Analysing a spreadsheet with AI.

Goal: Answer one real business question from a spreadsheet export using AI, and prove the answer is right before you act on it. Who it is for: Owners, office managers and operations staff who export data from Xero, ServiceM8, a point-of-sale system or a CRM and want answers without building a report. Time: 30 minutes, plus a CSV or Excel export with at least a few hundred rows (or the practice data below).

An assistant will happily tell you that leak repairs are your least profitable job type. It will not tell you, unless asked, that it read every date as month first, counted a subtotal row twice, or drew the conclusion from ten jobs. Spreadsheet analysis with AI is fast and genuinely useful, and the speed is exactly why the checks matter.

How to get reliable answers from a spreadsheet

Reliable AI spreadsheet analysis follows four moves: profile, ask, show, verify. Skipping the first or last is how wrong numbers reach a meeting.

Most current assistants can run code on an uploaded file rather than estimating. When yours does, ask to see the code; when it cannot, ask for the spreadsheet formula and run it yourself.

Exercise: answer one question from your own export

1. Write the question first. Make it specific, such as “Which job types ran over quote most in the last quarter, by dollars and by percentage?” A vague question gets a vague tour of the data.
2. Export and clean a copy. Export from your system to CSV. Delete customer names, phone numbers, emails and street addresses; keep a suburb or region if location matters. Keep the original untouched. The NCSC’s AI guide for small businesses gives the same advice: remove or anonymise personal details before using an AI tool.
3. Profile before you analyse. Upload the file to ChatGPT, Claude, Copilot or Gemini and use the profiling part of the prompt below. Compare its row count and column totals with your source system.
4. Ask your question and demand the working. Ask for the method, the code or formula, and a small results table.
5. Verify two numbers. Rebuild two figures yourself with a pivot table or SUMIFS. If either differs, find out why before going further.
6. Ask what would change the conclusion. Small samples, one unusual job and missing data are the usual answers. Note them beside your result.

Worked example: where do quotes run over?

Illustrative example: a five-van Wellington plumbing business exports completed jobs for July to September 2026. Here are ten rows you can paste into any assistant to practise. The business and data are fictional.

job_date,suburb,job_type,quoted_ex_gst,invoiced_ex_gst,labour_hours,callback
03/07/2026,Karori,Hot water cylinder,2400,2650,6.5,N
05/07/2026,Newtown,Blocked drain,350,350,1.5,N
11/07/2026,Johnsonville,Hot water cylinder,2400,2980,8,Y
14/07/2026,Miramar,Leak repair,280,410,2.5,N
02/08/2026,Karori,Blocked drain,350,520,3,Y
09/08/2026,Island Bay,Leak repair,280,280,1,N
12/08/2026,Newtown,Hot water cylinder,2600,2600,6,N
20/08/2026,Kilbirnie,Bathroom renovation,14500,16900,62,N
04/09/2026,Johnsonville,Leak repair,280,355,2,N
10/09/2026,Miramar,Blocked drain,350,350,1.5,N

Step 1. Profile this data before analysing it. Report: row count, each
column and its data type, the date range, blank cells and anything odd.
Dates are New Zealand format, DD/MM/YYYY. Amounts exclude GST.

Step 2. Then answer: which job types ran over quote, by total dollars
and by percentage of the quoted amount? Show the code or formula you
used, give a results table, and tell me what would change the answer.

Example output (illustrative, abridged):

Profile: 10 rows, 7 columns. Dates 3 July to 10 September 2026 (read as
day first). No blanks. One job (Bathroom renovation) is far larger than
the rest.

Over quote by job type (excl GST):
Job type             Jobs  Quoted   Invoiced  Over    Over %
Hot water cylinder   3     $7,400   $8,230    $830    11.2%
Leak repair          3     $840     $1,045    $205    24.4%
Blocked drain        3     $1,050   $1,220    $170    16.2%
Bathroom renovation  1     $14,500  $16,900   $2,400  16.6%

Total over quote: $3,605 on $23,790 quoted (15.2%). The single bathroom
job is $2,400 of that. Excluding it: $1,205 on $9,290 (13.0%).
Both callbacks were on jobs that ran over quote.
Caution: 10 jobs is too few to draw conclusions. Rerun on the full year.

That output is checkable. Leak repairs have the highest overrun percentage but the smallest dollar impact, and one renovation dominates the total. Both readings are true, and a decision depends on which one you care about. If you need GST-inclusive figures, multiply by 1.15; to find the GST inside an inclusive amount, multiply by 3 and divide by 23.

Where spreadsheet analysis goes wrong

• Dates read month first. 03/07/2026 becomes 7 March. State the format in the prompt and check the reported date range.
• Total rows counted twice. Accounting and job system exports often include subtotal and grand total rows. Delete them, or tell the assistant which rows to ignore.
• Partial reads. Large files can be sampled or cut off. Matching row counts and column totals is the cheapest check there is.
• Mental arithmetic. An answer with no code or formula behind it is an estimate. Ask for the working every time.
• Small samples and outliers. Assistants rarely volunteer that ten rows prove nothing. Ask what would change the conclusion.
• Causation from correlation. “Callbacks happen on jobs that run over” does not tell you which causes which. Treat it as a question for the team.

The same discipline applies to text: see checking AI answers against sources. For connecting assistants to live data instead of exports, see our guides to AI and Xero and AI and ServiceM8. When a team needs the same analysis every month, a built report usually beats repeated prompting; our AI reporting and business intelligence work covers that. The next step in turning numbers into commentary is using AI to prepare management reports.

Self-check

• You wrote a specific question before uploading anything.
• Personal details were removed from the copy you uploaded.
• The assistant’s row count and column totals match your source system.
• The date range it reported matches the period you exported.
• Every figure you will use has visible code or a formula behind it.
• You recalculated two figures yourself and they matched.
• You noted at least one reason the conclusion could be wrong.

Next lesson: Summarising documents with citations.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.