30 AI Prompts for Data & Analytics
Copy-ready prompts that assume you will attach the dataset, the schema or the results — engineered to query faster, clean smarter and turn numbers into decisions.
Each prompt gives the AI an analyst persona, references your attached data or schema, and specifies the method and output so you get analysis you can trust and use. These prompts are tool-agnostic and optimised to work across all major AI assistants — Claude, Microsoft Copilot and ChatGPT (and Google Gemini). Replace the [bracketed placeholders] with your own details, and always review AI output before you act on it.
SQL and queries
Business-Question-to-SQL Query
senior data analyst
the table schema or CREATE TABLE statement for the relevant tables
- Confirm what business question the query must answer (e.g. monthly revenue by region for the last 12 months)
- Write a SQL query against the given schema
- Handle nulls, duplicates and ties explicitly
- Explain the query logic in plain English
- Flag any assumption made about the data
- Query must be readable, not just correct
- State the SQL dialect used (e.g. MySQL, PostgreSQL)
- Do not guess at ambiguous business rules without flagging them
The SQL query, followed by a plain-English explanation and a bulleted list of assumptions.
See example AI output
SELECT region, DATE_TRUNC('month', order_date) AS month, SUM(revenue) AS monthly_revenue FROM orders WHERE order_date >= CURRENT_DATE - INTERVAL '12 months' GROUP BY region, month ORDER BY region, month; This query (PostgreSQL dialect) sums revenue per region per month over the trailing 12 months, using DATE_TRUNC to bucket dates. Rows with NULL region are excluded — 214 orders (1.8% of total) fell into this bucket. Assumption: revenue already excludes refunds; if not, results overstate March by roughly 4%.
SQL Query Optimisation Review
query optimisation specialist
the slow-running SQL query you want reviewed
- Explain what the query currently does, step by step
- Identify the costly operations (full scans, nested subqueries, missing indexes)
- Suggest indexes or rewrites that preserve the same result
- Provide the optimised version of the query
- The result set must stay identical
- Explain the reasoning behind each change
- Note any trade-off (e.g. extra index storage)
A plain-English explanation of the original query, a list of bottlenecks, and the optimised SQL with inline comments.
See example AI output
The original query joins orders to customers then filters on customer.status, forcing a full scan of 2.1M rows before filtering. Main bottleneck: no index on customers.status, and a correlated subquery re-runs per row (est. 40s runtime). Suggested fix: add an index on customers(status), replace the subquery with a LEFT JOIN, and filter early. Optimised version reduces estimated runtime to roughly 3s. Trade-off: the new index adds about 80MB and slightly slows inserts on customers, acceptable given read-heavy usage.
Debug an Incorrect SQL Query
SQL debugger
the query, plus what result it should return versus what it actually returns
- Compare the expected result to the actual result
- Identify whether the flaw is in the join, grouping, filter or null handling
- Explain why the flaw produces the wrong output
- Provide the corrected query
- Reason about the underlying data, not just query syntax
- Confirm the fix against the original business intent
The identified flaw with a short explanation, followed by the corrected query.
See example AI output
The flaw is in the JOIN: an INNER JOIN between orders and promotions drops the 340 orders with no promotion, which is why total revenue undercounts by RM 18,200 (about 6%). It should be a LEFT JOIN so unpromoted orders are retained, with promotion fields defaulting to NULL/0. Corrected query switches to LEFT JOIN promotions p ON o.promo_id = p.id and wraps discount in COALESCE(p.discount, 0). Re-running against the sample confirms the total now matches the finance export.
Translate Business Question into SQL
translator from business question to query
no file - describe the business question and the relevant tables/columns
- Restate the business question in your own words
- Identify the grain (row-level meaning) and time window implied
- Flag any ambiguity in the question and how you resolved it
- Write the SQL query that answers it
- Confirm the grain and time window before finalising
- State interpretation choices explicitly, do not silently assume
A short statement of interpretation followed by the SQL query.
See example AI output
Interpreting 'top customers by growth' as customers whose spend increased most in Q2 2026 versus Q1 2026, ranked by absolute RM increase (not percentage, since new customers would dominate a percentage ranking). Grain: one row per customer. Query compares SUM(revenue) between the two quarters via a self-join on customer_id, computes the delta, and returns the top 20 by delta descending. Assumption: customers with zero Q1 spend are included as new-customer growth — flag if you'd rather exclude them.
Refactor SQL for Readability
code documenter
the existing SQL query to refactor
- Break the query into logical steps using CTEs
- Add comments explaining each CTE's purpose
- Rename cryptic aliases to descriptive names
- Verify the refactored query returns the same result
- Result set must be unchanged
- Structure must be clearer to a new team member, not just shorter
The refactored query using CTEs with inline comments.
See example AI output
WITH monthly_sales AS (-- Step 1: aggregate raw orders to month/region SELECT region, DATE_TRUNC('month', order_date) AS month, SUM(revenue) AS revenue FROM orders GROUP BY region, month), ranked AS (-- Step 2: rank regions within each month SELECT *, RANK() OVER (PARTITION BY month ORDER BY revenue DESC) AS rnk FROM monthly_sales) SELECT * FROM ranked WHERE rnk <= 3; Same output as the original nested subquery, now readable in two clear stages instead of one 40-line block.
Data cleaning and prep
Data Quality Assessment Report
data-quality engineer
a sample of the messy dataset (CSV or spreadsheet export)
- Scan the sample for duplicates, inconsistent formats, missing values, outliers and mixed types
- Identify the likely cause of each issue
- Recommend a fix for each issue found
- Prioritise issues by impact on downstream analysis
- Do not alter the data silently — flag every change needed
- Prioritise by impact, not by ease of fixing
A data-quality report table (issue, likely cause, fix, priority) followed by a short cleaning plan.
See example AI output
Found 4 issues: (1) 312 duplicate customer_id rows (18%), likely from repeated CSV exports — dedupe on id+timestamp, high priority. (2) date_joined mixes DD/MM/YYYY and MM/DD/YYYY formats — standardise to ISO 8601, high priority since it affects tenure calculations. (3) 5% of revenue values are negative, likely refunds miscoded — flag and separate into a refunds column, medium priority. (4) postcode field has both numeric and text entries — low priority, cosmetic. Recommended order: dedupe first, then dates, then revenue sign, then postcode.
Standardise Inconsistent Value Formats
data transformation assistant
a sample of the inconsistent values (e.g. date or currency formats)
- Identify every distinct format pattern present in the sample
- Write the transformation logic (formula or pseudo-code) to standardise them
- Explain the rule applied for each pattern
- Flag any value that doesn't fit a recognisable pattern
- Prefer reversible transformations where possible
- Explicitly flag ambiguous cases rather than guessing
The transformation logic/pseudo-code, with a short note on the rule and the edge cases to watch.
See example AI output
Detected 3 date patterns: DD/MM/YYYY (61%), MM-DD-YYYY (34%), and 'March 3, 2026' text form (5%). Pseudo-code: try parsing with locale=en-MY first (DD/MM/YYYY); if invalid, try MM-DD-YYYY; if still invalid, parse as long-form text; else flag as unparseable. Edge case: dates like 03/04/2026 are genuinely ambiguous between the two numeric formats — 23 rows fall in that overlap and should be manually verified against the source system rather than auto-converted.
Data Profiling Summary
profiling analyst
a sample of the dataset to profile
- Check for duplicate records
- Check for outliers in numeric columns
- Check for missing values by column
- Check for referential gaps (e.g. foreign keys with no match)
- Quantify how widespread each problem is
- Base every finding on evidence visible in the sample
- Do not assume causes not supported by the data
A profiling summary table with column, issue type, and percentage affected.
See example AI output
Profiled a 5,000-row sample: order_id — 0 duplicates. customer_email — 7.2% missing, concentrated in orders placed via the guest checkout flow. unit_price — 12 outliers (>RM 50,000) that look like currency-entry errors (missing decimal point). ship_region — 3.4% reference a region code not present in the regions lookup table, suggesting a stale mapping. Recommend prioritising the region mapping fix since it affects downstream reporting joins the most.
Missing Data Handling Recommendation
missing-data adviser
no file - describe the column, what it represents, and the percentage missing
- Assess whether the column should be dropped, imputed or flagged
- Explain the trade-off of each option
- State the impact each choice has on downstream analysis
- Recommend one approach with reasoning
- State the assumption each option makes about why data is missing
- Do not recommend imputation without naming the method
A short recommendation with the reasoning and the assumption behind it.
See example AI output
Column 'income_bracket' is 22% missing, likely because it's an optional survey field (missing-not-at-random risk if lower earners skip it). Dropping loses a fifth of usable rows — not ideal. Imputing with median assumes missingness is random, which may bias segment analysis toward higher earners. Recommended approach: flag missing values as a distinct 'Unknown' category rather than imputing, so the segment itself becomes analysable and no false precision is introduced. Revisit if missingness later drops below 5%.
Convert Unstructured Text to a Table
structuring assistant
the unstructured text block to convert (e.g. pasted notes, emails, or free-text responses)
- Infer the logical columns implied by the text
- Extract and standardise the values for each row
- List any row that could not be parsed with confidence
- Present the result as a clean table
- Preserve the original meaning, do not invent values
- Flag uncertainty rather than silently guessing
A structured table plus a short list of exceptions/rows that need manual review.
See example AI output
Parsed 40 free-text order notes into columns: customer_name, item, quantity, requested_date. 37 rows parsed cleanly. 3 exceptions: row 12 has no quantity mentioned ('a few boxes') — left blank and flagged; row 19's date 'next Friday' is relative and needs a reference date to resolve; row 31 mentions two items in one note and was split into two rows, marked as inferred. Recommend manually confirming the 3 flagged rows before loading into the system.
Analysis and insight
Top Three Insights for Decision-Makers
lead analyst
the dataset (or summary results) to analyse
- Identify the three most important, decision-relevant insights in the data
- Support each with the specific number behind it
- Recommend an action for each insight
- State your confidence level in each finding
- Insights must be significant, not trivial
- Distinguish correlation from causation explicitly
Three insights, each with the finding, the supporting number, the recommended action, and a confidence note.
See example AI output
1. Repeat-customer revenue grew 28% QoQ while new-customer revenue fell 6% — action: shift Q3 budget toward retention campaigns (high confidence, consistent across all 4 regions). 2. Klang Valley outlets show 15% higher basket size on weekends only — action: test weekend-specific bundles there before rolling out nationally (medium confidence, one region's sample is thin). 3. Support-ticket volume correlates with a recent app update but causation is unconfirmed — action: investigate before concluding the update is at fault (low confidence, needs a controlled comparison).
Pre-Conclusion Question Checklist
critical thinker
the data or analysis you're about to draw conclusions from
- List the likely biases in how the data was collected
- List gaps or missing context in the data
- Identify potential confounding variables
- State what would change the answer if it were different
- Be rigorous — the goal is to protect against a wrong conclusion
- Every question must be specific to this dataset, not generic
A question checklist grouped by bias, gaps, confounders, and sensitivity.
See example AI output
Bias: the survey was only sent to customers who opted into email, likely skewing toward more engaged/loyal customers. Gaps: no data on customers who churned before the survey window, so the 'satisfaction' number may be survivorship-biased. Confounders: satisfaction scores rose the same month a price cut launched — is it the cut or the newly rolled-out support chatbot? Sensitivity check: if the 200 lowest-scoring respondents were removed, does the headline 4.2/5 average still hold, or is it propped up by a small vocal group?
Diagnose Drivers Behind a Trend
diagnostic analyst
the trend data or chart you want explained (e.g. declining conversion rate over 6 months)
- List the likely drivers behind the trend
- Rank the hypotheses by likelihood
- For each hypothesis, state what data would confirm or rule it out
- Flag if the trend could be a data artefact rather than real
- Rank hypotheses, do not present them as equally likely
- Name the specific data needed to verify each
A ranked list of driver hypotheses, each with a verification step.
See example AI output
Most likely: checkout page load time increased 1.8s after the March redesign — verify by correlating page-speed logs with daily conversion rate. Second: a competitor launched a 20%-off campaign in the same window — verify via traffic-source split, checking if the drop is concentrated in paid-search visitors. Third (less likely): seasonal dip — verify against the same months last year; if last year showed no dip, seasonality is ruled out. Least likely: tracking or analytics misfire — verify by spot-checking raw event logs against the reported numbers.
Recommend the Right Analytical Method
methods adviser
no file - describe the business question and the data available
- Recommend the statistical or analytical approach that fits the question and data
- Explain why this method fits better than alternatives
- State the assumptions the method requires
- Suggest how to sense-check the result once produced
- Method must match both the data type and the question type
- Do not recommend a method whose assumptions the data clearly violates
The recommended method, the reasoning, the assumptions, and a sense-check suggestion.
See example AI output
For 'does the new onboarding flow increase 30-day retention', recommend a difference-in-differences comparison between the cohort that got the new flow and a matched control cohort, rather than a simple before/after comparison — this controls for seasonal retention trends. Assumes the two cohorts are otherwise comparable (similar acquisition channel mix) and that no other change launched simultaneously. Sense-check: run the same comparison on a placebo metric unrelated to onboarding — if that also shows a 'lift', something else is driving the result.
Plain-English Interpretation for Business Readers
translator for non-technical audiences
the analysis results to interpret (paste the numbers, chart description, or summary stats)
- Explain what the result means in plain language
- Explain what the result does NOT mean (common misreading to avoid)
- State the one action the result actually supports
- Note the level of uncertainty in the finding
- No statistical jargon
- Be honest about uncertainty rather than overstating confidence
A short plain-English interpretation paragraph plus the one supported action.
See example AI output
What it means: customers who used the mobile app in their first week were about twice as likely to still be active after 3 months. What it does NOT mean: the app itself necessarily caused this — customers who download the app in week one may simply be more engaged to begin with, so we can't yet say the app 'creates' loyal customers. Supported action: prioritise nudging new customers toward the app in week one and measure the effect with an actual test, rather than rolling out a costly app-first strategy on this evidence alone.
Visualisation and dashboards
Chart Type Recommendation
data-viz specialist
no file - describe the relationship you want to show (e.g. sales over time by product)
- Recommend the best chart type for this relationship
- Explain why this chart type fits better than common alternatives
- Specify the encoding — axis, colour, ordering
- Flag the one design choice that would mislead if done wrong
- Prioritise clarity over decoration
- Justify the recommendation against the specific data, not generically
The recommended chart type, the encoding spec, and a caution about a common mistake.
See example AI output
For sales over time by product (5 products, 24 months), recommend a multi-line chart over a stacked area chart — lines let you compare individual product trajectories without the visual distortion stacking introduces. Encoding: time on x-axis, revenue on y-axis, one colour per product with a direct end-of-line label instead of a legend. Caution: do not start the y-axis above zero — with only a 15% range between products, a truncated axis would visually exaggerate differences that aren't actually large.
Dashboard Specification Design
dashboard designer
no file - describe the dashboard's purpose and its primary audience
- Recommend the key metrics the dashboard must show
- Define the layout — what goes top-left through to bottom-right
- Recommend the filters the audience will need
- Confirm the dashboard answers the audience's main question at a glance
- No vanity metrics that don't drive a decision
- The audience's main question must be answerable within 5 seconds of looking
A dashboard spec: metrics list, layout description, and filter list.
See example AI output
For a weekly ops dashboard for store managers: top-left shows this week's revenue vs. target (single big number, red/green), top-right shows footfall trend (sparkline, 12 weeks). Middle row: top 5 and bottom 5 SKUs by sell-through. Bottom: staffing hours vs. sales-per-labour-hour. Filters: store selector and date-range toggle (week/month). Deliberately excluded: total lifetime sales and follower counts — neither changes what a manager does this week.
Dashboard Critique and Redesign
viz reviewer
a description or screenshot of the existing dashboard
- Flag any misleading scales or truncated axes
- Flag chart-type mismatches (wrong chart for the data)
- Flag visual clutter that obscures the main message
- Propose the specific fix for each issue
- Every critique point needs a specific, actionable fix
- Prioritise issues that could lead to a wrong decision
A critique list paired with the improved design for each point.
See example AI output
Issue 1: the revenue bar chart's y-axis starts at RM 80,000 not 0, making a 4% dip look like a 40% collapse — fix: rebase to zero or add a clear axis-break marker. Issue 2: using a pie chart for 9 product categories is unreadable — fix: switch to a sorted horizontal bar chart. Issue 3: 6 KPI cards compete for attention with no hierarchy — fix: enlarge the single most important metric (net margin) and demote the rest to a smaller secondary row.
DAX or Spreadsheet Formula Builder
formula writer
no file - describe the metric you want to calculate and the relevant columns/tables
- Write the DAX or spreadsheet formula to calculate the metric
- Explain what each part of the formula does
- Note the edge case that could break the formula
- Confirm the formula adapts if the underlying table grows
- Formula must be correct against the stated definition of the metric
- Must be adaptable, not hard-coded to current row counts
The formula, a line-by-line explanation, and the edge case to watch.
See example AI output
For 'rolling 3-month average revenue per region': Rolling3M = AVERAGEX(DATESINPERIOD('Date'[Date], MAX('Date'[Date]), -3, MONTH), CALCULATE(SUM(Sales[Revenue]))). DATESINPERIOD sets the trailing 3-month window relative to the current filter context; AVERAGEX then averages monthly totals within it. Edge case: for the first 2 months of data, the window is incomplete and will silently average over fewer months — consider a visual flag (e.g. greyed-out) for periods with less than 3 months of history.
Single-Message Visual Design
storytelling designer
the data you want to visualise (paste the figures or describe the dataset)
- Identify the single main message the data should communicate
- Recommend the chart that makes that message obvious at a glance
- Suggest the annotation that reinforces the message
- List what to strip away from the chart
- One message per chart — do not try to show everything
- Annotation must point at the message, not just label the data
The main message, the chart recommendation, the annotation text, and the elements to remove.
See example AI output
Main message: marketing spend efficiency has doubled since switching channels in Q1. Chart: a simple line of cost-per-acquisition over 8 months, with a vertical dashed line marking the Q1 channel switch. Annotation directly on the chart: 'Channel switch — CPA fell from RM 42 to RM 19.' Strip away: gridlines, the secondary axis showing total spend (distracting from the efficiency story), and the legend (only one line, doesn't need one).
Reporting and storytelling
Story-Driven Leadership Summary
data storyteller
the analysis results to turn into a narrative
- Set the context briefly
- State the key finding
- Explain the so-what — why it matters to the business
- End with the recommendation
- Lead with the insight, not the methodology
- Every claim must be tied to a specific number
A short narrative structured as context, finding, so-what, recommendation.
See example AI output
Context: we tracked customer retention across all 6 regions since the loyalty programme launched in January. Finding: enrolled customers show 34% higher 90-day retention than non-enrolled customers (68% vs 51%). So-what: at current enrolment (22% of the base), this is already worth an estimated RM 310,000 in retained revenue this quarter, with most of the gap still unrealised. Recommendation: fund a targeted enrolment push for the remaining 78%, starting with the two regions where enrolment lags most — Penang and Johor.
Executive Summary of an Analysis
executive-summary writer
the full analysis to summarise
- Identify the so-what and the decision it supports
- Write the summary leading with that so-what
- Cut anything that is methodology rather than conclusion
- Under 150 words
- No methodology dump — decision-makers want the conclusion
A single executive summary paragraph under 150 words.
See example AI output
Customer churn in the SME segment has risen from 4.1% to 6.8% monthly since Q4, driven mainly by price-sensitive accounts under RM 500/month spend — this tier accounts for 71% of the churn increase. If unaddressed, this trend puts roughly RM 480,000 in annual recurring revenue at risk over the next two quarters. Recommend a targeted retention offer for this tier before the next renewal cycle in September, alongside a pricing review — the current tier appears priced above what this segment tolerates. Full analysis and segment breakdown attached for reference.
One-Line Chart Annotations
chart annotator
the report with charts that need annotation (describe or paste each chart)
- For each chart, identify what it shows
- Write a one-line annotation stating the takeaway, not just a description
- Ensure the annotation would make sense even without reading the whole report
- Takeaway, not description — say what it means, not what it is
- One line per chart, no exceptions
A list of chart annotations, one per chart.
See example AI output
Chart 1 (monthly revenue by channel): organic search overtook paid ads as the top revenue channel in May, and the gap is widening. Chart 2 (customer age distribution): the 25-34 segment now makes up 44% of customers, up from 31% a year ago — the core demographic has shifted younger. Chart 3 (support ticket resolution time): resolution time improved 22% after the new ticketing workflow launched in April, closing most of the gap to the 24-hour target.
Three Headline Findings
insight distiller
the data or analysis to distil
- Identify the three most memorable, decision-relevant findings
- Write each as one sentence including the supporting number
- Order them by importance
- Must be memorable — a leader should recall it later
- Must remain accurate, not simplified to the point of being misleading
Three headline sentences, each with a number, ordered by importance.
See example AI output
1. Repeat customers now generate 61% of revenue, up from 48% two years ago. 2. Average delivery time in East Malaysia is 2.3 days slower than West Malaysia, the single biggest driver of low ratings there. 3. The referral programme costs 40% less per acquired customer than paid social, but currently accounts for only 8% of new signups.
Findings Talking Points and Q&A Prep
presentation coach
the findings you need to present and defend
- Draft talking points to present the findings clearly
- Anticipate the three hardest challenges someone might raise
- Draft the response to each challenge
- Be honest about the analysis's limitations
- Responses must be substantive, not deflective
Talking points followed by a three-question Q&A prep list.
See example AI output
Talking points: open with the RM 2.1M revenue impact, then the driver (checkout redesign), then the recommended fix. Q&A prep — Q1: 'Could this just be seasonality?' A: we compared against the same period last year and the pattern doesn't appear, ruling seasonality mostly out. Q2: 'Sample size for East Malaysia stores?' A: only 4 stores, so treat that regional finding as directional, not conclusive — flagged in the appendix. Q3: 'What did the fix cost to test?' A: the A/B test ran for 3 weeks at no incremental cost, using existing traffic.
Modelling and productivity
Plain-Language Method Explainer
methods teacher
no file - describe the technique and the problem you're applying it to
- Explain in plain language when to use this technique versus the alternative
- State the assumptions the technique requires
- Flag the common mistake practitioners make with it
- Keep it practical and decision-oriented, not academic
- Ground the explanation in the specific problem given
A short plain-language guidance note covering when to use it, assumptions, and the common mistake.
See example AI output
For predicting 'will this customer churn' (yes/no) versus 'how much will this customer spend next month' (a number) — use classification (e.g. logistic regression) for the first, regression for the second; the mistake is treating churn as a number between 0 and 1 and running linear regression on it, which produces predictions above 1 or below 0. Classification also needs a class-imbalance check: if only 5% of customers churn, plain accuracy will look great while the model predicts 'no churn' for everyone — check precision and recall instead.
Feature Engineering Suggestions
feature-engineering adviser
the dataset (list of available columns) and the outcome you're trying to predict
- Suggest features that might predict the outcome
- Explain the rationale behind each suggested feature
- Flag the risk of leakage or bias in each
- Justify every feature against the outcome, not just intuition
- Explicitly call out any feature that risks leaking future information
A feature list, each with the rationale and a leakage/bias flag.
See example AI output
Predicting 'will customer churn next month': (1) days_since_last_order — strong rationale, recency is a classic churn signal, low leakage risk. (2) total_support_tickets_last_90d — rationale: friction predicts churn, low risk. (3) cancellation_reason_code — high leakage risk, this field is often only populated after a customer has already churned, exclude it. (4) customer_tenure_months — rationale: newer customers churn more, watch for bias against a recently-launched acquisition channel with naturally short tenure so far.
Peer Review of an Analysis Approach
peer reviewer
a description of the analysis approach used (sample, method, assumptions)
- Check whether the sample is representative of the population in question
- Check whether the method fits the question and data
- Check whether the assumptions hold
- Suggest specific improvements
- Be rigorous but constructive
- Every flaw flagged needs a suggested fix, not just criticism
A structured review covering sample, method, assumptions, and improvement suggestions.
See example AI output
Sample: the survey only reached customers with a valid email, excluding roughly 30% of the base who signed up via phone — this likely skews results toward a more digitally engaged segment; suggest weighting or a follow-up phone survey. Method: a simple average was used for satisfaction scores across regions with very different sample sizes (Sabah n=12 vs Selangor n=340) — recommend a weighted average or reporting regional confidence intervals. Assumptions: the analysis assumes responses are independent, but the same customers were surveyed twice — deduplicate before aggregating.
Pandas Pseudo-Code for a Data Task
pandas assistant
no file - describe the dataframe's columns and the transformation you need
- Write pandas pseudo-code to accomplish the task
- Explain each step in the code
- Note the edge case or gotcha to watch for
- Code must be readable, not just short
- Correctness matters more than cleverness
The pandas pseudo-code with inline comments, plus a short explanation and gotcha note.
See example AI output
df has columns customer_id, order_date, revenue. df['order_date'] = pd.to_datetime(df['order_date']); monthly = df.groupby([pd.Grouper(key='order_date', freq='M'), 'customer_id'])['revenue'].sum().reset_index(). This converts order_date to a real datetime, then groups by calendar month and customer to sum revenue. Gotcha: pd.Grouper with freq='M' labels each group by month-end date, not month-start — reformat with .dt.strftime('%Y-%m') if you want a cleaner month label for reporting.
Data Dictionary for a Field Set
documentation writer
the list of fields/columns to document (paste the field names and any known notes)
- Define the meaning of each field
- State the data type of each field
- List the allowed values or valid range for each field
- Note any caveat (edge case, known data quality issue, deprecation)
- Definitions must be clear enough for someone outside the team to use correctly
- Cover every field given, no omissions
A data dictionary table: field, meaning, type, allowed values, caveat.
See example AI output
customer_tier — meaning: loyalty tier assigned at signup; type: string; allowed values: Bronze, Silver, Gold; caveat: legacy records before 2024 use Standard instead of Bronze, not yet backfilled. order_status — meaning: current state of an order; type: string; allowed values: Pending, Shipped, Delivered, Cancelled; caveat: Cancelled does not distinguish customer-initiated from system-initiated cancellations. region_code — meaning: 2-letter Malaysian state code; type: string; allowed values: standard ISO state abbreviations; caveat: 3% of rows have an XX placeholder from an incomplete migration.
More prompt packs by function
View the full Prompt LibraryFrequently Asked Questions
Each gives the AI an analyst role, assumes you will attach the schema, dataset or results, and specifies the method and output. That produces correct queries, real insight and clear reporting — and the prompts explicitly ask the AI to state assumptions and separate correlation from cause so you can trust the result.
Yes — the prompts are written to work from your attached schema, data samples and results. Claude and ChatGPT can even run analysis on uploaded files. Never paste confidential or personal data into public consumer tools; use approved enterprise AI and anonymise where you can.
Copilot is powerful inside Excel and the Microsoft data stack; Claude and ChatGPT are excellent for SQL, analysis and explanation, and both can run code in their analysis modes. These prompts work across all of them.
Yes — more than ever for judgement. AI writes queries and suggests methods, but you need the understanding to know if the answer is right and to ask the correct questions. AI amplifies skilled analysts; it does not replace the thinking.
Yes. AITraining2U's AI Analytics course is HRD Corp SBL-KHAS claimable for eligible Malaysian employers.
Not without verification. Large language models can state a percentage or total with total confidence even when it's fabricated or miscalculated — this is a known limitation, not a rare glitch. Treat any AI-generated number as a first draft: ask the AI to show its calculation or the exact rows it used, then spot-check against the source data yourself. For anything going into a board report, regulatory filing or client invoice, always recompute the key figures independently before relying on them. AI is excellent at explaining and structuring analysis; it is not yet a reliable calculator on its own.
It depends on the plan and settings, not the tool name. Free consumer accounts may use your inputs to train future models unless you opt out; paid business tiers (ChatGPT Team/Enterprise, Claude for Work, Copilot with Microsoft 365 commercial data protection) typically don't. For Malaysian SMEs handling customer PDPA-covered data, use an enterprise-tier account, strip identifying fields (names, IC numbers, phone numbers) before uploading where possible, and check your data-processing agreement. When in doubt, anonymise the data or use aggregated summaries instead of raw customer-level exports.
No — they solve different problems. Dashboards give you a live, governed, always-on view of the numbers that doesn't depend on someone asking the right question. AI chatbots are best for ad-hoc exploration, explaining a spike, drafting a query, or turning a chart into a narrative — work a static dashboard can't do. The strongest setup pairs them: keep your dashboard as the single source of truth for daily monitoring, and use AI on top of its exports or underlying data for one-off deep dives and reporting write-ups.
Ask the AI to show its work — the exact filter, formula or rows behind the number — rather than accepting the headline figure alone. Then recompute it yourself in Excel, SQL or a pivot table using the same definition, and confirm the two match. Watch especially for mismatched time windows, double-counted rows, and percentages calculated on the wrong base. If the AI can't reproduce its own working when asked, treat the number as unverified and don't put it in front of clients, management or regulators until you've confirmed it independently.
Start with what your data already sits in. If your numbers live mostly in Excel and SharePoint, Copilot's native integration makes it the fastest win with the least data movement. If your team writes SQL or works with exported CSVs and needs strong reasoning on messy analysis, Claude or ChatGPT tend to explain and debug more thoroughly. Many Malaysian SMEs end up using both: Copilot for day-to-day spreadsheet work, and Claude or ChatGPT for deeper one-off analysis — pick based on where your analysts already work, not brand preference.
Go beyond prompts — train your team
Prompts are the start. AITraining2U runs hands-on, HRD Corp SBL-KHAS claimable AI training for Malaysian teams — from everyday AI productivity to building agents that run data & analytics workflows end-to-end.