The technical core still matters, just differently
You still need SQL, Excel, and one BI tool. What changed is that the bar moved from 'can you build it' to 'can you verify it.' In an AI-heavy workplace the model drafts the query, so your value is catching the error before a manager sees it. The core skills are the lens you use to check the draft.
Think of the core as the grammar. AI gives you fluent sentences, but if you don't know the grammar you can't tell a fluent lie from a fluent truth. That is why SQL practice still pays, even when a bot writes half your queries for you.
The practical way to keep the core sharp is to use it every week even when AI drafts for you. I still write one query by hand each Monday so the syntax stays in my fingers. A common mistake is letting the bot do everything until the day an interview asks you to type live, and your hands have forgotten the JOIN.
Skill 1: validate model output by hand
When the assistant returns a total, pick one region and one month and sum the raw rows in Excel. If it says West March is $85,000 and your manual sum is $81,300, you found a refunds assumption. This 5-minute check is the single most hired-protecting habit.
Run COUNT(*) with the same filter before any aggregation. If you expect about 12,000 orders and get 2 million, a join key is wrong. A common mistake is trusting the pretty chart and never counting the rows underneath it.
Read the WHERE clause and restate the period in your own words next to the query. A model that used 2025 instead of 2026 makes every number wrong but plausible. Stating 'Q1 2026' in writing lets a teammate catch the slip.
I recommend you keep a one-line validation note with every deliverable: 'checked West March by hand, matched to $81,300.' In my experience, managers who see that note stop second-guessing you and start forwarding your work upward.
Skill 2: write precise prompts for analysis
Open with the real schema: 'Table Orders has order_id, region, amount, order_date.' A vague prompt like 'analyze my sales' makes the model invent columns that don't exist, and you waste an hour mapping them back to your database.
Write 'one row per region per month, total amount' so the model doesn't return one annual row for the whole company. Specifying the grain is the difference between a usable query and a confident wrong one.
Request both the query and a one-line description of each clause. Forcing the model to justify its join in words makes your review step faster and exposes bad assumptions on the spot. See our AI tools for data analysis overview for reusable templates.
Prompting is not a soft skill you either have or don't. It is a checklist you reuse. The analysts who look natural at it just saved the checklist and stopped rewriting it each time.
Skill 3: read SQL you didn't write
When you inherit a query, list each ON clause and confirm the keys. Orders.customer_id = Customers.id is fine; joining on a name field multiplies rows and inflates revenue by 300 percent. Reading the join is faster than debugging the chart later.
Ask the model or a teammate: 'What does this query assume about nulls and refunds?' Those assumptions are exactly where your report will disagree with finance, and they are invisible until you name them out loud.
If you can't explain a query in one sentence, you don't understand it yet. Practice with SQL for data analysis exercises until explaining a window function feels normal. That fluency is what interviews test when they ask you to read code live.
A quick drill that builds this skill: take any query from your work, cover the SELECT line, and predict the columns before you look. If you're wrong, you didn't understand the join. Do this ten times and reading inherited SQL stops feeling like decryption and starts feeling like reading.
Avoid shipping a query you can't explain just because the bot wrote it cleanly. This will break the first time a stakeholder asks 'why did the Southeast region drop to zero' and you have no answer. If you can't explain it, don't publish it.
Skill 4: explain numbers to non-analysts
Open with the decision: 'We should cut the discount, it's leaking $40,000.' Then put the method as a footnote. A common mistake is leading with 'I used a window function,' which loses the manager in sentence one and buries the insight.
Pick the single visual that carries the point and drop the rest. In an AI-heavy workplace where anyone can generate ten charts, the analyst who shows one right chart looks senior. Restraint reads as confidence to a non-analyst audience.
Before any meeting, write the finding, the evidence, and the ask in three sentences. Say them out loud. A hiring manager at a logistics firm told me they decide on candidates in the first three sentences of a presentation, not the deck.
This skill is the one AI can't fake. A model can write the chart; it can't stand in the room and feel the room. The analysts who get promoted are the ones the manager trusts to translate, not just to compute.
Skill 5: decide what to hand to AI
Hand the bot the boilerplate: first query, formula, cleanup map. Keep for yourself the judgment: which question, which assumption, which number to trust. A useful rule is 'AI drafts, I decide.' That split protects both your time and your name.
Keep HR, finance, and customer-PII rows inside enterprise tools with data terms from Microsoft or Google. A free public chat that trains on your rows is a compliance incident, not a shortcut. Know your boundary before you paste anything.
Save AI help for the report you run every Monday. For a one-off strange request, do it by hand so you stay sharp. Automating the novel task before you understand it is how errors hide in the pipeline for months.
A simple test for what to delegate: if a mistake would be caught by a formula, let the bot draft it; if a mistake would be a wrong business call, you do it. Sorting tasks this way took me from racing the clock to actually owning the analysis. In my experience the clarity alone cuts rework in half.
I recommend you write down one sentence per task: 'draft' or 'decide.' In my experience, analysts who never make that call end up letting the bot decide by default, and the bot decides on speed, not on what the business needs.
A weekly plan to build these skills
Re-check a report you already shipped by hand. If it still holds, you practiced the habit. If you find a slip, you just saved next month's meeting. Either way Monday buys you credibility.
Turn a recurring ask into a saved prompt with table, grain, and metric stated. Reuse it Friday. By week four you have four templates and a faster week, and you've stopped rewriting the same English every time.
Fifteen minutes twice a week beats a weekend cram. The goal is reflex, not trivia. When the habit is automatic, the AI-heavy workplace feels like an upgrade instead of a threat.
Next steps for your skill stack
Pick the five skills above and rate yourself honest: draft or decide on each. Most analysts are strong on the core and weak on validation and explanation. Close those two gaps first — they are the ones recruiters screen for in 2026. Pair your practice with AI-assisted SQL and the Excel for data analysis basics, then benchmark your target on our 2026 salary guide.
