Your AI hands you a lot. Count how much of it you use

TL;DR

• Google let its new model work alone overnight, for eleven hours, on a Google project. Of six proposed changes, two shipped, and Google presented that as the workflow doing its job.
• The tally is one letter per thing AI hands you: used as is, redone, or thrown out. One week, logged as it happens.
• On Friday you read it by type of task. What you toss most, you stop asking for. What you redo most needs a different instruction. What you use as is more than half the time is the only thing worth turning into a template.
• If you work in English and Spanish, keep two tallies. AI can help in one language and create work in the other, and a single tally hides that.
• You'll finish with your weekly tally: three numbers per task type and per language, plus your first change: something to stop asking AI for, or an instruction to rewrite.

The short answer. To find out whether AI is actually helping your business, give everything it hands you for a week one letter: U (used as is), R (redone) or T (thrown out). Broken down by type of task and by language, that tally tells you what to stop asking for, which instruction to rewrite and what's worth turning into a template.

Monday's weekly roundup ended with one job for the week: notice where each thing AI gives you actually ends up. If you started, you already have a few marks. If you didn't, today is early enough.

1. Google left its AI alone for eleven hours, and the team kept two of six changes

Monday's roundup had the news: on October 13, a three-person Google Cloud team published an account of leaving Gemini 4 Argon, Google's new model, working alone overnight, eleven hours straight, on one of their open-source projects. By morning, six proposed improvements were waiting. Two shipped.

What matters here is how they cut the other four. The write-up gives the reason for each: one could hand back stale data, one leaned on the private internals of an outside library that could break on any update, one made the code harder to read without making anything measurably faster, and one only helped on test runs, not in everyday use. Four no's, each with its reason written down.

Google treated keeping two as proof the process worked

Their own summary: "the agent explores, the maintainer subtracts." And when Google reported on its internal use of the model, it put two numbers side by side: agent-written code submissions up more than 35%, and code rollbacks down by nearly a third. The first says how much got produced. The second says how much of it had to be undone.

Neither number tells you how AI will do in your business. It's Google's code, checked by Google's engineers. What does carry over is the habit of counting what stays.

2. Hours and volume count what AI did, not what you kept

It's easy to measure AI by its output: how many posts, how many emails, how fast it answered, how many hours it "saved."

If your assistant drafts ten posts for the week and you publish one, you don't have ten posts. You have one, plus the time it took to read the other nine and decide they wouldn't work.

A number that doesn't tell you what stayed can't help you decide anything

"It saved me hours" gives you nothing to change. "I redid nine of twelve client emails" tells you exactly where the work is.

3. One letter per draft: used it, redid it, or tossed it

The tally is short on purpose. Whenever AI gives you back something you asked for, add a line to a phone note:

day - what it was - language - letter

The letter is one of three:

• U (used as is). You sent or published it the way it came, or with under a minute of touch-ups: a typo, a name, a date. What it says didn't change.
• R (redone). You kept something, an idea, the order, a paragraph, but it took more than a minute of work or you changed what it says.
• T (thrown out). You used none of it.

One rule for the gray zone: every time you ask for the same thing again, that's a new draft, and each draft gets its own letter for what you did with it. Three tries at an email, and you sent the third one untouched? That's T, T, U. If you lifted one line from the first try, make that one an R.

Here's a Tuesday for someone who runs a boat-detailing business in Fort Lauderdale:

Tue - reply to a customer on WhatsApp - EN - U Tue - weekend Instagram post - EN - R Tue - service description for the website - ES - T

Log it in the moment, or by Friday you'll only remember the wins

A tally from memory isn't worth much. Five seconds per line, right then, without deciding yet what it means.

4. How to read the tally on Friday

At the end of the week, group the lines by type of task (customer replies, posts, emails, website copy, proposals, whatever you actually use it for) and add up the U's, R's and T's for each type.

Two rules before you read anything:

1. A task type with fewer than five drafts doesn't tell you anything yet. Five is a floor we chose, not a statistical rule: below it, each draft moves the shares more than twenty points, and one can flip the whole reading. Keep counting that type next week.
2. Read each type on its own. An average across everything hides the one thing you're looking for: that AI helps a lot with one job and barely at all with another.

Then read each type like this:

• More than half are U's: that already works. It's a template candidate.
• If not, and there are more T's than R's: that request isn't working, and it probably isn't worth repeating the same way.
• If not, and there are more R's than T's, or they're tied: AI gives you a starting point, but you're still doing the work. There's an instruction to rewrite.

The only number you compare yourself to is your own, next week

We haven't seen any study that says what share of "used as is" counts as good for a service business, and we're not going to make one up. Google's two of six is Google's code, not your target. This week's tally is your baseline. Next week's tells you whether the change worked.

5. Subtract on purpose: what to stop asking for, what to rewrite, what to template

The part of Google's story most people skip is what happened before the eleven-hour night. There was a first round of about two hours with no instructions about how the team works. It found real speedups, but it missed several of the conventions the project's maintainers follow. For round two, the team didn't switch models. They wrote down how the team works and gave it to the agent, the way you'd hand an onboarding guide to a new engineer, added two sentences to the request, and turned on a mode that keeps the goal in view for hours.

Almost everything they changed was the instruction. The tally takes you to the same place, and leaves you three moves:

• Where you toss more than you redo, stop asking for it, or take it to a different tool. A speaker who tosses every bio AI writes for them doesn't need a better prompt. They need to write that bio themselves, because the bio is what sets them apart.
• Where you redo as much as you toss, or more, rewrite the instruction. Look at what you changed in this week's R's. If you keep changing the same thing (the tone, the length, a fact the AI doesn't have), that's what's missing from the request.
• What you use as is more than half the time is the only thing worth templating or automating. It already works for you, and automating something that doesn't work yet just makes it fail faster. Before you let anything go out without a look, decide which jobs an assistant can finish alone and which ones you approve first.

Dropping a request is a legitimate way to use AI well

When a tool is always within reach, it feels like you should ask it for everything. You don't have to. Whatever you stop asking for is one less thing to read every week, and that time goes back to the work only you can do well.

6. What we found when we counted ourselves

We ran the tally on our own work. We took the twelve pieces AI drafted for us between Monday, October 12 and Tuesday, October 20: seven blog posts, last Wednesday's guide, interactive tool and email sequence, a prompt pack, and the newsletter. Every piece goes through a separate review before it ships, and what that review finds gets written down.

None of the twelve went out the way AI drafted it. In three, the fixes didn't touch anything central. The other nine we rebuilt somewhere that mattered: the closing exercise, the framework, the headline, or the founders' conversation. We didn't throw out a single piece entirely.

In the tally's letters, that's twelve R's, no U's and no T's. One caveat so the number doesn't mislead: we counted by piece, not by draft, with both languages together, and reconstructed it afterward from what each review wrote down, not line by line as we went. Your tally will be sharper than ours.

What forced the rebuilds was almost never the writing itself, except in the English, which we'll get to. It was what the piece said: an exercise that repeated one we'd already published, a claim our own records didn't back up, an interactive tool that contradicted what the reader had just clicked.

In two languages, the work hides in the second one

We write in Spanish first and English second. In five of the twelve pieces, the review found the English had been written with the Spanish open next to it, line by line or with the same skeleton, and it had to be rewritten or fixed. That finding didn't come up once for the Spanish.

One tally for both languages would have buried that in the average. That's why it's two.

7. What this tally doesn't measure

It doesn't tell you whether what you used as is was any good. A U only means you accepted it. If you're publishing mediocre work without touching it, your tally will look great and your business won't.

It doesn't measure time, sales, or whether a customer wrote back. And it doesn't settle who should review the work; we covered that in the people who built it can't be the only ones grading it. The question here is smaller, and it's yours: which number to look at on Friday.

Last week we looked at what you hand the tool before it starts working. This tally looks at the other end, what comes back. If one task type keeps scoring badly, sometimes the fix is in what went in.

A short, honest tally beats a perfect measurement nobody keeps

Three letters and one week. Everything else you can measure later, once you know where to look.

The what-stays conversation

Todd: Thirty years of managing projects taught me one rule I still use. A deliverable isn't done when it arrives. It's done when somebody signs off on it. Whatever AI hands you that you don't sign off on is still open work, and now you also have to read it first.

Naty: It's the same in every business I've run, and I've built them in more than fourteen industries. What's easy to count is what comes in. What tells you how the business is really doing is what stays. With AI it's very easy to watch only what comes in: how many posts, how many emails, how fast.

Todd: And what stays is the number that improves, as long as somebody looks at it on a schedule. In a project, the figure someone reviews every week gets better. The rest just sits in the report.

Naty: Writing down a T is harder for most people than writing anything else. It feels like a wasted afternoon, or like paying for something useless. But the T is the letter that tells you where to stop spending energy.

The one thing to do this week: your tally

Five seconds per draft, and fifteen minutes on Friday.

1. Start a new note on your phone. If you work in two languages, the language goes on every line.
2. Each time AI gives you back a draft, log the day, what it was, the language and one letter: U, R or T.
3. On Friday, add it up by task type. Set aside any type with fewer than five drafts, and any type where more than half are U's. Of the rest, take the one with the most R's and T's combined and decide one thing for Monday: stop asking for it, or rewrite the instruction.

The guide picks up from there: how to name your task types, how to read a tally that comes out tied, and what to put in a new instruction based on what you keep changing. It comes with the tool that does the math and hands you your numbers and your first change.

Send the word and it lands in that same chat.

Send KEPT

How do I know if AI is actually saving me time? Instead of estimating hours, count what happens to each thing it gives you. For one week, log every draft with a letter: used as is, redone, or thrown out. If almost everything in one type of task is redone or thrown out, AI isn't saving you work there, however fast it answers.

What counts as "used as is"? Anything you sent or published the way it came, or with under a minute of touch-ups: a typo, a name, a date. If it took more than a minute, or you changed what it says, the order, or more than half of it, it's redone. If you asked for another version, each version counts separately.

How many drafts do I need before the tally means anything? At least five per task type, a floor we chose ourselves. If a type doesn't reach five in a week, keep counting it the next.

Is there a good percentage to aim for? No study we know of has measured this for service businesses, and Google's two of six is Google's code, not a target. Compare yourself only to yourself: this week against next, after you change one thing.

I work in English and Spanish. One tally or two? Two, or one with the language on every line. AI can help in one language and create work in the other, and a single tally averages that away. It happened to us: the extra work was in the English.

Wrapping up

The headlines went with the eleven unattended hours. The useful part came the next morning, when the team read six proposals and kept two.

Your version of that fits in a phone note. One letter per draft, one week, and one decision on Friday.

What you keep from AI is the only part that counts as work done.

One more thing: whichever task type surprises you most on Friday is the one we'd like to hear about. Tell us, and we'll both read it.

Keep reading

• When AI acts on its own, who has to answer for it? - this week's roundup, where the tally started
• Automating a broken process does not fix it, it speeds it up - why you only automate what already works by hand
• Give a program a job it can't finish, and no way to say so, and it will find another way out - the three decisions before you switch an assistant on
• SEO vs. GEO vs. AEO: which one does your business need in 2026? - the full map
• Optimizing a 200,000-line Python codebase overnight with Gemini 4 Argon, Google Developers Blog - the full write-up

Sources

• Optimizing a 200,000-line Python codebase overnight with Gemini 4 Argon, Google Developers Blog, George Weale, Bo Yang and Andrew Biernat, October 13, 2026 - https://developers.googleblog.com/optimizing-a-200000-line-python-codebase-overnight-with-gemini-4-argon
• Gemini 4 Argon: Built to solve your hardest problems, Google, Koray Kavukcuoglu, October 13, 2026 - https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/