AI can explain, clean, classify and visualise data — but only if it has understood what each column actually is.
Do now: Profile one small table, then recalculate one figure by hand.
If you have not checked one number yourself, you have a chart, not a finding.Numbers get believed. A wrong sentence invites argument; a wrong figure in a clean chart gets acted on. The check costs two minutes and prevents the most expensive category of mistake these tools produce.
Profile before you analyse
Ask what each column is, and where the tool is guessing. Column names in real files are abbreviations, legacy terms, or someone's initials from four years ago. A tool that assumes 'amt' means the amount charged when it means the amount paid will produce fluent, confident nonsense.
Then ask for the mess: missing values, duplicates, inconsistent units, mixed date formats, categories that mean the same thing spelled three ways. Every real dataset has these. A tool that reports none has usually not looked.
Finally, ask what the data cannot answer. A sales table shows what was sold, not what was wanted and unavailable. Naming the absence prevents the most common analytical error, which is answering a question the data was never able to address.
Every real dataset is messy. If the profile comes back clean, be suspicious of the profile, not impressed by the data.
State the decision before opening the file
Write down what you would do differently depending on the result. If no answer changes any action, the analysis is entertainment. This one habit eliminates most wasted analysis, and it also tells you how precise the answer needs to be.
Then define what would count as good evidence for that decision, before you see the numbers. Deciding afterwards is how people end up choosing the cut of the data that agrees with them without ever noticing they did it.
Verify by hand, and keep the assumptions visible
Take one figure and recalculate it independently — on paper, in a separate formula, or against a source document. If it does not match, you have found something important. If it does, your confidence in the rest is earned rather than assumed.
Write down the assumptions the analysis rests on: how missing values were treated, which rows were excluded, what period was used, which definition of the metric applies. These are choices, not facts, and different reasonable choices produce different answers.
Keep the original file untouched and work on a copy. Cleaning is a series of decisions, and you will want to revisit some of them.
One hand-checked number is worth more than a page of generated commentary.