BXUP Track progress & test yourself — free
AI at Office · Hands-on lab · ~5 min read
AI AT OFFICE · HANDS-ON LAB

Lab 03 · Sales Data Detective

Clean a messy sales export with AI, answer five business questions in plain English, and build three honest charts — reconciling every AI-computed figure against a manual spot-check, because the export contains data-entry errors and a hidden pattern, and the AI will confidently compute right answers on top of wrong data.

Kavya Iyer is the sales operations coordinator at Sunbeam Beverages Pvt Ltd, a fictional regional drinks distributor. The national sales head wants "a quick read" on last year's order data before Thursday's planning meeting: what's selling, where, who's performing, and anything odd. The export below came straight out of the order system — which means it came out dirty. Two entries are wrong in ways that will silently corrupt any total that includes them, and there is one genuine pattern hiding in the dates that the sales head will want to know about. AI data analysis is the fastest way through this — and also the fastest way to a confident wrong answer, which is why every number you report gets a manual spot-check.

You will need

  • An AI that can analyze dataChatGPT (upload the CSV, free tier works), Claude (paste it), or Copilot in Excel / Gemini in Sheets.
  • A spreadsheet (Excel or Google Sheets) open beside the AI — this is your reconciliation bench.
  • Timeabout 75 minutes.

The data

Treat this 36-row extract as the full export for the exercise. Save it as sunbeam_orders.csv.

order_id,order_date,region,rep,product,units,unit_price_inr
SB-1001,2025-01-14,West,Ramesh,Mango Fizz 500ml,120,32
SB-1002,2025-01-20,North,Alka,Lime Sparkle 500ml,80,30
SB-1003,2025-02-03,South,Joseph,Mango Fizz 500ml,150,32
SB-1004,2025-02-11,West,Ramesh,Cola Classic 1L,200,55
SB-1005,2025-02-25,North,Alka,Mango Fizz 500ml,90,32
SB-1006,2025-03-09,South,Divya,Lime Sparkle 500ml,140,30
SB-1007,2025-03-18,West,Suresh,Cola Classic 1L,180,55
SB-1008,2025-03-30,Nroth,Alka,Cola Classic 1L,110,55
SB-1009,2025-04-12,South,Joseph,Mango Fizz 500ml,260,32
SB-1010,2025-04-19,West,Ramesh,Mango Fizz 500ml,310,32
SB-1011,2025-04-27,North,Prakash,Lime Sparkle 500ml,150,30
SB-1012,2025-05-08,South,Divya,Mango Fizz 500ml,340,32
SB-1013,2025-05-16,West,Suresh,Mango Fizz 500ml,380,32
SB-1014,2025-05-24,North,Alka,Cola Classic 1L,160,55
SB-1015,2025-06-05,South,Joseph,Mango Fizz 500ml,290,32
SB-1016,2025-06-14,West,Ramesh,Lime Sparkle 500ml,170,30
SB-1017,2025-06-500,North,Prakash,Mango Fizz 500ml,210,32
SB-1018,2025-07-02,South,Divya,Cola Classic 1L,150,55
SB-1019,2025-07-15,West,Suresh,Cola Classic 1L,190,55
SB-1020,2025-07-28,North,Alka,Lime Sparkle 500ml,120,30
SB-1021,2025-08-09,South,Joseph,Cola Classic 1L,170,55
SB-1022,2025-08-21,West,Ramesh,Cola Classic 1L,210,55
SB-1023,2025-09-04,North,Prakash,Cola Classic 1L,180,55
SB-1024,2025-09-17,South,Divya,Lime Sparkle 500ml,130,30
SB-1025,2025-09-29,West,Suresh,Mango Fizz 500ml,160,32
SB-1026,2025-10-08,North,Alka,Cola Classic 1L,320,55
SB-1027,2025-10-15,South,Joseph,Cola Classic 1L,360,55
SB-1028,2025-10-22,West,Ramesh,Cola Classic 1L,410,55
SB-1029,2025-10-29,North,Prakash,Cola Classic 1L,380,55
SB-1030,2025-11-06,South,Divya,Cola Classic 1L,350,55
SB-1031,2025-11-14,West,Suresh,Cola Classic 1L,-390,55
SB-1032,2025-11-21,North,Alka,Mango Fizz 500ml,140,32
SB-1033,2025-12-04,South,Joseph,Lime Sparkle 500ml,110,30
SB-1034,2025-12-12,West,Ramesh,Cola Classic 1L,230,55
SB-1035,2025-12-12,West,Ramesh,Cola Classic 1L,230,55
SB-1036,2025-12-19,North,Prakash,Mango Fizz 500ml,100,32

The five questions from the sales head:

  1. What was total revenue for the year, and which region contributed most?
  2. Who is the top rep by revenue, and by how much?
  3. Which product drives the most revenue, and is that consistent across regions?
  4. Is there any seasonality we should plan stock around?
  5. Is there anything in the data itself we should worry about?

Steps

  1. Profile before you compute. Give the AI the CSV with: "Before answering any business questions: profile this data. List every distinct value in region and product, the date range, min and max of units, and any rows that look invalid, impossible, or duplicated. Do not fix anything — report only." A profiling pass is how professionals meet a dataset; totals come later. Read the profile line by line against the raw CSV.

  2. Rule on every flagged row yourself. For each row the AI flagged — and any you spot that it missed — decide the treatment: correct it (only when the right value is obvious from context), exclude it (with a note), or keep it. Write a cleaning log: row, problem, decision, reason. There are more problems in this file than any single profiling pass typically catches, so scan the date, region, and units columns yourself before closing the log.

  3. Produce the clean dataset. Instruct: "Apply exactly these cleaning decisions [paste your log] and output the cleaned CSV. Tell me the final row count and how many rows were changed or dropped." Load the cleaned CSV into your spreadsheet. Confirm the row count matches what the log predicts — if the AI dropped or altered anything beyond your instructions, that is a finding about the tool, and it goes in your notes.

  4. Answer questions 1–3, with reconciliation. Ask the AI each revenue question against the cleaned data, requesting the arithmetic: "Show revenue as units × unit_price, give per-region and per-rep tables, and state which rows are included." Then reconcile in your spreadsheet: compute total revenue yourself with a formula (SUMPRODUCT of units and price), and hand-check one region and one rep by filtering. The rule: no AI-computed figure enters your report until your spreadsheet reproduces it. If they disagree, the spreadsheet wins, and you find out why.

  5. Hunt the seasonality. Ask: "Group cleaned revenue by month and by product-month. Is there a seasonal pattern? Which products and regions drive it?" Verify the shape yourself: sort the spreadsheet by month, subtotal, and eyeball it. Then push one level deeper than the AI's first answer — which product creates the spike, and is the spike a one-month event or a quarter-long ramp? Planning stock needs the second answer, not the first.

  6. Build three charts that each say one thing. In your spreadsheet (or ask the AI for the chart, then rebuild it): monthly revenue trend (line), revenue by region (bar), and product mix in the peak months vs the rest of the year (two bars or a grouped bar). Each chart gets an assertion title stating its finding — "Festive quarter drives the year" — not a label title like "Revenue by month". If a chart needs a paragraph to explain, it is two charts.

  7. Write the one-page read. Five numbered answers, one chart reference each where relevant, and a final section headed "Data quality" listing what you found in the raw export, what you excluded, and what should be fixed at the source (order-entry validation). The data-quality section is question 5's real answer — the sales head asked "anything odd" and the honest response covers both the pattern and the pipeline.

  8. The reconciliation stamp. Add one line at the bottom of the report: "All figures reconciled against manual spreadsheet computation on [date]; cleaning log available." Then actually make it true — re-verify your headline total one last time. This line is the difference between analysis and vibes, and once you start shipping it, people notice.

Remember:

Chat models miscount — when a number matters, make the tool compute it with real code or a spreadsheet, and reproduce it yourself before it leaves your hands.

Verify your work

  • A profiling report and a cleaning log with a ruling on every flagged row — including rows the AI missed.
  • Cleaned dataset with a row count that matches the log.
  • Total revenue, top region, top rep, top product — each reproduced independently in your spreadsheet before reporting.
  • Seasonality identified at product level with the months named, not just "Q4 is higher".
  • Three charts with assertion titles, each making exactly one point.
  • A data-quality section that names the source-system fixes, and the reconciliation stamp on the report.
▸Answer key — planted errors, the hidden pattern, and reference figuresoptional deep dive — click to expand

Planted error 1 — negative units (SB-1031). −390 units of Cola Classic in November. Physically impossible as a sale; probably a return or keying error. Including it understates November revenue by Rs 21,450 and dents the seasonality signal. Correct handling: exclude (or recode as a return, separately) — never silently flip the sign, because you don't know which it was.

Planted error 2 — exact duplicate (SB-1034 / SB-1035). Same date, region, rep, product, units, price; consecutive IDs. Classic double-entry. Keep one, drop one, note it. If both stayed, December revenue is inflated by Rs 12,650.

Planted error 3 — impossible date (SB-1017). 2025-06-500 is not a date. The AI often "helpfully" reads it as June; your log should either correct to a June date with a flag (day unknowable) or exclude from any date-based analysis while keeping it in totals. Either ruling passes — silent correction does not.

Planted error 4 — region typo (SB-1008). Nroth for North. Left unfixed, region totals split North's revenue across two labels and can flip the "top region" answer. Fix and log.

The hidden pattern: October–November is a festive-season spike driven almost entirely by Cola Classic 1L — five of the six largest orders land in a four-week window across all three regions, roughly doubling normal monthly revenue. Mango Fizz shows a milder summer ramp (April–June). The stock-planning insight: build Cola inventory from September; the spike is category-specific, not a general lift.

Reference figures (after cleaning: drop SB-1031 and SB-1035, fix Nroth, rule on SB-1017): total revenue ≈ Rs 3,48,180 with SB-1017 included in totals (Rs 6,720 of Mango Fizz). West is the top region and Ramesh the top rep (West's festive Cola orders decide both). Cola Classic is the top revenue product overall despite Mango Fizz leading unit counts in summer. Your exact totals may differ slightly by your SB-1017 ruling — what is graded is that your report states the ruling.

Grading: all four data errors caught and logged = distinction; the seasonality named at product level = pass on analysis; any reported figure that your own spreadsheet cannot reproduce = fail that question and redo it.

Key Takeaway:

Dirty data in, confident nonsense out — profile first, rule on every anomaly yourself, and reconcile every AI-computed number in a spreadsheet before it reaches a decision-maker. The reconciliation stamp at the bottom of a report is what separates an analyst from an autocomplete.

This is the free reading view. Open BXUP to track your progress, run the hands-on labs, and prove it with original, exam-pattern assessments.
© Anvi Innovate · BXUP — Be Exponentially Upskilled · BXUP is an independent learning platform, not affiliated with or endorsed by AWS, Microsoft, Google, NVIDIA, IBM, Salesforce, Databricks, Oracle or Anthropic. All trademarks belong to their respective owners. · bxup.in