Will AI replace data analysts? A careful argument, not a prediction

Why a prediction isn't useful here

Headlines often quote forecasts about jobs "at risk". Such forecasts depend heavily on assumptions — how fast tools improve, how quickly companies adopt them, and whether saved time leads to fewer jobs or more work. Different reasonable assumptions give very different answers. A more useful approach is to reason about tasks, because tasks are what tools actually change.

The argument, step by step

Premise 1: A job is a bundle of tasks. An analyst's week mixes data extraction, cleaning, querying, building reports, investigating anomalies, explaining findings, and advising decisions.

Premise 2: AI tools change tasks unevenly. They are strongest where the task is well specified and output can be checked cheaply — drafting a query from a clear request, summarising text, formatting. They are weakest where the question is unclear, context is unwritten, or a mistake is expensive and hard to spot.

Premise 3: Checking stays with people when being wrong is costly. An organisation can delegate drafting to a tool, but someone must still be accountable for a number in a board pack or a regulatory report. Checking requires understanding the data.

Premise 4: Cheaper analysis often increases demand for it. When producing an analysis becomes faster, managers tend to ask more questions, not fewer. Whether that offsets task automation varies by company and cannot be assumed.

Conclusion: The analyst role shifts towards question framing, verification, context and communication. Analysts whose work is mainly the automatable tasks face real pressure; analysts who own the judgement tasks become more productive and, plausibly, more valuable.

Where the argument could be wrong

A careful argument names its weak points:

• If tools become reliable at verification, Premise 3 weakens. Today, tools still make confident errors on numbers; see analyst tasks you should not hand to an LLM.
• If companies cut headcount rather than expand analysis, Premise 4 fails locally. Some will.
• If business users query data directly with natural-language tools, fewer requests reach analysts — though someone must still maintain definitions and data quality behind those tools.

Hold the conclusion with the confidence it deserves: reasonable, not certain.

Score your own role

List the tasks you did in the last two weeks and score each. Be honest about time spent.

The scores are a rough guide, not a model. A quick calculation shows the share of your time in higher-exposure tasks:

tasks = [
    ("weekly report refresh", 6, 2, 2, 0),
    ("ad-hoc SQL requests",   5, 2, 1, 1),
    ("conversion drop dig",   4, 0, 0, 2),
    ("presenting findings",   2, 0, 0, 2),
    ("vendor file cleaning",  3, 1, 1, 1),
]
total = sum(t[1] for t in tasks)
exposed = sum(h for _, h, s, c, x in tasks if s + c - x >= 2)
print(f"{exposed}/{total} hours in higher-exposure tasks ({exposed/total:.0%})")
assert (exposed, total) == (11, 20)

In this example, 11 of 20 hours sit in higher-exposure tasks. That is not a verdict on the person; it is a map of where to change.

What to do with your score

Skills that hold their value

• Translating a vague business question into a measurable one.
• Knowing your company's data — definitions, quirks, where it breaks.
• Checking results: reconciliations, sanity checks, spotting impossible numbers.
• Explaining uncertainty and recommending an action.
• Using AI tools well, including knowing when not to.

The Data Analyst programme runs 3–4 months and teaches SQL, Power BI and Excel alongside GenAI-assisted analysis, with an emphasis on verifying AI output. GenAI and the analyst's daily tasks maps which everyday tasks to automate, assist or keep.