You will be able to have an assistant clean, summarise and chart a dataset and confirm the results are right.
On Friday, Priya's regional manager sends a spreadsheet export from the delivery system and asks which routes have the most late deliveries and whether it is getting worse. The file has about 4,000 rows (an invented example, as are all of Priya's figures), dates in two formats, blank cells where drivers had not closed jobs, and a few hundred duplicates from a system glitch in July, so cleaning it by hand would take most of a morning.
Many assistants can now handle a file like this by writing and running code in the background. You upload the spreadsheet and describe what you want, and the assistant cleans the data, calculates totals, builds tables and draws charts. You see the results and can usually ask to see the code it ran.
With formula help in lessons 5.1 and 5.2, you got a formula and ran it in your own sheet. A code-running assistant does the work on a copy of your data on the provider's servers, which makes one-off analysis faster. Because the data goes to the provider's servers, the data-handling rules from module 2 apply. Upload only data your company's rules allow in that tool, and redact or remove anything the task does not need.
Priya's question is about routes and dates, so driver names and customer addresses come out before the upload. The task description follows the pattern lesson 5.1 used for formulas: what the columns mean, what you want to know, and any rules you already know about the data. Priya gives the assistant three rules. A job is late if Delivered at is after Promised by. Rows where Status is Cancelled are ignored. Each Job ID counts as one delivery even if it appears twice.
Jobs like this are where an assistant saves you the most time, and also where a confident wrong number is most likely to slip through. Results come back quickly and look finished, and the cleaning behind them stays hidden unless you ask about it. Each time, ask the assistant to report:
the number of rows at the start and at the end which rows it removed, and the reason for each how it handled blank cells and duplicates any assumptions it made about dates or about a column it was unsure of
Most problems show up in this report. An assistant might drop every row with a blank in any column, including rows that are blank only in an irrelevant column such as notes. It might read 03/07 as 7 March when your system meant 3 July. It might count duplicates as separate deliveries if you did not mention the glitch that caused them. You would not see any of these in a chart, yet each one changes the answer. A wrong filter gives you no error message, and the wrong number arrives as neatly formatted as a right one would.
Priya's report says the assistant started with 4,012 rows, removed 318 duplicate Job IDs and 41 cancelled jobs, and kept 3,653 (4,012 - 318 - 41 = 3,653). It also shows the assistant had counted blank Delivered at cells as late. Those were jobs drivers had not closed, so Priya tells the assistant to exclude them and run the analysis again. The assistant confirms it read dates as day then month, which matches Priya's system.
Before you use any result, check at least two figures yourself. A pivot table in your own spreadsheet is the quickest way. Build it on the same cleaned data, or on the original file with the same filters applied by hand, and compare a total or count with the assistant's figure. Choose one figure that is easy to check, such as late deliveries on one route in one month, and one that matters most, such as the route the assistant names as the worst. When both match, you have more reason to trust the rest. When either one differs, find the cause and fix the rule before you send anything.
Priya rebuilds late deliveries on the busiest route in August, and the pivot table figure matches. Total late jobs is off by 27. The gap comes from jobs delivered on the promised day but a few minutes after the promised time. The assistant counted these as late, and the regional manager's definition of late leaves them out. Priya clarifies the rule with the assistant, and the figures line up.
Some spreadsheet apps include a built-in assistant, such as Copilot in Excel or Gemini in Google Sheets, depending on your company's licences and settings. It works inside your file instead of on an uploaded copy, which can be convenient. The same rules apply. Check that your company allows it on work data before you use it, since an assistant that sits inside a familiar app has not necessarily been approved. Treat its results as you would any other assistant's: ask what it did and verify a figure or two.
For your own practice, start with data that cannot cause trouble. Public tables from SingStat, the national statistics agency, suit this well because they are real, structured and free of anyone's personal details. Choose one now that connects in some way to your work.
Upload a non-confidential dataset, such as a public SingStat table, ask for three summary figures and one chart, and verify two figures yourself.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).