Use two complete, comparable date windows, reconcile the source exports, and prepare a short report with observed changes, unresolved questions, and a next action. Start manually in Grist; add automation after the checks work.
In this guide
- Prepare two complete reporting windows
- Import and inspect the file in Grist
- Check the totals and the limits of the comparison
- Give AI the checked figures and ask for a bounded draft
- Save the repeatable steps before adding automation
- What should you do when the figures disagree?
- Run your next update from the kit
Prepare two complete reporting windows
This workflow is for a small-business operator preparing a weekly update. Record the timezone, reporting dates, currency, source system, and meaning of each metric. Keep the exports unchanged beside your working file.
The practice data is fictional. It covers October 5–11 and October 12–18, 2026, in America/Los_Angeles. It is not ToolFlowKit analytics or a merchant report. Ad spend is blank throughout and must stay unknown.
Import and inspect the file in Grist
- Open your Grist workspace and create a document by importing weekly-business-sample.csv. Follow the official import guide for your chosen hosted or self-hosted setup.
- Check that 14 dates arrived once each, the timezone is consistent, and the numeric columns have the intended types. Preserve the blank ad_spend_usd cells.
- Separate the two seven-day periods and total sessions, orders, gross sales, and refunds for each. Use the column/formula helper if adding net sales: gross sales minus refunds.
- Compare the working totals with the reference report before asking AI to explain anything.
Check the totals and the limits of the comparison
- Compare each date and source total, not only the final grand total.
- Calculate percentage change as (current − previous) ÷ previous. A previous value of zero needs a different explanation, not an infinite percentage.
- Keep refunds visible even when gross sales rise.
- Separate missing data from an observed zero.
Recorded sessions: 800 → 960
+20%. This is a reported traffic total, not proof that a particular channel produced the orders.
Reported orders: 32 → 40
+25%. Gross sales go from USD 1,600 to 2,000. Refunds go from USD 50 to 100.
Net sales: USD 1,550 → 1,900
Gross minus refunds, +22.58% after rounding. More net sales does not explain why the increase happened.
Advertising performance: unknown
No ad-spend export or channel-level order allocation was supplied. Keep ad return and channel conversion conclusions out of the report.
Give AI the checked figures and ask for a bounded draft
Paste the approved totals, dates, metric definitions, and missing fields into your own AI service. Keep customer-level and order-level personal information out of the prompt.
TASK: Draft a weekly business update from these checked aggregate figures. WINDOWS AND TIMEZONE: [Complete current and previous dates] SOURCES AND DEFINITIONS: [Where each metric came from] CHECKED TOTALS: [Paste approved table] MISSING DATA: [Fields unavailable; keep them unknown] Return: observed changes; checks or unresolved discrepancies; up to two plausible explanations clearly labeled as hypotheses; the evidence needed to test them; one or two practical next actions. Check percentage calculations. Do not attribute orders to a channel or query without supporting attribution data. Do not treat missing values as zero or publish the report.
Save the repeatable steps before adding automation
Keep a dated input folder, field definitions, the checked totals, prompt, draft, and reviewer decision. On the next run, replace the inputs and check that headers, dates, and source definitions still match.
If using Activepieces, begin with a test flow: choose a trigger, read sample data through the relevant connection, validate required fields, and prepare a draft. Confirm which connectors and review features your account supports. Add an external sharing action only after the review process is explicit.
What should you do when the figures disagree?
- A person approves the source totals and the final draft.
- The report identifies what is known, unknown, and inferred.
- The next action requests the missing evidence rather than guessing a cause.
The two exports use different date boundaries
Align complete days and timezones before comparing them. A partial current week stays provisional.
Rows are missing or repeated
Check the source export, filters, pagination, and date/key uniqueness before editing the report's explanation.
Sessions and orders seem inconsistent
Check their different definitions and sources. Keep them separate until the data supports an attribution relationship.
Run your next update from the kit
References
Continue the task
Use the checked report to discuss a concrete decision while keeping measurement limits and operating costs explicit.