Start with one question, identify the records it needs, and verify that those records are complete, current and consistently defined. More uploaded data does not automatically produce a more trustworthy answer.
Begin with a question you already ask
“Which items should we buy again?” sounds simple. To answer it, a shop may need sales by size and colour, sellable stock, incoming purchases and the remaining selling season. A sales total alone cannot establish a purchase quantity.
Write down the decision first. Then list the inputs. If an input is missing, keep the proposed answer at the level the records support. You can still get a useful sales summary without pretending it is a complete buying plan.
A readiness sheet for one task
| Check | Example question | If it fails |
|---|---|---|
| Scope | Which company, store or book is included? | Select and label the intended scope |
| Time | When were these records last updated? | Refresh or label the older snapshot |
| Identity | Are two similar party names different accounts? | Match stable account references |
| Units | Are quantities pieces, metres or kilograms? | Separate units; document any conversion |
| Completeness | Are returns and pending receipts included? | State the omission before using the result |
| Access | Should this person see every company's records? | Restrict access to an agreed scope |
Why a correct addition can still mislead
Consider a fictional export with ₹80,000 of sales and ₹8,000 of returns. Sales after returns are ₹72,000 if both amounts belong to the selected period and use the same tax basis. If the return belongs to another company, subtracting it gives the wrong answer even though the arithmetic is correct.
Record what the figure means beside the result. “August sales after returns, Shop A, excluding tax” is checkable. “Business performance: ₹72,000” leaves too much open.
Test an ordinary case and an awkward one
Use one known transaction, one account total and one business total. Compare each with the matching source report. Then try duplicate names, an empty date range and a request for data you have not supplied. Keep discrepancies rather than correcting the reference answer just to make the trial pass.
The NIST generative AI profile describes risks including confidently presented false content. This worksheet is our practical way of making a business answer easier to challenge; it is not a certification under that framework.
You can begin with paper records
For a first task, a carefully maintained sheet may be enough. Use one row per transaction or item movement, retain its reference and date, and avoid placing totals inside the detail rows. Have the person who knows the records check the sheet before relying on an automated summary.
Once the trial works, agree who updates it and when. “Kal ka data hai” can be perfectly useful when everyone knows it is yesterday's position. An unlabeled old answer is the problem.
Questions you might ask
Can I start without an ERP?
Yes, a well-defined sheet can support a limited task. Its records still need consistent dates, units, references and an update process.
How do I know whether an AI business total is reliable?
Check the company, date range, calculation basis and source entries against the matching report. Correct arithmetic alone is not enough.