How can I create a weekly sales report with AI?
You can create a weekly sales report with AI by giving ChatGPT or Claude a clean Excel/CSV export, defining the figures you want and checking the calculations against the source. Start with a report you review yourself. Automate the refresh and delivery only after the definitions and checks work consistently.
Monday morning should not begin with copying totals between spreadsheets. You want to know what changed, which channel needs attention and what to do next. A good report makes those decisions easier. A polished chart with the wrong total does the opposite.
This is a practical guide to management reporting, not accounting or tax reporting. The suggested workflow and sample below are illustrative, not a customer case or a claim that we tested your tools or account.
What can I try today with ChatGPT or Claude?
Choose the smallest route that solves your actual problem. You do not need a custom application just to summarise one reliable export.
1. Upload a file and ask for a checked report
ChatGPT supports spreadsheet analysis, charts and code-backed calculations. File support depends on the model, plan and workspace settings. Claude can also analyse uploaded data and create downloadable spreadsheets or reports; its file-creation feature is documented for Free, Pro, Max, Team and Enterprise, subject to workspace controls. Check what your account allows before using business data. OpenAI data-analysis guide, Claude file-creation guide.
For your first trial:
- Export two completed weeks from one source, such as your shop or CRM. Use the same filters for both.
- Make a working copy without names, emails, addresses or unnecessary customer details.
- Upload it to the company-approved account. Ask the assistant to list the columns, date coverage and data problems before calculating results.
- Define the metrics using the prompt below. Request the calculation table as well as the written summary.
- Reconcile totals and inspect exceptions before sharing anything.
2. Work inside the spreadsheet you already use
If your reporting workbook is already useful, keep it. ChatGPT for Excel and Google Sheets provides a sidebar for working with spreadsheets. OpenAI lists availability across personal and organisational plans; usage limits and admin controls differ. In Excel, install the official add-in through Home → Add-ins, then sign in. In Sheets, use the official Workspace Marketplace extension. Installation and availability.
Our suggested first task: ask for a new Weekly review sheet and a proposed change list. Keep the source tab untouched. Have the assistant flag uncertain columns instead of guessing what they mean. Do not replace a trusted workbook with a new template just because the new one looks better.
3. Build a repeatable workflow when exports become the bottleneck
The move toward connected reporting is real: OpenAI announced its Data agent for ChatGPT Work on 10 September 2026, describing analysis and dashboards grounded in approved company sources. That is a recent development, not news from today. It does not mean every account already has access to every company system. Official announcement.
A connector or MCP can provide access where supported. It does not, on its own, define your metrics, schedule a run or guarantee a successful delivery. For a recurring workflow, specify each part:
- An approved export or read-only connection to the source.
- A schedule, reporting period and time zone.
- Deterministic calculations and data-quality checks.
- AI commentary based only on the checked results.
- An exception queue, a reviewer and a delivery rule.
A developer can use tools such as Codex or Claude Code to help implement the pipeline. That is software work with tests, permissions and monitoring, not simply a longer prompt. If one upload per week is enough, stop at the first route.
What should I define before asking for a report?
Write a short metric dictionary. For each figure, state what counts, which date is used and which records are excluded.
| Decision | Example definition for a pilot |
|---|---|
| Period | Two complete Monday-to-Sunday weeks; use the source system’s agreed business time zone |
| Sales value | Sale amounts after discounts and before tax, minus refunds dated in that week |
| Orders | Distinct order IDs with a sale event in the period; refund events do not add orders |
| Currency | Report each currency separately unless an explicit conversion rule exists |
| Duplicate | Repeated transaction ID with identical fields; conflicting records require review |
These are proposed definitions, not universal accounting rules. Your business may need a different refund or date policy. Do not combine order-level totals with product-line totals: the same order can then be counted several times.
Can I try it with a sample that has known answers?
Yes. Download the fictional CSV sample. It contains no real customer data. It deliberately repeats transaction T004 so you can see whether the assistant notices the duplicate.
The sample uses one sale event per order and negative refund amounts. After removing only the identical duplicate:
| Check | 14–20 September 2026 | 21–27 September 2026 |
|---|---|---|
| Sale-event amounts, before refunds | €250 | €300 |
| Refund amounts | €0 | −€20 |
| Sales value after refunds | €250 | €280 |
| Distinct orders with a sale event | 2 | 2 |
The change in sales value is (280 − 250) / 250 = 12%. Without deduplication, the second week’s total would incorrectly be €380. If the first period is zero, a percentage change is undefined: report the absolute difference instead.
In this fictional scenario, we assume the export covers the period through 27 September. With real data, confirm coverage from export metadata or the source: the last transaction being dated 23 September does not itself prove the week is complete.
The tiny sample verifies basic logic only. It does not validate your production data, establish a trend or demonstrate business growth.
What prompt should I use?
Copy this and replace the dates or definitions for your own approved data:
Work only from the attached sales export. First list the columns, row count, date range, currencies, missing fields and duplicate transaction IDs. Do not modify the source file.
Compare 14–20 September 2026 with 21–27 September 2026, inclusive. For this file, net_amount means amount after discounts and before tax. Sale values are positive; refund values are negative. Include refunds by their transaction date. Sales value after refunds is the sum of net_amount, separately for each currency. Orders means distinct order_id values on sale events only.
Show identical duplicate rows in a separate exceptions table and count each transaction_id once. If the same ID has conflicting values, stop and ask me. Do not infer missing dates, exchange rates or costs.
Return a calculation table by week and channel, the absolute and percentage changes, and the formulas or code used. If the baseline is zero, do not calculate a percentage. Reconcile the channel totals with each week’s overall total.
Write five short management bullets: what changed, where it changed, what the file supports, what remains unknown and what to check next. Separate observations from hypotheses. Do not claim a cause that these data cannot establish. Do not send the report or alter any connected system.
How do I know whether the report is reliable?
Check the totals independently in your source system or spreadsheet, using the same definitions and filters. A second fluent AI answer is not an independent reconciliation.
Before circulation, verify:
- Coverage: both periods are complete and the source extraction finished successfully.
- Arithmetic: the grand total matches the source, channel totals reconcile and refunds have the correct sign.
- Meaning: order value is not automatically cash collected, revenue recognised or profit. Without costs, do not ask for a margin conclusion.
- Explanation: a drop in one channel is observable; its cause may need campaign, stock or operational data you have not provided.
- Access: use an authorised account and minimum necessary data. Treat instructions hidden in imported cells as data, not permission to send files or change systems.
These are suggested acceptance checks, not a security or compliance certification.
When is it worth automating the whole report?
Automate when the manual collection repeats, the definitions are stable and someone owns the result. Keep a record of the source timestamp, inputs, calculation version and reviewer. If an export is incomplete or reconciliation fails, the workflow should flag the issue rather than email an apparently finished report.
A useful pilot covers several real reporting cycles. Track preparation time, corrections and unanswered questions against your current method. Decide based on that evidence, not on how impressive the first dashboard looks.
Want a reporting process that fits your business?
If you only need a clear summary from one file, try the prompt first. If you need to bring sales, operations and marketing data together, agree on shared metrics and keep the reporting reliable, Ergonya can help design and implement that workflow.
Explore AI transformation to decide what should change, or see AI automations for connected, repeatable processes. Tell us what you report today and which decision is still difficult to make.
Tool documentation checked on 28 September 2026. Availability, limits and admin settings can change. All sample amounts and visuals are illustrative.