What you can hand to an AI data analyst, and what you cannot

Safe to delegate

• Aggregation: totals, means, medians and quantiles by month or category
• Visualisation: line, bar, scatter and heatmap
• Preparation: checking nulls, type conversion, joins, duplicate detection
• Descriptive statistics: distribution shape, outliers, correlation coefficients
• Repetition: the same treatment applied to another period or segment

Within that range, a correctly phrased instruction returns a correct result almost every time. Work that takes tens of minutes by hand takes a few minutes, so delegate it without hesitation.

Not safe to delegate

1. Causal judgement

'Ad spend and revenue correlate at 0.8' is a correct computation. 'So increasing ad spend will increase revenue' is a different claim, and the data alone cannot support it. Seasonality may be moving both, or revenue may have driven the spend rather than the other way round.

Asked directly, the model will write prose that reads as if causation were established. Change the question — 'can causation be claimed from this correlation, and what rules it out?' — and sensible caveats appear. Use it knowing the answer depends on how you ask.

2. Where the data came from

When a revenue column contains zeros, nothing in the data says whether nothing was sold, whether collection failed, or whether a system migration dropped the records. Only the person who owns the process knows.

Treating those zeros as real sales drags the average below reality, and the AI will not make the distinction. Confirming the meaning of nulls and zeros per column, with the business, is a step that comes before analysis.

3. Interpretation in context

'Churn spikes on Fridays' is a finding. Whether that reflects people reviewing subscriptions before the weekend or the billing cycle closing on Thursday cannot be judged without knowing how the business runs. Producing the fact and assigning it meaning are separate jobs.

Three things to ask before any analysis

The third earns its place most often. Three days missing after an outage, unnoticed, produces a monthly average below reality. Checking for gaps in the dates prevents it outright.

Choosing the tool

For one-off exploration, use a conversational tool; for an aggregation you rerun monthly, use one where the code survives. Julius and Hex are the respective examples, compared in detail in our Julius vs Hex piece.

Pasting a CSV into a general-purpose AI chat also works. Up to a few thousand rows it is adequate and requires no new subscription — subject to file size limits and to how you feel about the data leaving your systems.

Tables containing personal data

• Drop name, email, phone and address columns before analysis
• Pseudonymise customer IDs where the real ID is not needed
• Check that inputs can be excluded from model training
• Check whether internal policy restricts uploads to external services

Most analysis survives the removal of identifying columns. The cases that genuinely need 'who' are rarer than expected; usually 'which segment' is enough. Making removal the default is safer than deciding case by case.

Frequently asked questions