An overstated sales figure can lead a business to buy stock it does not need. An AI assistant may cite genuine invoices and still count the same sale twice. Before using its figures to make decisions, management needs to know what was counted and how the calculation was checked.
The opportunity is worth pursuing. Staff can get answers faster without asking the finance team to prepare a new report each time. Start with one figure the business regularly uses, agree what it means and test the calculation behind it. A small Norwegian company can do this with its accountant and system provider.
How NOK 150,000 becomes NOK 250,000
Consider a hypothetical Norwegian equipment supplier reviewing invoiced sales before ordering more stock. There are two invoices and no credit notes. Both invoices fall within the selected period, and all amounts exclude VAT.
| Invoice | Amount | Recorded deliveries |
|---|---|---|
| A | NOK 100,000 | Two partial deliveries |
| B | NOK 50,000 | One delivery |
| Correct invoice total | NOK 150,000 | Three deliveries |
The assistant links each invoice to its delivery records to show which goods have been dispatched. Because invoice A has two deliveries, its full amount appears on two rows. Adding the invoice amounts across those rows gives NOK 100,000 + NOK 100,000 + NOK 50,000 = NOK 250,000.
The total is NOK 100,000 too high, even though every invoice and source link is genuine. A manager relying on it could mistake the inflated total for stronger sales and order too much stock.
The calculation must count each relevant invoice once, regardless of how many deliveries it covers. Removing repeated amounts is not a safe fix: two different invoices can have the same value. The system must distinguish the invoices themselves, for example by invoice number within the correct company.
This can go wrong in conventional reporting too. An AI assistant introduces another way to build calculations from company data. The check should therefore cover the calculation itself, not just whether the answer sounds reasonable or includes sources.
What the recent research does and does not show
In a 14 October 2026 article, AtScale reports preliminary tests in which AI models answered business questions more accurately using its semantic layer. This is a shared set of definitions and calculation rules that the assistant uses when retrieving figures.
There is an important qualification: the vendor's main comparison uses the best of three runs with its product against one run without it. AtScale also reports improvements on the first run. The findings justify testing the approach, but do not establish how much a Norwegian business would gain or whether this particular product is the best choice.
The underlying BIRD-Interact study tests database tasks, including ambiguous questions and clarification through a simulated user. It contains 600 tasks, of which 410 concern business intelligence. It does not establish how an assistant will perform with your data or Norwegian business terminology.
The practical implication is to test the definitions and calculations alongside the model. A more capable model may help, but its benchmark score cannot tell you whether your sales figure is correct.
Agree what the figure means
For the equipment supplier, “sales this month” is not precise enough to guide purchasing. Does it mean orders received, invoiced sales or payments received? Which company and period are included? How are credit notes and cancellations handled? If this month is incomplete, is the comparison with last month meaningful?
The finance manager and the people using the report should agree these definitions. The technology provider should implement them, rather than decide through undocumented defaults.
Write a short specification for the first figure: its purpose, included transactions, period, exclusions, currency treatment, source system, update time and responsible person. Connect it to the actual calculation the assistant will use. For the simplified example above, the rule is to sum each invoice once for the selected company and period, excluding VAT. The two test invoices must produce NOK 150,000 even when delivery details are included.
Norwegian terms also need attention. When someone asks about “omsetning”, “salg” or “innbetalinger”, the assistant should use the agreed definition or ask which figure they need. If an export business combines EUR and NOK, the exchange-rate method and date must be clear. Fluent Bokmål is not evidence of a correct calculation.
The example concerns internal management information. It does not determine when revenue should be recognised in statutory accounts.
Test the calculation before relying on the answer
Have the finance manager establish expected results using a separate, checked calculation. Include partial deliveries, a credit note, two invoices with the same amount, an incomplete period and a question that needs clarification. Reserve some cases for the final test so the assistant is not assessed only on examples used to adjust it.
Compare configurations using the same data, questions, access rights and number of attempts. Record failures as well as successes. A selected demonstration does not show how consistently the system will work in everyday use.
Check whether another authorised person can reproduce the answer from the same data version. Does the assistant state the period and scope? Does it ask for clarification or say that it lacks sufficient data when necessary? A consistently repeated answer can still be wrong if the underlying report is wrong.
Start with read-only access. Retrieving totals should not require permission to change invoices or place orders. Check that summaries and follow-up questions do not expose information beyond each user's existing access rights.
Measure the time needed to obtain and verify an answer, errors that could change a decision, unresolved questions and ongoing maintenance. Before the trial, agree what would prevent the figure from being used. In the invoice example, an unexplained discrepancy with the checked total is a reason to investigate before the figure guides a stock purchase. Asking a necessary clarification should count as correct behaviour, not as a failed answer.
Start with a solution the business can maintain
A small AS may get enough value from an assistant that explains a checked export from an existing finance report. A business with several systems may need a shared service that performs agreed calculations. A larger organisation may justify a dedicated semantic layer. Choose according to the problems found in testing and the resources available to maintain the solution.
Waiting for every company dataset to be perfect can also hold back useful work. Start with one well-defined figure. Connect the trial to a specific business workflow and measure what the time saved enables.
Management should own the definitions, implementation and investigation of errors. If the figures materially influence budgets or investments, the board should ask what checks support them, which limitations remain and who is accountable for correcting discrepancies. Directors do not need to inspect database queries. The guide to AI in board preparation addresses the separate question of using AI to read and summarise papers.
For the next management meeting, choose one figure that regularly informs a decision. Ask the team to show what it includes, how it is calculated and whether the result matches a checked report. That gives the business a concrete starting point for making reliable figures easier to access with AI.
Sources checked on 17 October 2026. The equipment supplier and invoice example are hypothetical. The testing approach and recommendations are the author's analysis. The vendor's findings are preliminary and do not estimate performance in Norwegian businesses. The illustration is AI-generated.