How to Stop AI Making Up Numbers in Your Business Reports

AI chatbots make up numbers because large language models generate text by predicting likely words, not by calculating. Ask one to total a column and it may produce a figure that looks right but isn’t. The fix is to separate the jobs: calculate every figure with code (a spreadsheet, a database query or a reporting tool), then let the AI write the narrative using only those pre-computed numbers. This is often called compute-then-narrate.

Why language models get numbers wrong

A large language model (LLM) such as the ones behind ChatGPT, Claude or Gemini is trained to produce plausible text. That is remarkable for writing and summarising, but it creates specific problems with figures:

• Arithmetic is pattern-matching, not calculation. Unless the tool runs code behind the scenes, the model predicts digits rather than computing them. Long sums and percentages are where errors creep in.
• Large files don’t fit. A model can only read so much text at once. Paste in 50,000 rows and it may see only part of the data, then answer as if it saw all of it.
• Gaps get filled. If you ask for ‘growth vs last year’ and last year isn’t in the data, a model may produce a sensible-sounding percentage anyway.
• Confidence doesn’t signal accuracy. A wrong number is written with the same fluency as a right one.

How to spot a hallucinated figure

1. Reconcile the totals. Compare the report’s total revenue, row count and date range with your source system. These three checks catch most problems.
2. Look for orphan numbers. A figure in the text that doesn’t appear in any table or chart is suspect.
3. Check comparisons. Every ‘up 12%’ needs both periods present in the data. If only one period was provided, the comparison was invented.
4. Recalculate one percentage. Pick a claimed share or growth rate and do the division yourself.
5. Ask ‘where did this come from?’ A trustworthy tool can point to the calculation or column. A chatbot will often apologise and give a different number.

The compute-then-narrate approach

The reliable pattern is to give the AI facts, not data. It works in three stages:

1. Compute

Code calculates every figure the report needs: totals, averages, period-on-period changes, top and bottom performers, shares, trends and data-quality counts. The same input always gives the same output, and results can be tested.

2. Constrain

The AI receives those computed results, often as a structured list of facts, with instructions to use only these figures and not to calculate or estimate new ones. It doesn’t need the raw rows at all, which also reduces privacy risk.

3. Narrate and verify

The AI writes the summary, insights and recommendations. Charts and KPI tables are drawn from the computed figures, never from the AI’s text. A final check can compare numbers mentioned in the narrative against the computed set and flag anything that doesn’t match.

What you can do today with a chatbot

If you use a general chatbot for reporting, you can apply the same principle manually:

• Build your pivot tables and KPIs in Excel or Google Sheets first.
• Paste only the summarised results into the chat, not the raw file.
• Instruct it: ‘Use only the figures below. Do not calculate new numbers. If something is not in the data, say so.’
• Check every number in the reply against your table before sharing.

It works, but it’s manual, and it depends on you remembering every time. We compare the two approaches in ChatGPT vs an AI reporting tool.

How Summarix handles it

Summarix is built on compute-then-narrate. Every number and chart in a report is computed by Summarix’s own code from your data; the AI writes the narrative around those figures and doesn’t invent numbers. The AI sees a statistical profile and a small masked sample (at most 15 rows) rather than your whole file, with personal information removed first. Each report includes a step-by-step build log so you can see how it was produced, plus data-quality notes flagging missing or suspicious values. More detail is on the security page.

The takeaway

Don’t ask AI to be a calculator. Let code do the maths, let AI do the writing, and keep a human checking the result. Clean inputs help too: our list of data quality issues in reports covers the problems that trip up even accurate calculations.