Automate Your Weekly Business Review With AI: A Solo Template

Every guide to an AI weekly business review assumes a team: a scorecard with owners, a Monday meeting, a Slack channel where the report lands. If you run the business alone, there is no meeting, and the review only matters if it changes what you do next week.

This article gives you a paste-in weekly review built for one person: four prompts run in order, an input block, and a fixed output of three numbers, what changed, one decision, and next week’s calendar blocks. It also gives you the export path for each input, taken from Google’s, Stripe’s and PayPal’s help pages on October 5, 2026, and the rules for what the AI may and may not decide.

What a weekly review is for when the business is one person

A team review exists so people agree on numbers and assign owners. You have no one to agree with. Every owner column says your name.

So the solo version has a narrower job. It answers one question: given what happened last week, where do next week’s hours go?

I run several online businesses alone, on Claude, GPT, Hermes and Google Calendar. The hard part of my week is rarely a bad number. It is several businesses wanting the same Tuesday, and a calendar that shows I spent most of last week on the one that moved least. That is why the template below treats last week’s calendar as data, next to revenue and traffic.

The output is fixed at four parts:

1. Three numbers, the same three every week.
2. What changed, at most three lines, each marked fact or guess.
3. One decision, proposed by the AI and confirmed by you.
4. Next week’s calendar blocks, sized to the free hours you actually have.

If a metric cannot change a decision this week, leave it out of the review. Watch it monthly instead.

This is the backward-looking half of the week. The forward-looking half, planning the coming week from a calendar export, has its own template in using Claude for business operations. Run this review first and feed its decision into that plan.

The four inputs and where to export them (checked October 5, 2026)

A solo review needs four inputs: revenue, one demand number from analytics, last week’s calendar, and your open task list. The task list is whatever you already keep, pasted as text. The other three come out of tools whose export buttons are easy to lose.

Export paths and limits, collected from each vendor’s help page on October 5, 2026:

Three details in that table change how you run the review.

First, timing. Stripe’s Dashboard reports include only complete days, GA4’s daily numbers are still settling for a day or two, and Search Console marks its newest data as preliminary. A review at 8 a.m. Monday reads a short Sunday. Run it Monday after lunch, covering Monday through Sunday, and use the same window every week.

Second, people’s data. Stripe’s itemized report offers customer name and email as optional columns, off by default. Keep them off. For PayPal or any export where you cannot pick columns, delete name and email columns before the file goes anywhere near a chat window. The review needs totals per business, not buyers. If you sell through both processors, the fee differences are covered in Stripe vs PayPal for digital products; for this review, net is the number that counts.

Third, the calendar. If you use a personal Gmail account, Time Insights is not available to you, so the hours-per-business number has to come from the events themselves. Put the business name at the start of every event title. That one habit is what makes the calendar countable.

The weekly review template (copy the whole thing)

The template has three parts: a standing context block you write once, an inputs block you fill every Monday, and four prompts you run in order. Do not merge the prompts into one. The point of the chain is that you check the numbers after step 1, before the AI is allowed to explain anything.

Before you start, check one setting. Claude reads CSV uploads, but XLSX files need code execution and file creation switched on, which lives under Settings, then Capabilities on Free, Pro and Max plans. ChatGPT accepts .csv, .xls and .xlsx and runs Python on them. CSV is the safer default for both.

Standing context (write once, paste at the top or into Project instructions)

WEEKLY REVIEW: STANDING CONTEXT
I run [number] online businesses alone:
- [Business A]: [what it sells]; revenue arrives through [Stripe / PayPal / other]
- [Business B]: [what it sells]; revenue arrives through [...]

The three numbers I review every week:
1. Net revenue per business (gross minus fees and refunds), in [currency]
2. One demand number: [Search Console clicks / GA4 sessions on the sales page / signups]
3. Hours I spent on each business, counted from my calendar

Review window: Monday 00:00 to Sunday 23:59, [time zone].
Calendar events start with the business name, e.g. "[Business A] - outreach".

Rules for every answer in this review:
- Use only numbers that appear in what I paste. If one is missing, write MISSING. Never estimate it.
- Label every explanation FACT (visible in the data) or GUESS (your inference).
- You may propose. You may not decide. End every decision line with "for me to confirm".

Inputs (fill every Monday)

INPUTS FOR [dates]
REVENUE THIS WEEK: [Stripe Balance summary activity CSV and/or PayPal activity CSV, customer columns removed]
REVENUE LAST WEEK: [the three numbers from last week's review]
DEMAND: [this week's and last week's export rows or totals]
LAST WEEK'S CALENDAR: [events with day, start, end, title, or the .ics contents]
OPEN TASKS: [per business, with any real deadline]
NEXT WEEK'S FIXED EVENTS: [calls, travel, deadlines already booked]
OPEN DECISION LOG LINES: [any earlier decisions whose review date has arrived]

Prompt 1: read it back before you analyze

Step 1 of 4. Do not analyze yet.
1. Restate every total you read: gross, fees, refunds and net per business; the demand number for both weeks; hours per business.
2. For hours, count only events inside the review window and assign each one by the business name at the start of its title.
3. Show each total as a sum I can check against the file.
4. List anything MISSING, any row you could not read, and any event you could not assign.
Stop here and wait for me to confirm or correct.

This is the step most reporting prompts skip. Compare the net figure with the total line in the export itself before you type “confirmed”. If the AI misread a column, you find out now, not after it has built a story on the wrong number.

Prompt 2: what changed

Step 2 of 4. Use only the totals I confirmed.
1. Put this week next to last week for the three numbers, as an amount and a percent.
2. Name at most three changes worth my attention. For each, the evidence (FACT) and one possible cause (GUESS).
3. Compare where my hours went with where the revenue and demand came from. Say it plainly if one business took most of my time and little of the result, or the reverse.
4. If nothing material changed, say "nothing material changed" and stop. Do not fill the section.

Item 4 matters more than it looks. A model asked “what changed” will find something. Giving it permission to say nothing is how you keep a quiet week from producing a fake crisis.

Prompt 3: one decision

Step 3 of 4. Propose the single decision these numbers most support. Not a list of ideas.
- The decision, in one line.
- Two alternatives, one of them "change nothing for another week".
- For each: hours it costs next week, money it costs, and whether I can undo it within a week.
- The result next week that would prove it wrong.
- Then one line for my decision log, in this format:
  [date] | [decision] | [numbers it rests on] | [review on: date] | for me to confirm
If the data cannot support any decision, name the input that would and stop.

Prompt 4: next week’s calendar blocks

Step 4 of 4. Turn the decision I confirmed into next week's calendar.
1. From NEXT WEEK'S FIXED EVENTS, list the free blocks of an hour or longer on working days.
2. Give the decision its blocks first, then open tasks with a real deadline, then nothing else.
3. Write each block as: day | start | end | title beginning with the business name.
4. If the free time cannot hold the decision and the deadlines, say what gets dropped. Do not shrink blocks to make it fit.
Do not create, move or delete calendar events. I will add them.

What the finished review looks like

WEEKLY REVIEW: [dates]
3 NUMBERS: Net revenue [A / B] (last week [A / B]) | [Demand] [x] (last week [y]) | Hours [A / B]
WHAT CHANGED: [up to three lines, each tagged FACT or GUESS]
ONE DECISION: [line], confirmed / changed / rejected on [date]
NEXT WEEK'S BLOCKS: [day | start | end | title]
DECISION LOG: [the line, copied into your log file]

Keep last week’s three numbers in the inputs block, not in the chat’s memory. A fresh conversation every Monday with the numbers pasted in is easier to check than a thread that has quietly accumulated six weeks of context.

What the AI may decide, and what it may not

The review is only safe to act on if the line between proposing and deciding is written down before the first Monday. This checklist is that line.

The AI may:

• Read, total and compare the numbers you pasted.
• Label causes as guesses.
• Propose one decision with alternatives and a kill signal.
• Draft calendar blocks as text, and draft any message you will send yourself.

The AI may not:

• Change a price, a plan, a discount or a product page.
• Issue or refuse a refund, or answer a dispute.
• Send an email, DM or post, including a “quick update”.
• Create, move or delete calendar events without your approval.
• Pay, buy or move money.
• Edit or delete the exports or the decision log.

Before you accept the review, check three things:

• Net revenue matches the total line in the export you downloaded.
• One day of calendar hours matches what you remember doing.
• Every cause in “what changed” carries a FACT or GUESS label.

The vendors’ own documentation draws the same line. OpenAI’s help page tells you to review the code, outputs and assumptions ChatGPT used before relying on an analysis. Anthropic’s Cowork safety page says not to schedule tasks that access sensitive files, send messages on your behalf, make purchases or take other hard-to-undo actions. Claude’s Gmail connector asks for approval before each send by default. Leave that default alone.

There is a quieter risk as well. Anthropic warns that instructions hidden in external files or websites can trick Claude into leaking data when code execution is on. An export you downloaded yourself from Stripe is low risk. A CSV someone emailed you is not, so do not feed a review from files you did not pull.

The decision log is the piece most people skip. Keep it in a sheet or doc you own, one line per confirmed decision, with a review date. When that date arrives, the line goes back into the inputs block, and the review has to say whether the decision worked. Without the log, an AI review becomes a weekly mood report.

How much of this to automate

Automate the collection. Do not automate the decision.

Start fully manual for the first few weeks: download, strip, paste, run four prompts. You learn what your numbers look like and where the AI misreads them, and that is what tells you which parts are safe to hand off.

After that, take the steps in this order:

1. Calendar first. Claude’s Google Calendar connector can view your events and is available to all users, so step 1 can count hours without an export.
2. Revenue second. Stripe lets you schedule the same Balance summary export daily, weekly or monthly, so the file is waiting on Monday instead of you fetching it.
3. Scheduled runs last, if at all. A scheduled task that prepares a draft review for you to read is fine; the setup and limits are in the Claude business operations guide. A scheduled task that sends the review to anyone, or acts on it, crosses the line in the checklist above.

Which assistant runs it matters less than the chain. Claude and ChatGPT both read a CSV; how their plans and limits compare with Gemini’s is in ChatGPT vs Claude vs Gemini for solopreneurs. Claude is where I would run it, because a Project can hold the standing context. The raw payment files, before the customer columns are gone, I would rather keep on Hermes, the open-weights model I use for private numbers; why that split works is in my AI tool stack for solopreneurs.

When the review keeps saying the same thing

Sooner or later the review proposes the same decision three Mondays running. That is not the AI repeating itself. It is the review doing its job: the decision keeps being right and keeps not getting done.

The usual cause shows up in prompt 4. The blocks were proposed, then lost to calls and small tasks during the week. When this happens, change the input, not the prompt. Put the decision’s blocks on the calendar before anything else on Monday, and protect them like a booked call. If outside booking links keep eating those hours, the scheduling tools for solopreneurs comparison covers how to cap what others can grab.

If a decision survives three reviews without getting its hours, drop it on purpose. Write that in the log too. A decision you rejected is useful history; one you keep postponing is noise in every review after it. More workflows built this way are under AI workflows.

This Monday after lunch, download last week’s Stripe Balance summary or PayPal activity, remove the customer columns, and run only Prompt 1. Check its net total against the file before you run anything else.