Build an AI Spreadsheet Analysis Workflow

A dependable AI spreadsheet analysis workflow turns an unfamiliar workbook into a traceable analysis with verified calculations, data-quality findings, meaningful comparisons, explainable anomalies, and clearly separated facts, hypotheses, and next questions.

The goal is not to upload a spreadsheet and accept whatever summary sounds convincing. The goal is to preserve the original file, understand what each row and column represents, validate formulas and totals, perform calculations through reproducible methods, and use AI to organize findings and explain patterns without inventing missing context.

Why spreadsheet analysis produces unreliable conclusions

Spreadsheets often look structured while hiding inconsistent definitions, broken formulas, manual overrides, duplicated records, mixed reporting periods, and formatting that changes how values are interpreted.

Common failure patterns include:

Analyzing the first visible sheet without checking the full workbook.
Treating formatted tables as clean source data.
Assuming one row represents one transaction when it represents a subtotal or snapshot.
Combining currencies, time zones, units, or reporting periods.
Reading numbers stored as text as valid numeric values.
Ignoring hidden rows, columns, sheets, and filters.
Using totals that exclude newly added rows.
Accepting formulas with broken references or manual constants.
Counting duplicated records as separate events.
Treating blanks, zeros, and “not applicable” as equivalent.
Calling every extreme value an error.
Explaining anomalies without checking operational context.
Using AI-generated calculations that cannot be reproduced.
Publishing summaries without reconciling them to the source.

A reliable workflow should answer:

What business question is this analysis intended to answer?
What does one row represent?
Which columns are authoritative inputs, formulas, labels, or notes?
Are the records complete, unique, valid, and comparable?
Can every result be reproduced from the source data?
Which unusual values require investigation?
What is known, what is inferred, and what remains unanswered?

Step 1: Define the analysis question

Do not begin by asking AI to “find insights.” Define the decision, comparison, population, period, and expected output first.

Analysis field	Question to define
Business question	What decision or investigation should the analysis support?
Population	Which transactions, customers, products, projects, or periods are included?
Measure	Which values, rates, counts, or outcomes matter?
Comparison	Target, previous period, budget, forecast, baseline, or segment?
Time period	Which exact start and end dates apply?
Decision owner	Who will approve the interpretation and act on it?
Output	Quality report, anomaly list, trend summary, reconciliation, or decision brief?
Risk level	Could the result affect customers, money, compliance, staffing, or public claims?

Turn vague requests into testable questions

Vague request	Testable analysis question
Analyze sales	Which products and regions explain the variance between actual and budgeted revenue for the current quarter?
Find unusual expenses	Which approved expense records exceed category-specific limits or differ materially from the employee’s normal pattern?
Summarize project performance	Which projects missed their committed dates, and which blockers or ownership patterns are associated with those delays?
Check customer data	Which records contain duplicate identifiers, missing required fields, invalid status values, or conflicting account ownership?

Step 2: Protect the original file and sensitive data

Preserve the original workbook before cleaning, editing, sorting, filtering, or correcting it. Analysis should occur on a controlled working copy.

Save the original file as read-only when possible.
Create a dated working copy with a clear version name.
Record the file owner, source, reporting period, and collection time.
Preserve original sheet names, formulas, comments, and metadata.
Do not overwrite source values during cleaning.
Record every transformation applied to the working data.
Limit access according to the information contained in the workbook.
Remove unnecessary personal or confidential fields before using an external AI service.

Check for sensitive content

A workbook may contain information that is not obvious from its title.

Customer names, email addresses, phone numbers, or account identifiers
Employee details, salaries, performance records, or attendance
Payment, invoice, tax, banking, or transaction information
Private notes, comments, and internal assessments
Credentials, access tokens, or hidden connection details
Confidential pricing, forecasts, margins, or contract terms
Health, legal, safety, or other restricted information

Use the minimum information required for the analysis. Aggregate, redact, or replace identifiers when detailed records are unnecessary.

Step 3: Inventory the entire workbook

Before analyzing values, inspect the structure of the file. The visible sheet may depend on hidden calculations, lookup tables, imports, named ranges, or manual assumptions elsewhere in the workbook.

Workbook element	What to record
Worksheets	Name, purpose, row count, column count, and visibility
Tables and ranges	Structured tables, named ranges, and manually selected ranges
Formulas	Formula columns, summary cells, lookups, and external references
Pivot tables	Source range, refresh status, filters, and calculated fields
Charts	Source range, displayed measure, and filters
Hidden content	Hidden sheets, rows, columns, and filtered-out records
External connections	Linked files, queries, databases, or imported data
Comments and notes	Manual explanations, exceptions, and assumptions
Macros or scripts	Automations that modify, calculate, or export data

Create a sheet inventory

Sheet	Purpose	Source or output	Owner	Analysis status
Raw_Orders	Imported order records	Source	Sales operations	Requires quality checks
Product_Map	Product category lookup	Reference	Commercial operations	Verify uniqueness
Monthly_Summary	Management totals	Calculated output	Finance analyst	Reconcile to source
Dashboard	Charts and KPI presentation	Presentation output	Operations lead	Do not use as raw data

Step 4: Define what one row represents

The row grain determines how records may be counted, joined, filtered, and summarized. Never assume the grain from the column names alone.

Possible grain	Example
Transaction	One row per invoice line or payment
Order	One row per customer order
Customer	One row per customer account
Task	One row per work item
Status event	One row per status change
Daily snapshot	One row per item per day
Monthly aggregate	One row per segment and month
Subtotal	One row summarizing several detailed records

Mixing transaction rows with subtotals can double-count values. Combining current-state records with historical snapshots can multiply the number of customers, projects, or balances.

Identify a unique record key

Document the field or combination of fields that should uniquely identify one record.

Transaction ID
Order ID plus line number
Customer ID
Project ID plus snapshot date
Employee ID plus payroll period
Invoice number plus item code

When no reliable key exists, the workflow should report that limitation rather than silently treating every row as unique.

Step 5: Create a data dictionary

A data dictionary records what each column means, where it comes from, how it is formatted, and whether it may be used in calculations.

Column field	Required information
Column name	The exact spreadsheet header
Business meaning	What the value represents
Data type	Text, integer, decimal, date, datetime, Boolean, or category
Unit	Currency, percentage, minutes, items, kilograms, or another unit
Source	System, form, user entry, lookup, or formula
Allowed values	Approved categories, range, or format
Blank meaning	Missing, unknown, not applicable, not yet entered, or zero
Formula	Calculation rule when the column is derived
Owner	Person responsible for definition and quality
Sensitivity	Public, internal, confidential, or restricted

Example data-dictionary record

Column	Net_Revenue
Meaning	Recognized revenue after discounts and approved refunds
Type	Decimal
Currency	Currency defined in the Currency_Code column
Formula	Gross revenue minus discount minus approved refund
Blank meaning	Calculation unavailable because a required source value is missing
Owner	Finance operations

Step 6: Profile data quality before analysis

Measure the quality of the data before calculating business results. Quality findings should remain separate from performance findings.

Quality dimension	Check
Completeness	Required fields are populated
Uniqueness	Record keys are not duplicated unexpectedly
Validity	Values follow the approved format and range
Consistency	Related columns and sheets do not conflict
Accuracy	Values reconcile to authoritative sources when available
Freshness	Records cover the expected reporting period
Referential integrity	Identifiers match valid reference records
Comparability	Units, definitions, and periods are compatible

Profile each column

Total row count
Blank count and blank percentage
Distinct-value count
Duplicate count
Minimum and maximum
Median and relevant percentiles
Most common categories
Invalid values
Unexpected data types
Earliest and latest dates
Formula versus constant count

A column that appears numeric may contain currency symbols, spaces, commas, dashes, error values, or text comments. Do not calculate until the type has been verified.

Step 7: Normalize types, units, dates, and categories

Formatting is not the same as stored value. Standardize the data on the working copy while preserving the original representation.

Numbers and percentages

Distinguish decimal values from formatted percentages.
Confirm whether 15 means 15 units, 15%, or 0.15.
Remove thousands separators only after identifying the locale.
Keep original precision when rounding is not part of the business rule.
Do not convert blanks into zero automatically.

Dates and time

Confirm day-month-year versus month-day-year formats.
Separate dates from datetimes.
Document the reporting time zone.
Identify dates stored as text.
Check for impossible or default dates.
Confirm which date controls inclusion in the analysis.

Currencies and units

Keep currency code separate from numeric amount.
Do not add different currencies without an approved conversion rule.
Record the exchange-rate source and effective date when conversion is required.
Standardize units such as minutes and hours before aggregation.
Separate quantities, prices, rates, and percentages.

Original values	Normalized value
Complete, Completed, Done	Completed
In progress, In-Progress, WIP	In Progress
UK, United Kingdom, GB	United Kingdom
N/A, Not applicable	Not Applicable

Step 8: Audit formulas and calculated cells

Calculated values should be inspected before they are used as evidence. A workbook may contain correct-looking outputs generated by incomplete ranges or inconsistent formulas.

Identify cells containing formulas and cells containing hard-coded values.
Check whether formulas are consistent down each calculated column.
Find formulas that reference blank, deleted, or incorrect cells.
Inspect absolute and relative references.
Check whether new rows fall outside fixed calculation ranges.
Identify hidden error handling that converts errors into zero or blank.
Review lookup formulas for duplicate or missing keys.
Confirm whether subtotals include filtered or hidden rows.
Check external workbook references and their update status.
Confirm whether calculated values have been replaced manually.

Common formula risks

Risk	Example
Incomplete range	A total stops at row 500 while new records extend to row 620
Inconsistent formula	One row uses gross value while the rest use net value
Broken lookup	A missing key returns an empty string and hides the error
Manual override	A calculated result is replaced with a typed value
Circular reference	A summary depends directly or indirectly on itself
Stale external link	The workbook displays values from an older linked file

Do not assume that a displayed value is reliable because the spreadsheet application shows no visible error.

Step 9: Reconcile the workbook to trusted totals

Before interpreting trends, confirm that the working dataset agrees with an approved source or control total.

Reconciliation	Example
Record count	Workbook transaction count matches the source-system export
Financial total	Invoice amount agrees with the approved ledger report
Customer count	Distinct active customers match the CRM control total
Period coverage	Earliest and latest records match the requested dates
Category totals	Segment totals add to the approved overall value
Opening and closing balance	Opening balance plus movement equals closing balance

Record the difference, permitted tolerance, explanation, and owner when reconciliation does not match.

Step 10: Separate source data, transformations, and outputs

A reproducible workbook should make the analysis path visible.

Source layer: Original imported or entered records.
Reference layer: Approved mappings, categories, and targets.
Transformation layer: Cleaning, standardization, joins, and derived fields.
Analysis layer: Metrics, comparisons, distributions, and anomaly rules.
Presentation layer: Tables, charts, and written summaries.

Do not mix manual comments, source values, and calculated results inside the same column. Keep transformation steps documented and repeatable.

Step 11: Calculate approved metrics through reproducible methods

Use spreadsheet formulas, queries, scripts, or another deterministic calculation method for exact results. AI may help draft formulas or explain outputs, but every number should be reproducible independently.

Metric requirement	Required definition
Metric name	The approved business label
Formula	The exact calculation
Numerator	The measured quantity
Denominator	The population used for a rate
Inclusions	Records included in the result
Exclusions	Cancelled, test, duplicate, invalid, or incomplete records
Time rule	The date determining the reporting period
Unit	Count, currency, percentage, hours, or another unit
Comparison	Target, budget, baseline, forecast, or previous period
Owner	The person approving the metric definition

Use distributions, not only averages

An average may hide the experience of most records or a small set of serious exceptions. Depending on the question, calculate:

Count
Sum
Mean
Median
Minimum and maximum
Percentiles
Standard deviation
Interquartile range
Category distribution
Period-over-period change
Rate with a clear denominator

Step 12: Compare results fairly

Every comparison should use compatible definitions, populations, and time periods.

Comparison	Useful when	Risk
Previous period	The operation is stable and seasonality is limited	Periods may have different numbers of working days
Same period last year	Seasonality matters	Definitions or processes may have changed
Budget	Financial control is required	The budget may be outdated or built on different assumptions
Target	An approved operating standard exists	The target may not fit every segment
Forecast	Expected performance needs review	The forecast may contain uncertain assumptions
Segment	Different customer, product, or team experiences matter	Small groups may create unstable results

Label partial periods clearly. Do not compare a full month with an incomplete month without an approved adjustment.

Step 13: Define anomaly rules before flagging records

An anomaly is a value or pattern requiring investigation. It is not automatically an error, fraud, or operational failure.

Use several anomaly types:

Anomaly type	Example
Rule violation	A negative quantity where only positive quantities are permitted
Threshold exception	An expense exceeds the approved category limit
Statistical outlier	A value is far outside the normal distribution
Trend break	A previously stable series changes sharply
Missing expected record	A scheduled monthly submission is absent
Duplicate pattern	The same identifier, amount, and date appear more than once
Sequence problem	A completion date occurs before the start date
Cross-field conflict	Status is completed while completion date is blank
Segment deviation	One region differs materially from comparable regions

Document every anomaly rule

Rule field	Required information
Rule name	Clear description of the exception
Population	Records eligible for the check
Condition	The exact formula or threshold
Minimum volume	Sample required before the rule applies
Severity	Informational, review, urgent, or critical
Evidence	Fields supporting the flag
Owner	Person responsible for investigation
Expected action	Verify, correct, explain, approve, or escalate

A statistical threshold should not override an approved business rule. A rare value may be valid, while a common value may still violate policy.

Step 14: Investigate anomalies with context

AI can group unusual records and suggest questions, but a human should verify the source and operational context before a conclusion is recorded.

Investigation field	Question
Flagged record	Which row, identifier, or group triggered the rule?
Rule triggered	Why was it selected?
Source evidence	What original values support the flag?
Comparison	How does it differ from peers, history, or policy?
Possible explanation	Which hypotheses are consistent with the evidence?
Verification required	Which system, document, or owner can confirm the explanation?
Impact	What decision, total, customer, or process is affected?
Disposition	Valid exception, data error, formula error, duplicate, unresolved, or approved adjustment?

Separate facts from hypotheses

Statement type	Example
Fact	Record 418 contains an amount 4.7 times above the category median
Hypothesis	The amount may represent an annual rather than monthly charge
Verification	Compare the record with the source invoice and contract
Conclusion	The record is valid and should be classified as an annual subscription

Do not convert a plausible hypothesis into a confirmed explanation without supporting evidence.

Step 15: Use visualizations to answer defined questions

Charts should support comparison, distribution, trend, or investigation. They should not replace the underlying calculations or quality notes.

Analytical need	Useful visual
Trend over time	Line chart
Compare categories	Sorted bar chart
Show distribution	Histogram or box plot
Compare actual with target	KPI table or variance chart
Investigate records	Detailed table with identifiers and filters
Show relationship	Scatter plot when both variables are suitable
Show composition	Stacked bar chart with a limited number of categories

Include the reporting period, units, filters, source, and data-quality limitation with every decision-critical visual.

Step 16: Build a traceable insight brief

The final summary should distinguish verified results, quality limitations, anomalies, hypotheses, and next questions.

Analysis question: The decision or investigation supported.
Dataset: File, sheets, period, population, and row grain.
Quality status: Missing, duplicate, invalid, stale, or conflicting records.
Reconciliation: Whether totals agree with approved controls.
Verified findings: Reproducible calculations and comparisons.
Anomalies: Flagged records, rules, severity, and status.
Possible explanations: Hypotheses requiring verification.
Limitations: Missing data, weak definitions, and sample restrictions.
Next questions: Additional evidence or analysis required.
Actions: Owners, deadlines, and expected outcomes.

Example insight record

Finding	Median delivery time increased from 4.2 to 6.1 days
Population	Completed standard orders in the current and previous full months
Evidence	1,842 current-period records and 1,779 previous-period records
Primary concentration	Orders routed through Warehouse C
Anomaly	Warehouse C contains an unusually high number of records waiting more than three days before dispatch
Hypothesis	A staffing or inventory constraint may be increasing pre-dispatch wait time
Verification	Compare staffing, inventory availability, and carrier collection records
Owner	Fulfilment operations manager

Step 17: Run human quality control

An analyst or business owner should review the workbook, calculations, and summary before the results are used.

The correct workbook version and reporting period were used.
Every included sheet and range has a documented purpose.
The row grain and unique key are understood.
Column definitions and units are confirmed.
Blanks, zeros, and not-applicable values are treated correctly.
Duplicate records have been reviewed.
Formula ranges and references are complete.
Manual overrides are identified.
Totals reconcile to approved sources.
Calculations can be reproduced independently.
Comparison periods and populations are compatible.
Anomaly rules are documented.
Extreme values are not automatically treated as errors.
Facts and hypotheses are separated.
Claims do not exceed the available evidence.
Sensitive information is protected.
Every recommended action has an owner.

Step 18: Preserve the analysis record

Store enough information for another analyst to reproduce and review the work.

Original file and checksum or immutable copy
Working file version
Analysis question and scope
Data dictionary
Transformation log
Formula and quality findings
Reconciliation results
Metric definitions
Anomaly rules
Approved output tables and charts
Human reviewer and approval date
Corrections and revised versions

Do not silently replace a previously approved workbook or result. Record what changed, why it changed, and which conclusions were affected.

Measure whether the workflow improves analysis

Success is not the number of rows processed or summaries generated. Measure accuracy, reproducibility, review effort, and decision usefulness.

Workbook acceptance rate: Percentage of files suitable for analysis without major repair.
Data-quality issue rate: Number of material issues per workbook or dataset.
Formula-error rate: Percentage of reviewed formulas requiring correction.
Reconciliation accuracy: Difference between calculated and approved control totals.
Duplicate-detection accuracy: Percentage of duplicate flags confirmed by reviewers.
Anomaly precision: Percentage of flagged records that require meaningful investigation.
False-positive rate: Valid records incorrectly flagged as problematic.
Insight correction rate: Percentage of AI-drafted findings materially changed.
Reproducibility rate: Percentage of results reproduced by another reviewer.
Analysis time: Time required from file intake to approved insight brief.
Action adoption: Percentage of verified findings entering an operational decision.
Decision usefulness: Whether the analysis improves a real business action.

Copy-and-use prompts

Workbook inventory prompt

You are helping me prepare a spreadsheet for controlled analysis.

Business question:
[QUESTION]

Workbook context:
[OWNER, SOURCE, PERIOD, AND PURPOSE]

Workbook inventory:
[PASTE SHEET NAMES, DIMENSIONS, TABLES, FORMULAS, CONNECTIONS, AND NOTES]

Create:

1. Workbook purpose summary
2. Sheet inventory
3. Likely source sheets
4. Reference and lookup sheets
5. Transformation sheets
6. Calculated output sheets
7. Presentation-only sheets
8. Hidden or external dependencies requiring review
9. Possible row grain for each source sheet
10. Possible unique keys
11. Missing documentation
12. Sheets that should not be treated as raw data
13. Human questions required before analysis

Rules:
- Do not infer business meaning from sheet names alone
- Do not assume visible sheets are complete
- Do not treat charts or summaries as source data
- Mark external links and hidden content clearly
- Keep uncertain conclusions labelled as hypotheses

Data-dictionary prompt

Create a draft data dictionary for this spreadsheet.

Analysis question:
[QUESTION]

Row grain:
[GRAIN]

Column names:
[COLUMNS]

Sample values:
[SAMPLES]

Known business definitions:
[DEFINITIONS]

For each column, return:

1. Exact column name
2. Proposed business meaning
3. Proposed data type
4. Unit or currency
5. Source type:
   - imported
   - manually entered
   - lookup
   - formula
   - unknown
6. Allowed values or expected range
7. Meaning of blank
8. Possible quality risks
9. Sensitivity level
10. Definition confidence:
   - confirmed
   - likely
   - unclear
11. Human clarification required

Rules:
- Do not invent business definitions
- Do not treat formatting as proof of data type
- Separate zero, blank, unknown, and not applicable
- Mark mixed types and ambiguous dates
- Do not assume currency from a symbol alone

Data-quality review prompt

Review this spreadsheet data-quality profile.

Analysis scope:
[SCOPE]

Row grain and key:
[GRAIN AND KEY]

Data dictionary:
[DICTIONARY]

Column profiles:
[PASTE COUNTS, BLANKS, DISTINCT VALUES, MINIMUMS, MAXIMUMS, AND TYPES]

Known validation rules:
[RULES]

Return:

1. Completeness issues
2. Duplicate-key issues
3. Invalid values
4. Type inconsistencies
5. Date and time issues
6. Currency or unit conflicts
7. Category inconsistencies
8. Referential-integrity problems
9. Stale or incomplete period risks
10. Cross-field conflicts
11. Severity:
   - informational
   - review
   - blocking
12. Affected rows or fields
13. Recommended correction or investigation
14. Human owner

Rules:
- Do not replace blanks with zero automatically
- Do not remove duplicates without verification
- Do not correct source values silently
- Separate definite errors from possible exceptions
- Preserve the original value and record identifier

Formula-audit prompt

Review this spreadsheet formula inventory.

Sheet and range:
[SHEET OR RANGE]

Expected calculation:
[BUSINESS RULE]

Formula patterns:
[PASTE FORMULAS OR FORMULA GROUPS]

Known row range:
[RANGE]

Return:

1. Formula purpose
2. Expected consistent pattern
3. Inconsistent formulas
4. Hard-coded values inside calculated ranges
5. Broken or missing references
6. Fixed ranges that may exclude new records
7. Lookup risks
8. Hidden error handling
9. External-link dependencies
10. Possible circular references
11. Cells requiring human verification
12. Recommended test calculation

Rules:
- Do not approve a formula because it returns a value
- Compare the formula with the documented business rule
- Mark uncertain findings for manual review
- Preserve exact cell references
- Do not change the workbook automatically

Analysis-planning prompt

Create a reproducible analysis plan for this spreadsheet.

Business question:
[QUESTION]

Decision owner:
[OWNER]

Dataset scope:
[SCOPE]

Row grain:
[GRAIN]

Approved columns:
[COLUMNS]

Metric definitions:
[METRICS]

Comparison:
[COMPARISON]

Known quality limitations:
[LIMITATIONS]

Create:

1. Required filters
2. Required exclusions
3. Calculations
4. Groupings and segments
5. Comparison periods
6. Distribution statistics
7. Reconciliation checks
8. Anomaly rules
9. Required output tables
10. Recommended visuals
11. Verification steps
12. Limitations to disclose
13. Questions that cannot be answered from the spreadsheet
14. Human approval points

Rules:
- Use deterministic calculations for all numbers
- Do not invent missing values
- Do not use averages when the distribution matters
- Keep quality findings separate from performance findings
- Make every result traceable to source records

Anomaly-investigation prompt

Prepare these spreadsheet anomalies for human investigation.

Analysis scope:
[SCOPE]

Approved anomaly rules:
[RULES]

Flagged records:
[RECORDS]

Comparison statistics:
[STATISTICS]

For each anomaly, return:

1. Record identifier
2. Rule triggered
3. Source values
4. Comparison value
5. Size of deviation
6. Severity
7. Possible explanations
8. Evidence supporting each explanation
9. Evidence still required
10. Potential impact
11. Recommended owner
12. Recommended disposition:
   - valid exception
   - data error
   - formula error
   - duplicate
   - approved adjustment
   - unresolved
13. Human decision required

Rules:
- Do not label an anomaly as an error without verification
- Do not allege misconduct or intent
- Separate facts from hypotheses
- Preserve the original row and source reference
- Do not hide conflicting evidence

Insight-summary prompt

Create a draft spreadsheet analysis brief for human review.

Analysis question:
[QUESTION]

Dataset summary:
[DATASET]

Quality findings:
[QUALITY]

Reconciliation results:
[RECONCILIATION]

Verified calculations:
[CALCULATIONS]

Comparisons:
[COMPARISONS]

Anomalies:
[ANOMALIES]

Known limitations:
[LIMITATIONS]

Create:

1. Executive summary
2. Dataset and scope
3. Data-quality status
4. Reconciliation status
5. Verified findings
6. Important comparisons
7. Confirmed anomalies
8. Unresolved anomalies
9. Possible explanations clearly labelled as hypotheses
10. Limitations
11. Next questions
12. Recommended actions with owners
13. Verification checklist

Rules:
- Use only supplied verified calculations
- Do not calculate new totals in prose
- Do not invent causes or missing context
- Separate facts, hypotheses, and recommendations
- Include denominators and sample sizes when relevant
- Disclose unresolved quality problems
- Do not describe correlation as proof of cause

Spreadsheet analysis quality-control prompt

Review this spreadsheet analysis before it is used.

Business question:
[QUESTION]

Source workbook documentation:
[DOCUMENTATION]

Metric definitions:
[METRICS]

Quality and reconciliation results:
[QUALITY]

Draft analysis:
[DRAFT]

Check for:

1. Wrong workbook version or period
2. Unclear row grain
3. Duplicate or subtotal records counted incorrectly
4. Blanks treated as zero
5. Mixed currencies, units, or time zones
6. Invalid date interpretation
7. Inconsistent or incomplete formulas
8. Totals that do not reconcile
9. Metrics without approved definitions
10. Mismatched comparison periods
11. Averages hiding important variation
12. Anomalies described as confirmed errors
13. Hypotheses presented as facts
14. Unsupported causal claims
15. Missing sample sizes or denominators
16. Sensitive information exposed unnecessarily
17. Results that cannot be reproduced
18. Actions without owners

Return:
- Blocking corrections
- Important corrections
- Calculations requiring reproduction
- Anomalies requiring investigation
- Claims requiring weaker language
- Missing limitations
- Final decision:
  - ready for internal use
  - minor revision
  - major revision
  - additional verification required
  - do not use

Do not approve the analysis merely because the summary sounds clear.

AI spreadsheet analysis workflow checklist

The analysis question and decision owner are defined.
The original workbook is preserved unchanged.
A controlled working copy is used.
Sensitive fields and access rules are reviewed.
Every worksheet and hidden dependency is inventoried.
Source, reference, transformation, analysis, and presentation sheets are separated.
The row grain is documented.
A unique record key is confirmed or its absence disclosed.
A data dictionary exists for important columns.
Blanks, zeros, unknowns, and not-applicable values are separated.
Data types, dates, currencies, units, and categories are validated.
Duplicate records are reviewed before removal.
Formula patterns, ranges, lookups, and external links are checked.
Manual overrides inside calculated areas are identified.
Totals reconcile to an approved source.
Every metric has a reproducible definition.
Calculations use deterministic methods.
Comparison periods and populations are compatible.
Distributions are reviewed when averages may mislead.
Anomaly rules are documented before records are flagged.
Extreme values are investigated rather than automatically removed.
Facts, hypotheses, and recommendations are separated.
Every finding is traceable to source records.
Limitations and missing data are disclosed.
A human reviews the final calculations and interpretation.
The analysis files, rules, and approvals are preserved.

Common mistakes to avoid

Starting with the dashboard sheet: Identify the true source records first.
Trusting formatting: Verify stored types and values.
Ignoring row grain: Confirm what one row represents before counting.
Deleting duplicates automatically: Verify whether they are errors or valid repeated events.
Treating blanks as zero: Preserve the business meaning of missing values.
Trusting visible formulas: Check ranges, references, lookups, and overrides.
Using AI for exact arithmetic: Calculate through reproducible spreadsheet or code logic.
Calling outliers errors: Investigate the operational context.
Explaining correlation as cause: Label possible explanations as hypotheses.
Publishing without reconciliation: Confirm totals and definitions first.

Final guidance

A dependable AI spreadsheet analysis workflow does not begin with a request for instant insights. It begins by protecting the source, understanding the workbook, documenting the row grain, validating the data, auditing formulas, and reproducing every calculation.

Use AI to organize workbook documentation, identify quality risks, prepare analysis plans, group anomalies, and draft clear explanations. Keep exact calculations, data corrections, anomaly decisions, sensitive-data handling, causal interpretation, and final approval under human control.