and Pinnacle Masters

P.S. I keep a running list of the exact prompts I use for the six plays above, including the one that turns a call transcript into a follow up that actually sounds like a person. Reply with the word AUTOMATE and I will send you the whole set.