The weekly report is one document with two jobs: it is the owner's Friday read of where the money and the work stand, and it is the only artefact a part-time bookkeeper or a lead hand can pick up on Monday without a briefing. You build it by exporting three or four files every Friday, pasting them into a fixed prompt that fills a fixed template, and checking the three numbers that matter against the source before you send it to anyone. The AI does the assembly and the plain-English summary; the numbers come from the exports and are pasted, not reasoned. Done that way, it is the task in the everyday list of what AI can take over where the input is already structured, the output is a short page, and the failure mode is a wrong number, which you catch at the check.
What the report is for
The first use is the owner's Friday read. A few minutes, coffee in hand, before the weekend. What is in the bank. What is owed and how old each item is. What is in flight and where each job sits. What is committed next week. Anything that broke.
It also does the handoff. A part-time bookkeeper does not need a meeting; they need the report. A lead hand does not need to ask "where are we with job 0412?"; the report says it. You can chase the overdue invoices on their own schedule without the owner sitting at the keyboard on a Sunday night, and the person who runs the jobs can see what is coming without a meeting notes action list being the only written record of what was agreed on Thursday.
What goes in
Five lines, in order:
- Cash in the bank, as of Friday close.
- Money owed to you, sorted by how old each item is.
- Jobs in flight, one line each, with the current stage.
- Next week's commitments, dates and deadlines.
- One line on anything that broke: a missed delivery, a client complaint, a tool that stopped working.
The last line is the one people skip and then wonder why the report felt like a spreadsheet. One sentence, in plain English. "The Xero export failed twice this week; we are on manual entry for September." That is the line a bookkeeper needs, and it is the line an AI can write if the export tells it what failed.
How AI assembles it
The loop runs once a week and, once the prompt is settled, takes less time than the coffee. You export, you paste, the AI fills the template, you check three numbers, you send.
Exhibit 1
The weekly loop is five steps, and the owner's only hands-on work is the check.
The three or four exports are the same every Friday. From QuickBooks Online you pull a bank report and an accounts-receivable list: Reports, then Standard reports, then Export/Print, then Export to Excel. From Xero the same two reports come out as Excel or PDF. From the job board, whether that is Trello, Notion, or a spreadsheet, you export the active items. Three or four files, each one small, each one in a format the assistant can read. Claude reads CSV and spreadsheet files and runs Python over them on every plan once code execution is switched on; ChatGPT reads .xlsx and .csv files and runs Python in a notebook environment. Either one will handle the files.
You paste them into a fixed prompt. The prompt is written once, saved in a project, and changed only when the template changes. It says: "Here are the bank report, the invoice list, and the job board for the week ending Friday. Fill the weekly report template. Use only the numbers in the files. Where a figure is missing, write 'not in the export.' Do not estimate."
The template is a fixed set of numbered rows. The AI's job is to fill them from the data and to write the plain-English summary for anything that broke. It does not invent a number and it does not round one. The bank report's figure goes in the row to the cent. If a job stage is missing from the export, the row says "not in the export," and you fix it before the report leaves your hands.
Exhibit 2
The template has five fixed rows, and the AI fills them only from the pasted exports.
You check three numbers against the source: the bank balance, the total owed, and the number of active jobs. If one is wrong, you find the row, you correct it, and the rest of the report is the AI's. The assembly and the prose are the AI's. The arithmetic is yours.
Which plan tier the exports may go into
The files you paste in carry client names, invoice numbers, and job references. That is real data, and where it goes is a plan decision before it is a privacy decision.
On the business tiers the position is the same at both vendors. Anthropic states that "by default, we will not use your inputs or outputs from our commercial products" to train its models, and its Team and Enterprise plans sit under that policy. OpenAI says that "by default, we do not train on any inputs or outputs from our products for business users, including ChatGPT Business, ChatGPT Enterprise, and the API." The business tier is the tier that lets you answer the question "will my client's invoice number end up in a training set?" with a written no.
On the consumer tiers the default differs by vendor. OpenAI's consumer plans improve by training on conversations unless you opt out. Anthropic's consumer plans carry a Model Improvement control under Privacy Settings, and the choice is yours to make. If you are the only user and the data stays in a chat that only you can see, the consumer tier is a choice you can make with your eyes open. If a bookkeeper or a lead hand is also reading the report, the business tier is the one that lets you say the same sentence to two people without a caveat.
If you would rather not send the exports to a cloud assistant at all, the question of which parts of the work can stay on your own machine is the one to answer before you pick a plan, and for a file as small as a weekly export it is a question you can probably answer with "the cloud is fine, the data is small, the business tier covers it."
What to strip is simpler than the plan question. A customer's name becomes a job number. The invoice number stays; the customer name goes. The job board export is the file to clean, because the other two come from your own accounting software and carry the names your bookkeeper already sees every week.
Exhibit 3
Job numbers, dates and amounts go into the prompt, and customer names stay in your files.
What the report is not
It is a snapshot, taken on Friday, and it does not update at 9 a.m. Monday when someone pays an invoice. If you need live numbers, the accounting software is the live document, and the report is the digest you write from it once a week. A dashboard answers "what is it right now"; the report answers "where did the week land, and what did we commit to next."
It does not replace the monthly bookkeeping rhythm. The Friday number is a snapshot. The month-end close is where the bank and the ledger get reconciled. A report can tell you the two disagree. It cannot tell you which missing transaction is the one you missed. That is the bookkeeper's job, or your job on a quiet Tuesday, and no amount of plain-English summary closes the gap.
The first four weeks
Week one: you do the exports by hand, you write the prompt, you paste, and the output is wrong in two or three places. The numbers are right; the plain-English line is clumsy, and one job is at the wrong stage because the job board export was from Thursday. You fix it. You save the corrected prompt.
Week two: the output is right in four of five places. The one that is wrong is the same kind of thing: a job stage that the AI read from the wrong column. You add a line to the prompt: "Job stages come from column D, not column F." You save.
Week three: the report is right all five times. You check the three numbers. You send it.
Week four: you realise you did not open the job board until the AI told you a job was at risk. The report caught something you would have caught on Saturday, a day later. The loop is working, and the check has become a read.
After week four you stop doing it by hand, because the export, the paste, the fill, the check and the send are a sequence you can hand off. On the week you are on a job site, the bank report is the thing you cannot do standing up in a customer's hallway. You hand it off.
The part of this that we work on is the template, the prompt, and the habit: getting the loop to the point where the owner checks three numbers and walks away. That is what AI enablement covers for a small firm.