Skip to main content
All retail guides

AI and accounting

How to verify AI answers about stock and sales

Before you use an AI answer to make a decision, check which records it includes and how it calculated the result.

By the Zemba team4 min read
Zemba artwork reading “Your AI’s answer should come with receipts.” beside a paper receipt on a dark green background. The image also says “Read the story.”

An AI answer about stock or sales should be checked against the records used to produce it. The calculation may be correct even when the selected records are incomplete or cover the wrong shop or period.

We recommend checking the product, location and period first, then reviewing the records behind the total. Here is how you can apply those checks.

Specify the product, location and reporting period

“How is the business doing?” leaves a lot open. “How many desk lamps did our Enugu shop sell from 1–7 September, excluding returns?” names a product, location, period and treatment to inspect.

Even that question needs care. Does the report use the shop’s local dates? Are returns dated when they happened or attached to the original sale period? Were any sales recorded late? A precise question reduces ambiguity; it does not make incomplete records complete.

Check the source records and calculation

For example, suppose you ask for a weekly stock summary. An assistant reports 18 net units sold: 20 units on completed sales less two units returned during the same period. The arithmetic works. Now inspect the source records.

Checklist for verifying an AI stock answer
CheckWhat you need to establish
ProductThe same item code and variant, not every product called “lamp”
LocationThe named shop only, not a company-wide total
PeriodThe intended dates, time zone and date field
StatusCompleted sales, with voids excluded and returns handled as stated
CoverageEvery relevant page of records included
Calculation20 − 2 = 18, with the source quantities visible

Check that every source belongs to the requested branch and period. Also check whether the assistant included the complete record set, including any additional pages of results.

Example: a missing sale changes the result

Suppose the assistant lists three sales records that total 20 units and a return for two. Then you notice that a fourth completed sale, for four units at the same shop during the same period, is missing.

Including the missing sale changes the total sold to 24 units. After deducting the two returned units, net sales are 22 units rather than 18. Request the corrected answer and complete source list before using the result.

Record the cause as a missing source transaction rather than an arithmetic error. This helps the reviewer check the part of the process that needs correction.

Review recommendations separately from historical results

Eighteen net units sold is a historical quantity. It is not automatically a recommendation to reorder eighteen units.

The next buying decision also needs usable stock on hand, existing orders, delivery times, reservations and current demand context. An assistant should say when those inputs are missing rather than turn a past sales number into a confident forecast.

The same distinction applies to accounting. Identifying a possible bank match does not approve it. Categorizing a transaction does not establish that it is new income or expense. Preparing a response does not mean the books changed.

AI answer verification checklist

Before acting on an answer, ask:

  • Which records did you use, and can I open them?
  • Which shop, period, currency and accounting basis does this cover?
  • What calculation or rule produced the answer?
  • What is missing, excluded or uncertain?
  • Is this a read, a proposed change or a completed action?
  • What still needs a person’s decision?

For any completed action, look for the resulting record and its audit history. Do not rely only on a chat message saying it succeeded. For a read-only answer, no business record should need to change just to prove it was read.

Summary

Verify the reporting scope, record coverage and calculation before using an AI answer in a business decision. Where information is missing, record the limitation and request a correction or further review. A recommendation may need additional information beyond the records used for a historical total.