How to Check Every Figure in an AI Business Analysis
Why AI answers about numbers are hard to trust
Large language models write fluent, confident text, and that is exactly the risk with business numbers. A model can round a figure, mix up two weeks or produce a plausible total that appears nowhere in the data, and the paragraph reads just as convincingly as a correct one.
When a manager acts on that number, for example chasing the wrong customer or cutting the wrong product, the cost is real. Checking it afterwards means rebuilding the analysis by hand, which removes the reason for asking the AI in the first place.
Trust comes from a check the reader does not have to perform: every figure tested against the rows that were actually read, any answer with an unsupported number refused, and each figure linked to its row so anyone can confirm it in seconds.
Why this happens
- AI text is checked by eye, if at all.
- Figures are not linked to the rows behind them.
- Rounded or invented numbers look like real ones.
- Nobody records which data an answer used.
- Refused or failed analyses leave no trace.
Biznsbook addresses this through figure check against the data read, source reference on every figure, The data it used tables, Runs log of every analysis in AI Analyst, reading your own company data with each figure linked to its source row.
None of this needs more data: the business records are already in the system. What is missing is a quick, trustworthy way to ask about them and see the rows behind each answer.
Step-by-step: Check Every Figure in an AI Business Analysis
Written for owners, finance managers and team leads using Biznsbook AI Analyst on their own company data.
- Run an analysis. Ask a question in AI Analyst → Ask, or click Run now on a brief in AI Analyst → Briefs.
- Read the figures in the findings. Summary and Findings each have their figures, and each figure shows where it came from, such as sales.weeks#1: the source and the row number.
- Open The data it used. On the right, The data it used shows every table the analyst read, with numbered rows. Find the row named in the figure's source.
- Confirm the figure. Compare the value in the answer with the row. A figure must match a supplied row, or be a simple sum, difference or percentage of rows it references.
- Read what it could not see. What the analyst could not see lists any data it was missing, so you know the limits of the answer.
- Handle a did not match the data failure. If an analysis fails with "did not match the data", the AI quoted a figure that was not in the data even after a second try. Run it again or reword the question.
- Review Runs. Runs lists every analysis, including ones that were refused, with the question, the kinds of data, the model and region, tokens and cost.
Review results after the first full monthly cycle. Adjust roles, mappings, or approvals where the same exception repeats.
Setup and screen steps are in the AI Analyst guide inside Biznsbook. This page focuses on the business question; the AI Analyst page summarises the add-on.
Common mistakes to avoid
- Mistake 1: AI text is checked by eye, if at all. Repeating this each month usually shows up first in decisions made on numbers nobody checked.
- Mistake 2: Figures are not linked to the rows behind them. Repeating this each month usually shows up first in decisions made on numbers nobody checked.
- Mistake 3: Rounded or invented numbers look like real ones. Repeating this each month usually shows up first in decisions made on numbers nobody checked.
- Mistake 4: Nobody records which data an answer used. Repeating this each month usually shows up first in decisions made on numbers nobody checked.
- Mistake 5: Refused or failed analyses leave no trace. Repeating this each month usually shows up first in decisions made on numbers nobody checked.
Each of these leaves managers acting on numbers nobody checked. Ask about the data where it lives, and read the source rows before acting on a figure.
Best practices that hold up as you scale
- Run an analysis — Ask a question in AI Analyst → Ask, or click Run now on a brief in AI Analyst → Briefs.
- Read the figures in the findings — Summary and Findings each have their figures, and each figure shows where it came from, such as sales.
- Open The data it used — On the right, The data it used shows every table the analyst read, with numbered rows.
- Confirm the figure — Compare the value in the answer with the row.
- Read what it could not see — What the analyst could not see lists any data it was missing, so you know the limits of the answer.
Ask specific questions, read The data it used before acting, and save the questions you repeat as briefs.
How Biznsbook supports this workflow
figure check against the data read is part of Biznsbook AI Analyst and reads the same business records the rest of Biznsbook uses, under the asking person's rights, with every figure checked against the data before the answer is shown.
source reference on every figure is part of Biznsbook AI Analyst and reads the same business records the rest of Biznsbook uses, under the asking person's rights, with every figure checked against the data before the answer is shown.
The data it used tables is part of Biznsbook AI Analyst and reads the same business records the rest of Biznsbook uses, under the asking person's rights, with every figure checked against the data before the answer is shown.
Runs log of every analysis is part of Biznsbook AI Analyst and reads the same business records the rest of Biznsbook uses, under the asking person's rights, with every figure checked against the data before the answer is shown.
AI Analyst is an add-on ($10 per company per month) and needs no other module; each analysis area needs its module, scheduled analysis also needs Business Automation, and your company pays its own AI provider for tokens.
Suggested implementation timeline
- Day 1: Add an approved AI model and API key in AI Analyst → Settings, tick the kinds of data allowed and turn analysis on.
- Week 1: Give the AI Analyst User role to two or three people and try figure check against the data read on real questions.
- Week 2: Open The data it used on each answer and check the figures you plan to act on.
- Month 1: Save the questions people repeat as briefs, and schedule the weekly ones if you use Business Automation.
- Ongoing: Building a habit of checking.
What the figure check does and does not cover
The check is about numbers. Every value in the figures of an answer must reference a row the server supplied and match it, within rounding, or be a sum, difference or percentage of referenced values. That stops invented totals, numbers taken from the wrong week and figures the model made up to complete a sentence.
The check does not judge the opinion around the numbers. Findings and What to do are the AI's explanation of the checked figures, written like an analyst's notes. Read them as a starting point: the figures are from your data, and you still decide what action makes sense for your business.
Because the data it used is kept with the answer in your company's records, an answer can be checked later too. Someone reviewing a decision made from an analysis can open the run and see exactly which rows it was built on.
Metrics to track monthly
- Questions asked per week and by whom
- Analyses refused or failed, and why
- Recommendations acted on
- Tokens and cost per month in Runs
Start with three metrics; trend direction matters more than a single point-in-time snapshot.
Manual analysis vs Biznsbook AI Analyst
Compare common manual analysis with Biznsbook figure check against the data read and the related AI Analyst capabilities.
| Capability | Manual / Spreadsheet | Biznsbook |
|---|---|---|
| Unsupported numbers | ❌ Read and believed | ✅ Answer rejected, one retry, then not shown |
| Source of a figure | ❌ Unknown | ✅ Each figure names its source row |
| Data behind the answer | ❌ Not kept | ✅ The data it used, with numbered rows |
| Derived figures | ❌ Calculated by the model | ✅ Only sums, differences or percentages of referenced rows |
| Failed analyses | ❌ Lost | ✅ Listed in Runs with the reason |
| Missing data | ❌ Silently ignored | ✅ Listed under What the analyst could not see |
Building a habit of checking
For the first few weeks, open The data it used on every answer and check two or three figures against their rows. Once people see the sources match, they check only the figures they will act on. Review Runs once a month for refused or failed analyses and reword questions that fail often.
Write down which questions each role asks every week and who acts on the answer.
Frequently asked questions
What happens when the AI quotes a wrong number?
The answer is rejected and the AI gets one retry with the errors. If the figure still does not match the data, the run fails with "did not match the data" and the answer is not shown.
What does a source like sales.weeks#1 mean?
It names the table the figure came from and the row number in The data it used, so you can find the exact row beside the answer.
Can the analyst calculate new numbers?
A figure must match a supplied row or be a simple computation, such as a sum, difference or percentage, of the rows it references.
Are follow-up answers checked too?
Yes. The analyst remembers the last three answers of a conversation, but every figure is still checked against fresh data.
Who can see all runs?
Each person sees their own runs. People with the View AI Analyst Audit right see everyone's runs.
Does the check make the answer always right?
It makes sure every figure comes from your data. The explanation and recommendations are still the AI's reading of those figures, so check them against what you know about the business.
How this differs by industry
Retail
A store manager checks that the sales drop quoted for Store 2 matches the weekly sales row before moving staff or stock.
Wholesale & distribution
A credit controller confirms the overdue balance in the answer against the receivables row before calling the customer.
Manufacturing
A planner checks the days of cover quoted for a finished good against the inventory row before changing a production plan.