Automate Weekly Reports with AI: From Raw Data to a Ready Draft

Every Monday, the same ritual: exports from the CRM, the ad account, a couple of spreadsheets, two hours of copy-paste and rewording, and by lunchtime a report your manager will skim in three minutes. AI can take almost all of that work off your plate, but the numbers stay your responsibility: you will still verify them by hand, and that is how it should be.

Let's be honest about the cost up front. The first setup takes an evening: building a template, describing the tone, mapping your data sources. After that, the weekly report takes about ten minutes, verification included. Here is the full walkthrough, using Claude as the example.

What you need

• Access to Claude: the web version or Claude Desktop, either works.
• Two or three of your past reports, ideally ones your manager accepted without edits.
• A clear picture of where your data comes from: CRM, ad platforms, spreadsheets, anything with an export button.

No coding required. Everything here is done with plain text.

Step 1: Build a project around the report (that one evening)

A project in Claude is a workspace with persistent context: the instructions and files you put into it apply to every new chat inside that project. Set it up once, then just use it. We covered how projects and context work in a separate post.

What goes into the project:

1. The report template. Take your best past report and strip out the numbers, keeping the skeleton: sections, their order, the recurring tables. Claude will assemble every future report on that skeleton.
2. Tone and rules. How your company writes: dry or with takeaways, which term you use for revenue, how many decimal places for percentages. The more specific this is, the fewer edits later.
3. A data source map. A short file listing which exports arrive, from which system, what columns they contain, and what each column means. The model should never have to guess what "amt_net" stands for.

A sample project instruction:

`` You prepare a weekly report for the sales director. Structure: strictly follow report-template.md. Tone: short, factual, no intro paragraphs, no opinions. Take numbers only from the attached exports. If data for a section is missing, write "no data". Never estimate and never reconstruct trends against past periods. Output: final text ready to paste into an email. ``

The two lines about missing data are the most important part of the whole instruction. I will explain why below.

Step 2: The weekly cycle (those few minutes)

From then on, the routine looks like this:

1. Export the week's data: a CSV from the CRM, the ad account report, the sales spreadsheet.
2. Open a new chat inside the project and attach the files.
3. Write a short prompt, for example: "Build the weekly report for Jul 27 to Aug 2 using the template. Attached: the sales export and the ads export. No refunds data this week."
4. Read the draft as an editor, not as a reader.
5. Cross-check the key numbers against the exports by hand.
6. Paste into an email or document and send.

Never skip steps 4 and 5, even when a month in it feels like the system runs perfectly. That is exactly the week it will slip.

What to verify by hand: the checklist

AI is excellent at structure and wording, but looser with numbers than you would like. Before sending, check:

• Totals. Open the export and compare against the summary row, or add up the key line items yourself.
• Trend percentages. Recalculate at least two or three: week-over-week growth, plan completion.
• Every "up" and "down". Make sure the direction matches the data, not the expectation.
• Names. Clients, managers, campaigns: the model can quietly substitute a similar name from the template for the real one in the export.

It sounds tedious, but it takes three minutes and protects you from the worst-case scenario: your manager finds a wrong number first, and trust in every future report dies with it.

Common failure modes

• Rounding. The model likes clean numbers: 19.7% quietly becomes "about 20%". In a management report that is a defect, so spell out the required precision in the project instruction.
• Invented trends. The most dangerous one. If the export has no data for the previous period, the model may "estimate" a comparison, and it will look convincing. That is why "if data is missing, say so" must be written into the instruction explicitly.
• Mixed periods. Drop in two exports from different weeks without labels and you get a report where numbers from different periods are added together. Name your files clearly: sales-2026-w31.csv, not "export (3).csv".
• A vague template. If the template says "a section about sales", the model writes filler. If it says "a table: plan, actual, deviation in percent", you get a table.

All four have the same cure: specifics in the template and hard rules in the instruction. Every mistake you catch makes the instruction stricter and the reports cleaner.

Next level: put it on a schedule

Once the manual cycle is stable and you have not touched the instruction for a couple of weeks, you can automate it: exports land in an agreed folder, a scheduled scenario builds the draft, and it is waiting in your inbox on Monday morning. Your role shrinks to checking the numbers and hitting send. We covered building scenarios like that in a post on automating routines with Claude.

One piece of advice: do not start with automation. Run two or three manual cycles first, until the template and the instruction settle. Automating a raw process just means getting a bad report faster.

The short version

One evening of setup, then minutes per week. A project with a template, tone rules, and a data source map; a weekly chat with fresh exports; a hand-check of the numbers before sending. The failure modes are predictable: rounding, invented trends, mixed periods, a vague template, and every one of them is fixed by a stricter instruction and a checklist.

If you would rather walk this path with a safety net, AGINE Academy teaches it hands-on: you play as a robot named UNIT and complete missions in the real Claude, not a simulation, leaving each lesson with a configured tool. An AI mentor tied to the lesson helps when you get stuck. The starter block of four lessons is open for free, no signup.