AI Skills for Analysts: From Data Wrangling to Insight

The analyst's version of the AI question isn't "can it do my job". It's narrower and more useful: which parts of the work does it do reliably, and which parts does it do convincingly but wrong?

The answer is fairly clean. AI is strong on the mechanical layer — reshaping data, writing the formula you can't remember, describing what a dataset contains, drafting the commentary once you know what you want to say. It's unreliable on the layer that makes analysis worth paying for: knowing which number is misleading, which comparison is unfair, and which finding is an artefact of how the data was collected.

The failure mode is specific. It doesn't return an error when it's out of its depth. It returns a clean, plausible, well-formatted answer with a wrong number in it.

Skill one: hand over shape, keep hold of meaning

The most reliable division of labour is structural work to AI, interpretive work to you.

Structural: pivot this, reshape that, find the duplicates, write the lookup, describe the distribution, convert this into that format. These are verifiable in seconds and the model is genuinely good at them.

Interpretive: is this trend real or seasonal, is this segment small enough that the percentage is noise, does this correlation have an obvious confound, would a stakeholder read this chart the way I intend. These depend on knowing the business, not the dataset.

Analysts who get the most out of AI aren't the ones with the best prompts. They're the ones who've drawn this line clearly and stopped crossing it.

Skill two: describe the data before you ask about it

A model looking at a spreadsheet sees columns and values. It doesn't know that status has five possible values but two of them were deprecated in March, or that revenue is booked at contract signature rather than delivery, or that the last three rows are a manual adjustment.

Every one of those facts changes the correct answer. None of them are visible in the file. Saying them out loud before asking your question is the single highest-return habit in AI-assisted analysis.

Skill three: check the arithmetic, always

Language models handle numbers differently from how they handle text, and the reliability is not the same. A total that's slightly off, a percentage computed against the wrong denominator, a filter silently applied — these are common, and they're invisible in a well-formatted answer.

The practical habit is cheap: spot-check one figure per output against the source. Not all of them. One. It catches most of what matters, and it takes thirty seconds.

Skill four: make the chart argue something

Ask for "a chart of this data" and you get a chart of that data. Ask for "a chart that shows whether the drop is seasonal" and you get something that makes a case. The second is the one you can put in front of a stakeholder.

Nadia — the number nobody questioned

Nadia is a senior analyst at a retail group. She used AI to build a quarterly performance summary across 40 store locations, and the output was excellent — clean structure, sensible groupings, plausible commentary.

One figure was wrong. Average basket size had been calculated as the mean of each store's average, rather than weighted by transaction volume. Two very small stores with unusual product mixes pulled the group figure up by enough to change the story from flat to growing.

The model hadn't done anything unreasonable — she'd asked for "average basket size by region" and it had picked an interpretation. She hadn't specified which average she meant, because in her head there was only one.

Her rule now: whenever an output contains an aggregate, she asks the model to state how it computed it. Not to check its work — to check her own question.

Callum — the correlation that wasn't

Callum is an insights lead at a subscription business. He asked AI to look at churn drivers and got a confident finding: customers who contacted support were substantially more likely to churn.

True, and useless. Customers contact support when something's broken. The support contact isn't causing the churn, it's a symptom of the thing that is. Acting on the finding as stated would have meant discouraging support contact — actively making the problem worse.

The model had accurately described a pattern in the data. What it couldn't do was know which direction the arrow points, because that requires knowing how the business works.

His habit now: for any relationship an AI surfaces, he writes down the mechanism before he acts on it. If he can't state a plausible mechanism, it goes in a "look into" list rather than a recommendation.

The one thing

The value of an analyst was never in producing the numbers. It was in knowing which numbers are trustworthy and which comparisons are fair — and that hasn't moved.

What has moved is how much of the mechanical work you should still be doing by hand. Probably far less than you are. The trade is worth making, as long as the thirty-second arithmetic check goes with it.