Eighteen months of LLMs in the analyst stack have been enough to separate the workflows where AI is genuinely load-bearing from the ones where it’s a demo that doesn’t survive contact with a real engagement. The marketing has not caught up to the reality. Here’s the honest read.
Where LLMs are real
SQL drafting against a known schema. This is the cleanest win. Given a schema dump, a few sample rows, and a clearly stated business question, a frontier model produces correct, often well-structured SQL on the first or second try. CTEs are formatted, window functions are used appropriately, and the joins respect grain. The analyst’s job shifts from typing to reviewing — checking that the join keys reflect the actual entity relationships, that the date logic matches the warehouse’s timezone conventions, and that the aggregations roll up at the level the stakeholder actually wants. That review step is non-negotiable, but the time savings on the drafting half are real and consistent. We measure 40–60% faster cycles on net-new exploratory queries.
Code review on transformation logic. dbt models, Python notebooks, and PySpark jobs benefit from a second pair of eyes that never gets tired. LLMs catch the bugs that humans pattern-match past: silent type coercions, off-by-one window boundaries, joins that fan out, COUNT(*) where COUNT(DISTINCT id) was intended. They miss the semantic bugs — “this metric is technically correct but doesn’t measure what the business calls ‘churn’” — but they catch the mechanical ones at a rate that genuinely changes the defect curve. Pair this with a working CI suite and the floor rises.
Narrative writing from validated findings. Once the analysis is done and the numbers are signed off, turning bullet-point findings into a coherent executive narrative is something LLMs do well. They produce prose that is structurally sound, appropriately hedged, and free of the awkward phrasings that creep into deck-writing at 10pm. The critical word is validated. The model is a writer, not an analyst. If you ask it to interpret findings, it will confabulate causality and oversell magnitude. If you give it a clean set of conclusions and ask it to render them in a particular voice for a particular audience, it works.
Exploration scaffolds. When you’re entering a new dataset cold, asking an LLM to propose a starter EDA — “what would you check first in a transactions table with these columns?” — produces a useful checklist of distributions to plot, integrity checks to run, and edge cases to verify. It’s not a substitute for thinking, but it lowers the activation energy of getting started, and that matters when the alternative is procrastination disguised as “scoping.”
Where LLMs still don’t help
Judgment under ambiguity. The hardest part of analytics is not the SQL. It’s deciding what question to answer when the stakeholder has asked the wrong one, or asked the right one with the wrong success metric, or asked a question whose true answer would be politically inconvenient. LLMs do not push back. They do not say “your framing assumes a causal relationship the data can’t support.” They produce confident output against whatever prompt they receive, including prompts that contain unexamined assumptions. A junior analyst will sometimes catch this. A senior will catch it most of the time. A model will not catch it at all, because it has no stake in the outcome.
Scoping a real engagement. Translating a vague executive ask (“we need to understand customer health”) into a tractable analysis plan with defined deliverables, data sources, and a timeline is judgment-heavy work that depends on knowing the business, the data warehouse, the political landscape, and the half-dozen prior attempts that failed for reasons no one wrote down. LLMs produce generic scoping documents that read fine and would not survive a 30-minute conversation with the actual VP of Customer Success.
Stakeholder alignment. The work of getting three executives, two product managers, and a finance partner to agree on what a metric means and how it should be governed is fundamentally about humans, trust, and the calibration of language across roles that don’t share vocabulary. AI can draft the meeting notes. It cannot do the meeting.
Novel methodology. When a problem requires methodology that isn’t already well-represented in the model’s training data — a custom attribution scheme, a non-obvious experimental design, a statistical approach that fits the specific structure of your data — LLMs produce plausible-sounding nonsense. They will cheerfully suggest a Bayesian hierarchical model when a difference-in-differences is what’s called for, or vice versa, with equal confidence. Methodological judgment is where senior analysts earn their keep, and it’s the area where LLMs have improved least over the past year.
What this implies for how teams operate
The teams getting the most leverage out of LLMs in 2026 are not the ones with the loudest “AI strategy.” They’re the ones who have quietly reorganized so that drafting, review, and rendering work is AI-assisted, while scoping, judgment, and stakeholder work remains human-led — and who treat the boundary between those zones as something to defend rather than blur.
The analyst who tries to use the model as an oracle gets bad analyses faster. The analyst who uses it as a tireless junior who needs supervision gets good analyses faster. That distinction is the entire game.