KilatBot Articles

Articles

Reports without made-up numbers

· · Updated

To make reports without made-up numbers, limit the data sources, ask for the calculation, and check the result against the original file. An AI agent can help prepare a report, but instructions do not guarantee accuracy. When data is missing, ask it to name the gap rather than fill it with a guess.

Choose the question before the summary.

Start with one question your data can answer. Example: what is the value of completed orders within a specific period? Do not ask for conclusions about the whole business from one sales file. Specify the start date, end date, order status, and value column to use.

Keep order value, money received, and profit separate. Choose one measure for your first test. If you want profit calculated, provide the necessary cost records. Do not ask the agent to treat absent costs as zero just to make the report look complete.

KilatBot's admin and reports template includes explaining a CSV. That is a starting point, not proof that every file format or business calculation is covered. Read the agent guide for working in chat and reviewing results.

Set a clear source boundary.

Keep a copy of the original file. For the first test, use sample records without customer names, phone numbers, or addresses. Name the file for its period. If several versions exist, choose one as the report source so your reference does not change during review.

Explain ambiguous columns. Does the date mean an order was placed or payment arrived? Is the value after discounts? Are cancelled orders still in the export? Answer from your business records, not the agent's assumptions.

Example instruction to test:

Use only the attached CSV. Calculate completed order value for the period I specify. Show the column names, filtering rules, and calculation. Do not use estimates or information outside the file. If information is insufficient, name what is missing and ask me.

This request sets proposed working rules. Still check whether the response follows them. The phrase “based on the data” is not enough to accept a result.

How should a report treat missing data?

If costs are absent, ask for the profit section to say “cannot calculate yet”. If the comparison period is absent, ask for no growth or decline claim. If a column is unclear, resolve its meaning before using any number that depends on it.

The agent may not always refuse to fill a data gap. Do not treat a confident answer as an automatic check. Test a request with deliberately incomplete material and inspect the actual response.

Example check: provide a sample CSV without costs, then request a sales and profit summary. Acceptance does not mean a page full of numbers. Accept only if sales can be traced and profit is marked unknown. If the agent guesses, do not use that report.

How do you check an AI report's numbers?

Ask for three separate parts: source figures, calculation details, and unknowns. This gives you a fixed review order. Check the numbers first; polish the writing once they match.

  • Scope: match the period, statuses, and excluded rows to your instructions.
  • Calculation: recalculate the main total from the original file using your usual tool.
  • Conclusion: require supporting material for claims about changes or causes, not merely a plausible sentence.

If results differ, ask which rows were counted. Do not correct only the final number. Find the differing rule, fix the instruction, and recalculate from the same source. Keep a change note so the next report uses the same definitions.

Do not call a report correct because the agent repeats your preferred number. Acceptance requires a calculation you can follow from source to result, including exclusions and unavailable information.

Repeat once the checks are clear.

Keep the request that passed review as a working reference. Before making it recurring, also try an empty file, a different period, and renamed columns.

For daily preparation, continue with the article on reports from CSV files. For recurring work, read the guide to scheduling agent tasks. A schedule does not replace source checks or your acceptance of the report.

Judge value using total time spent preparing data, checking numbers, and correcting output. Do not count savings from the speed of the first answer alone. Keep your own important files; agent storage is not a backup. Short answers on price, access and data are in the Q&A.

Prepare one report to test.

Choose one question, one sample file, and a reference result you have checked yourself. Bring all three to your first test. Get started.

Read next: An AI daily report generator that reads your CSV file · Recurring tasks: daily data analysis, content drafts

Built to work for you

Put your agent to work.

Start from a template and shape it to your business.

Get started