How to Subtract: What to Stop Asking AI For

A one-week tally, how to read it, and what to change on Monday

You won't get a grade at the end of this guide. You'll get one decision for Monday. All week, you jot down a short line every time AI hands you something. Friday, you read the lines in one sitting. Monday, you change one thing.

This isn't a verdict on whether AI is any good. It tracks what you do with what it gives you. So the only fair comparison is your own tally next week. Not another business, and not some number you read online.

New here? Read the article first. This guide has everything that didn't fit: the exact rules, the borderline calls, how ties break, and what to put in the new instruction.

A draft is anything AI hands back when you ask it for something: a social post, an email, a reply to a customer, website copy, a section of a proposal, a summary.

Chat answers you only read to understand something don't count. If AI explains a clause in a boat charter agreement for you, skip it. If it writes your reply to the customer who asked about that clause, log it.

All you write all week is one line per draft, in a note on your phone:

day - task type - language - letter

Three rules people forget:

1. Every new ask for the same thing is a new draft, and each one gets its own letter for what you did with it. Three tries at an email, and you sent the third one untouched? T, T, U. If you lifted one line from the first try, make that one an R.
2. The window is one week. Monday to Friday, or all seven days if you work weekends.
3. Log it as it happens. Not from memory on Friday. Memory keeps the wins and drops what you tossed.

Keep names out of the note, and out of the prompt. A line records the kind of task, not who it was for. Whatever you paste into AI gets brackets instead of names, phone numbers or a client's boat name: [client], [boat], [date]. You fill them in at the end, which takes under a minute, so the draft still counts as a U. Prices stay out too. AI writes the wording of the quote; you write the number.

• U, used as is. You sent or published it the way it came, or after under a minute of touch-ups: a typo, a name, a date. The message didn't change.
• R, redone. You kept something (the idea, the order, a paragraph) but changed the meaning, the structure, or more than half of it.
• T, thrown out. None of it made it.

Torn between U and R? More than a minute of work, or a changed message, makes it an R. Between R and T? If anything from that draft ended up in what you sent, it's an R.

Three made-up cases, so you can see how it works:

• U. A yacht charter broker gets a booking confirmation, fixes the departure time, and sends it.
• R. A fine-jewelry boutique asks for a collection description. It keeps the paragraph order but rewrites more than half the sentences, because they read like any store's catalog.
• T. A keynote speaker asks for a follow-up email to a conference organizer. Not one line is usable, so he writes it from scratch.

Friday's math adds things up by type. Write "post" one day, "Instagram" the next and "social" after that, and you end up with three types of two drafts each, none of them readable. So the types come from a fixed list, spelled the same way every time, and they match the tool:

Language means the language of the draft, EN or ES. Work in only one? Leave it off. Here's Monday for a yacht charter broker who works in both:

Mon - customer reply - EN - R Mon - customer reply - ES - U Mon - social post - ES - U Mon - proposal or quote - ES - T Mon - proposal or quote - ES - R

Step 1. Group by type and language

"Customer reply in English" and "customer reply in Spanish" are two separate reads. Count the U's, R's and T's in each. Add them up for the total.

Step 2. Under five drafts, it's not ready

Keep counting that type next week. Five is a floor we chose, not a statistical rule. With so few drafts, a single one swings the read too much.

One exception for bilingual businesses. If a type falls short of five in each language but reaches five with both added up, read them as one and label it "both languages together." You get a signal for the type, but nothing about language. If one language reaches five and the other doesn't, read only the one that made it.

Step 3. Work out the percentages, and keep the counts

Divide each letter by the total and round. Write both: "3 of 6, 50%." A percentage alone hides whether you mean 4 drafts or 40.

Step 4. Find each type's signal

Ask these in order and stop at the first yes:

1. Are more than half of them U's? Signal U. The request already works.
2. Are there more T's than R's? Signal T. The request isn't working.
3. Anything else, including an R and T tie: signal R. AI gets you started, but you're still doing the work.

The first question is stricter than it looks. U can be the biggest letter and still not win. With 3 U's, 2 R's and 2 T's, you redid or tossed more than half of what you got, so the signal is R. When a call is close, R wins because a rewritten instruction is the easiest thing to undo.

Step 5. Pick your first change. Just one

Look only at readable types with an R or T signal. Your first change goes to the one with the most R and T drafts combined, by count, not by percentage. If two tie, more T's wins. Still tied? Whichever comes first on the list in Part 3.

A U type stays out of this, even with plenty of R's and T's by count: that request mostly works. Only when nothing comes out R or T does your first change become saving your biggest U type as a template. And if no type reached five, your first change is to keep counting for another week.

Step 6. Check for a language gap

A gap exists only when the same type has five or more drafts in each language on its own, with a U signal in one and R or T in the other (Part 6). A both-languages read never tells you anything about a gap.

One week, start to finish

Here's the broker's Friday, tally already added up. It's an example we built, not a real client.

What it says:

• Social posts: 3 in Spanish, 2 in English. Neither reaches five alone, but together they do, so they get read as one and come out R. Nothing gets said about language for posts this week.
• First change: among the R and T types, R plus T comes to 4 for English replies, 3 for social posts and 5 for Spanish proposals. So it's proposal or quote in Spanish, a T: stop asking for it that way.
• Language gap: customer replies reach five in each language, U in Spanish and R in English. AI works for this broker in Spanish and makes work for him in English.

Now look at the Spanish replies. By count they have the most R's and T's of anything, six. But he used 9 of the 15 as is, so the signal is U and they're out of the running. The change goes where the request isn't working.

Signal T: stop asking for it that way

If tossing is what you mostly do with a type, a slightly better prompt rarely saves it. Pick one of three ways out:

1. Stop asking for it. If you throw it out every time, asking just adds the time it takes to read.
2. Use a different tool, one that has what the job needs: your documents, your data, your history with that customer.
3. Do it yourself, especially when that piece of writing is where your business stands apart.

The broker's Spanish proposals came out T. What sells a charter proposal is what he knows about that boat and that client, and none of it was in any request. His Monday change: write the next proposal himself.

A T doesn't mean AI is useless. It means that request, asked that way, isn't working for you.

Signal R: rewrite the instruction

If you keep redoing a type, odds are you make the same fix every time. That fix is what's missing from the request.

On Friday, ask yourself one thing for each R type: what do I almost always change? Not sure? Put two or three of those drafts next to what you actually sent. It's usually one of these five. In each example, the old request was just the first sentence; the rest is the fix the owner kept making, now written in. Each one comes in English and Spanish, for anyone working in both.

1. The tone. A fine-jewelry boutique.

In English: Write a reply to a client asking whether we can resize a ring she bought from us. Warm and straightforward, the way the owner would talk to her in the store. No exclamation points, and none of these words: "exclusive," "luxury," "elevated." Three sentences, max.

In Spanish: Escribe una respuesta para una clienta que pregunta si podemos agrandar un anillo que compró con nosotros. Tono cálido y directo, como le hablaría la dueña en la tienda. Trátala de usted. Sin signos de exclamación ni palabras como "exclusivo", "lujo" o "experiencia única". Tres frases como máximo.

2. The length. A keynote speaker.

In English: Write the program blurb for my talk on making decisions in a crisis. 60 words max, one paragraph; it's printed next to twenty other sessions. Leave out my credentials. Just say what the audience walks out with.

In Spanish: Escribe la descripción de mi charla sobre cómo decidir en una crisis, para el programa de un congreso. Máximo 60 palabras, un párrafo: va impresa junto a otras veinte charlas. Sin mis logros. Di solo qué se lleva el público.

3. A fact AI doesn't have. A yacht charter broker. Whatever AI doesn't know about your business, it fills in with something generic, or something untrue.

In English: Write a post promoting a yacht charter in November. Use these details exactly and don't add any: [length], room for [number] guests, departs from [marina], captain and crew included. If you need a detail that isn't here, leave a blank in brackets. Don't make it up.

In Spanish: Escribe una publicación para promocionar el alquiler de un yate en noviembre. Usa solo estos datos, tal cual: [eslora], capacidad para [número] personas, sale desde [marina], incluye capitán y tripulación. Si te falta un dato, deja un espacio entre corchetes. No lo inventes.

4. The order. A marine upholstery shop.

In English: Write the text for a quote to reupholster a 28-foot boat, in this order: what we'll do, in one sentence; what's included and what isn't; how many days we need the boat; what we need from the owner to get started. Don't write any price. I'll add that.

In Spanish: Escribe el texto de una cotización para cambiar la tapicería de una lancha de 28 pies, en este orden: qué vamos a hacer, en una frase; qué incluye y qué no; cuántos días necesitamos el barco; qué necesitamos del cliente para empezar. No escribas ningún precio.

5. It sounds translated. A private chef who cooks on board. Asking for "a better translation" won't fix it. Write the instruction in the language the draft should come out in, saying what needs to be said, instead of handing over text in the other language to convert.

The old request: Translate this email into Spanish for a couple who asked about a private dinner for twelve on their yacht.

When the Spanish sounds translated: Escribe un correo para una pareja que preguntó por una cena privada para doce personas a bordo de su yate. Escríbelo como lo haría un chef privado de Miami que creció hablando español, no como una traducción. Di tres cosas: podemos hacer la fecha que pidieron, necesitamos las alergias de los invitados antes del [fecha] y la degustación es en [lugar]. Menos de 120 palabras.

When the English sounds translated: Write an email to a couple who asked about a private dinner for twelve on board their yacht. Write it the way a Miami private chef who grew up speaking English would, not as a translation. Cover three things: we can do the date they asked for, we need guest allergies by [date], and the tasting is at [place]. Keep it under 120 words.

Changing something else? Put it in one sentence, the way you'd brief someone new on your team: "never promise a date," "always end with a question." If the fix is different every time, there's no instruction to write yet. Keep counting.

One fix at a time. Change the tone and the length on the same Monday, and Friday can't tell you which one moved the tally.

Signal U: template candidate

If you use more than half of a type as is, the request already works. That's the only thing worth saving as a template, and, down the road, automating.

A template is the instruction that earned those U's, saved word for word, with whatever changes each time in brackets. For the broker's Spanish replies, his saved instruction is in Spanish and boils down to "Write the booking confirmation for [client], departing [day] at [time] from [marina]..." He keeps it in a separate note and reuses it.

Two cautions:

• A template stays in the tally. If that type stops coming out U, it isn't a template anymore.
• Automating is a separate step. Only automate what already works by hand, because automating a process that doesn't work yet just makes it fail faster. And before anything goes out without your eyes on it, decide what an assistant can do on its own and what needs your sign-off.

If you work in English and Spanish, each language gets its own read. Lump them together and they average out, hiding the very thing you're looking for: AI pulling its weight in one language and making work for you in the other. The Step 2 exception gives you a signal when neither language reaches five. It never tells you anything about language.

When the rule in Step 6 holds, the read is:

For [type], AI works for you in [language A] and makes work for you in [language B]. Write the [language B] instruction in [language B], not translated.

• Don't ask AI to translate your good text. Ask for the language B version directly, with an instruction written in that language, like example 5 in Part 5.
• If you don't write comfortably in language B, have someone who does write or check that instruction once. Then save it and reuse it.
• If the rule doesn't hold, there's no gap to fix, even if one language "feels" weaker.

• Quality. A U only says you accepted it. Publish mediocre work untouched, and your tally looks great while your business doesn't.
• Time. It won't tell you how many hours AI saved you or cost you.
• Sales, whether a customer wrote back, or whether a post performed.
• How you stack up. There's no "good" share of U's. We don't know of a study that measures it for a service business, and we're not going to invent one.

What it does show you is where AI is still leaving you the work. The rest can wait until you know where to look.

Monday, you make one change: your first change. Then you count exactly as before, same types, same names.

The next Friday, check that type against your own tally from the week before. Fewer R's and T's means the change moved the tally. No difference? Try something else in that same type, again just one thing.

Types that didn't reach five keep adding up, and some won't be ready until week two or three. Your tally only gets compared to your tally.

Fill-in

This week's note

One line per draft, as it happens:

day - type - language - letter

Types: customer reply - social post - email or newsletter - website copy - proposal or quote - summary or notes - other

Letters: U used as is - R redone - T thrown out

Friday's read

Copy this line once for each type and language you used:

Type: ______________ - Language: ____ - U ____ - R ____ - T ____ - Total ____ - Signal: ____ (or "not ready yet")

More than half U's: signal U. Otherwise, more T's than R's: signal T. Otherwise, signal R, ties included.

Your first change

Type and language: ______________________ (among the R and T types, the most R's and T's combined, by count)

What I'll do: T, stop asking for it that way / R, rewrite the instruction / nothing came out R or T: I'm saving ______________ as a template / no type reached five: I keep counting

If it's R, what I almost always change: the tone / the length / a fact AI doesn't have / the order / it sounds translated / something else: ______________

My new instruction: ____________________________________________

Language gap? No / Yes, in ______________: I'll write the ______ instruction in ______

Next read: Friday ____________

What this guide doesn't cover

Which AI tool to use, or what to give it before it starts. If a type keeps coming back wrong, sometimes the fix is in what went in. That's in the real reasons your customers buy.

The math. That's what The Weekly AI Tally is for, the tool that comes with this guide: enter how many U's, R's and T's each type got in each language, and it gives back your percentages, each type's signal, your first change and, if there is one, the language gap.

Last thing

Short on time this week? Do just this: the next time AI hands you something, write one line. Day, type, language, letter. Another one tomorrow. By Friday you'll have somewhere to start.

Subtracting isn't using AI less. It's no longer asking it for the things you always end up doing yourself.

Want help reading your tally? Bring it to a free call and we'll pick your first change with you.

Book a free call

The AI Edge comes out Fridays at 4pm ET: one AI idea you can put to work that same week. Sign up here.