AI in 2025: an honest take from a data analyst

I use AI most days. I also distrust it most days. After a couple of years of folding these tools into real analytics work, I've landed somewhere unfashionably boring: AI is excellent leverage on top of good foundations and human judgement — and close to useless, sometimes worse, without them. Here's the honest version, from someone who ships numbers people make decisions on.

The discourse around AI tends to swing between two equally tiresome poles. One says analysts are finished; the other says it's all autocomplete and hot air. Neither matches my week. The truth is more specific and more useful: there are tasks where AI quietly saves me hours, and tasks where letting it lead would put a wrong number in front of a leadership team. The whole game is knowing which is which.

Where AI genuinely earns its keep

Start with the good news, because it's real. The places AI helps me most are the unglamorous ones — the plumbing, the first draft, the boilerplate that used to eat an afternoon.

Drafting SQL and pandas. When I know exactly what I want but don't want to hand-type a window function with three partition keys, AI gets me 80% of the way in seconds. I describe the shape — "rolling 28-day active users per market, deduplicated on account, ignoring internal test accounts" — and it produces a credible skeleton. I almost never ship its first version, but editing a draft is far faster than starting from a blank query.

Profiling unfamiliar data. Hand it a schema and a few sample rows and it's good at suggesting what to check first: which columns look like keys, where nulls probably hide, what a sensible distribution would be. It doesn't know my data — it pattern-matches against a million datasets it has seen — but as a "here's where to point your eyes" assistant, it's genuinely useful.

Turning numbers into narrative. This is the underrated one. I'll have a clean result set and a tired brain, and the gap between "the chart" and "the three sentences a director will actually read" is real work. AI is a strong rubber duck here — it drafts the framing, I correct the emphasis and strip the things that aren't true. The judgement stays mine; the blank page goes away.

Generating validation rules. Ask it to propose data-quality checks for a table and it will reliably surface the obvious-in-hindsight ones — referential checks, range bounds, freshness, uniqueness — that I'd otherwise write by hand. I treat the list as a checklist to prune, not a spec to trust.

Where it's overhyped — or quietly dangerous

Now the part the launch demos skip. The failure mode that costs you isn't AI refusing or stalling. It's AI being confidently, fluently wrong — and you not noticing because the answer looks right.

I've watched a model invent a column that doesn't exist, join on a field with the right name but the wrong grain, and silently double-count revenue because it didn't know two tables overlapped. None of these threw an error. All of them produced a number. A plausible, well-formatted, completely wrong number. In analytics, that's the dangerous shape: not the obvious failure, but the smooth one that survives a glance and lands in a deck.

There's a subtler cost too, and it's the one I worry about most for the long run: erosion of judgement. When a tool will always hand you an answer, it's tempting to stop forming your own first. The muscle that says "that number feels too high, let me sanity-check the join" is exactly the muscle AI makes easy to skip — and it's the muscle that separates an analyst from a copy-paste machine. Use the tool to move faster, not to think less.

And plenty of the hype is just hype. "Ask your data anything in plain English" demos beautifully and breaks the moment your schema has two columns that mean almost-but-not-quite the same thing, or a metric whose definition lives in three analysts' heads. The model can't read intent that was never written down. That gap isn't a model problem; it's a foundations problem wearing an AI costume.

Help here, humans there

So where do I actually draw the line? Roughly like this:

The dividing line isn't "easy versus hard." It's how expensive it is to verify. Where I can glance at the output and instantly know it's right, I let AI run. Where being wrong is cheap to catch, I lean in. Where a wrong answer is fluent, costly, and hard to spot — that's where a human stays firmly in the loop, and where "the model said so" is never an acceptable answer.

How I actually fold it into my week

In practice it's less of a philosophy and more of a habit. The workflow that keeps me honest looks like this:

• Frame it myself first. I decide what the question is and what a sane answer roughly looks like before I open the assistant. The expected order of magnitude is my tripwire.
• Let AI draft the tedious bit. The query skeleton, the transform, the test fixtures — whatever I'd recognise instantly but hate typing.
• Read every line as if a junior wrote it. I check the joins, the grain, the filters, the edge cases. Fluent ≠ correct, and I assume nothing.
• Validate against ground truth. Spot-check against a known number, a small manual count, or last period's figure. If it can't survive that, it doesn't ship.
• Own the output. Once it leaves my hands it's mine, not the model's. I have to be able to explain every number without saying "the AI did it."

None of that is exotic. It's just refusing to let speed quietly become carelessness.

So, is it overhyped?

Both things are true at once, which is what makes the honest answer unsatisfying. The grand claims — autonomous analysts, ask-anything dashboards, judgement-free insight — are oversold and will keep being oversold. The quiet, specific wins — fewer blank pages, faster drafts, less time on plumbing — are real and already in my workflow, and I'd be slower without them.

What hasn't changed is the part that was always the actual job: knowing what to measure, trusting the number, and being able to defend it. AI made the surrounding work cheaper. It made that core work, if anything, more valuable — because when everyone can generate a plausible answer in seconds, the scarce skill is telling a right one from a wrong one. That's the part I'm not handing over. How I use AI in practice is mostly a list of ways to keep it true.