Gains:
- Ability to analyze sales, cost and operation tables with artificial intelligence and detect trends, breaks and outliers
- Ability to verify against arithmetic errors and hallucinations by redoing or manually checking each calculation
- Ability to distinguish correlation from causation and question data quality and sample limits before interpretation
One of the most repeated phrases in management consulting is: "What do the numbers say?" A customer's sales data, cost table, customer segments or operational metrics; they all tell a story. The consultant's job is to find that story, prove it, and show it clearly to the decision maker. Artificial intelligence interprets a chart in seconds, points out trends, makes calculations, and even generates code (Python, Excel formula) for analysis. But there is a particular danger here: AI can silently make mistakes in arithmetic and present the wrong number with correct confidence. Therefore, the golden rule in data analysis is to redo or manually check each calculation.
Two ways to work with data
There are two primary methods of data analysis with AI, and knowing the difference is critical for accuracy.
- Textual interpretation of the model: You paste the data directly into the chat and say "comment". The model reads the patterns but does the calculations "mentally"; This is the path with the highest risk of error. Suitable for small tables and qualitative interpretation, not for precise accounting.
- Code/tool generation: You print analysis code (Python/pandas) or Excel formula to the model; The code does the calculation, not the model. This way is much more reliable because the arithmetic operates in a deterministic tool. Prefer it for big data and precise number.
Tip: If you need an exact number (total turnover, average, growth rate) ask the model to "write the Excel formula/Python code that calculates this", run the result yourself. Never consider the number given by the model as final.
How to set up analysis prompt
A good analysis prompt includes the data, the question asked, and the validation rule.
Your role: data analyst. Below is a 24-month sales chart (month, region, product, quantity, turnover). Task: 1) Summarize the total and monthly turnover trend. 2) Find the 3 fastest growing and fastest shrinking region-product breakdown. 3) Mark if there are any outliers (unexpected jump/fall). Rules:- Write the Excel formula that makes each calculation next to it; I will verify.- Separate the interpretation from the data: first “the data says”, then “possible meaning”.- Do not claim causality; show pattern only.
This request has three critical elements: requiring the formula (verifiability), separating data from interpretation (integrity), and prohibiting the claim of causality (the subject of the next section).
Correlation is not causation
The most common and costly logical error in analysis is assuming that one of two things acting together causes the other. Just because two variables increase together (correlation) does not prove that one causes the other (causation); There may be a third factor or coincidence. Classic example: as ice cream sales increase, so do drownings; The reason is not ice cream, hot weather raises both.
AI can easily write a correlation as “X increases Y.” When the consultant sees this sentence, he should stop and ask: "Is this really the reason, or is there a hidden factor?" Claim of causality; controlled comparison requires knowledge of time sequence and domain.
Your role: critical data analyst.The following analysis contains a "X increases Y" claim.Task:- Evaluate whether this relationship is correlation or justified causation.- List 3 possible latent (third) factors that may influence both.- Say what additional evidence (comparison group, time sequence) is needed to support causality.Use strict causal language; Write “possible” and “should be tested.”
Questioning data quality before interpretation
Good analysis does not come from bad data. Ask each chart these questions before commenting:
control
Why is it important?
How to care
missing data
Empty cells distort the mean
Count blank/zero ratio
outliers
A single outlier distorts the trend
Look at min-max and distribution
Definition consistency
Is the "turnover" the same every month?
Confirm column definition
time range
Does seasonality mislead?
Compare year-over-year
sample
Does the data represent the entire universe?
Ask for coverage rate
Before INTERPRETING this table, do a data quality check:- What is the missing/zero cell ratio?- In which rows are the extreme (outlier) values?- Are the column definitions consistent, are there any suspicious ones? Just give an audit report; Don't comment yet.
three mini cases
Case 1 — Silent arithmetic error. A consultant has the model collect 12 months' turnover; the model says “142.6 million.” The consultant checks in Excel: actual total 128.4 million; The model added two months incorrectly. If a formula had been requested instead of a mental calculation, there would have been no mistake. If the wrong number had been entered into the presentation, customer confidence would have been shaken.
Case 2 — Spurious causation. In a retailer, the model says, "The turnover increased in the months when the discount was offered, so the discount increases the turnover." The consultant stops: discounts are already made during the holidays; The main driver is the holiday demand. To measure the discount effect, a comparison with the holiday without a discount is required. False inference prevents a margin-burning suggestion like “let's discount every month.”
Case 3 — Outlier trap. A zone average order value appears abnormally high. The model interprets this as a "premium zone". The consultant opens the line: a single corporate order of 4.2 million inflated the average; When that line is removed, the region is ordinary. Using median would avoid this trap in the first place.
Weak prompt / Strong prompt
Weak prompt:
Analyze this sales chart and state the key findings.
The model calculates head-on, claims causality, and gives unverifiable findings.
Powerful prompt:
Your role: meticulous data analyst. First do a data quality audit (missing, outlier, definition consistency). Then: monthly trend, top 3 breakdowns, outliers.- Write the Excel formula for each account; I'll verify.- Also give median instead of mean (to see outlier effect).- Separate "data says" from "possible meaning"; DO NOT CLAIM causality. - Mark "unclear" where you are not sure.
Segmentation: the truth the average hides
One of the most powerful moves in consulting analysis is segmentation: breaking a total number into meaningful subgroups. The “average customer” is often a non-existent construct; The truth is hidden in the subgroups. For example, while the average profit margin appears to be 8%, 20% of customers may bring 25% margin and 30% make a loss; average completely hides this fact. Artificial intelligence helps you quickly break down a chart across different axes (customer type, region, product, channel) and find which breakdown shows the most striking difference.
Your role: data analyst. The table is below.Task: Break this data on the following axis: [customer type / region / product].For each group: total, share (%), mean AND median, group size (n).Rules:- If group n<30, mark "small sample - attention".- Write the formula for each account; I will verify.- Show highest and lowest group; Evaluate whether the difference is real or due to a single outlier. First let me tell you which breakdown I expect: [hypothesis].
Common mistakes
- Relying on the model's head calculation. Outsource the arithmetic to the deterministic tool (Excel/Python); Check the result independently.
- Mistaking correlation for causation. Don't write the sentence "X increases Y" without evidence; Ask about the hidden factor.
- Blindly looking at the average. The single outlier inflates the mean; See also median and distribution.
- Bypassing data quality. Interpretations resulting from incomplete, duplicate or inconsistent data are misleading; check first.
- Forgetting the example. Generalizing from a small or biased sample to the entire population will produce erroneous results.
- Confusing data with interpretation. Show separately what is measurement and what is inference.
Attention: If the result of the analysis of financial, health or operational safety data will be the basis for a decision, the calculations and interpretation must be confirmed by the relevant expert (financial advisor, actuary, field expert). Artificial intelligence produces preliminary analysis; The final evaluation and responsibility belongs to humans.
In summary
Artificial intelligence is a powerful aid in interpreting tabular data, finding trends and breakdowns, and generating analysis code. But since it can silently make mistakes in arithmetic, it is necessary to have each calculation re-done with a deterministic tool and checked manually. Separating correlation from causality, looking at the median next to the mean, checking data quality before interpretation, and observing sample limits are the basic disciplines of consultancy. Model shows pattern; People create meaning and responsibility.
Application task
Select an actual or representative sales/cost statement. First run the data quality check prompt and find at least one issue (missing, outlier, definition inconsistency). Then extract trends and breakouts with powerful analysis prompt; Also manually verify the formula for each account. Compare the mean to the median and detect an outlier effect. Finally, find a causality claim that the model writes and rephrase it as “correlation.”
checklist
- [ ] I re-ran each exact calculation with the deterministic tool (Excel/Python).
- [ ] I performed a data quality check before commenting.
- [ ] Besides the mean, I looked at the median and distribution.
- [ ] I preserved the correlation-causation distinction.
- [ ] I questioned the sample and time period limits.
- [ ] I separated "data says" and "possible meaning".
- [ ] I planned expert approval in decision-critical analysis.