LaurenceYang · Excel Intelligence & Controls · excel-ai-analyst
Preserved methodology deck · For business teams and managers · ← Back to home

Read spreadsheets like code.
Turn hard-won judgment
into rules the next analyst can reuse.

Faster management reporting · From one-off analysis to reusable knowledge assets

2026-07 · Preserved first working method; current local Agent authority is stable 1.34.0 with five Skills
The live Browser Analyzer now adds formula families, dependency evidence, Excel errors, and a bounded review queue
This deck contains no real client names, personal names, or financial figures
Evidence labels: MEASURED = observed in a run · TARGET = designed metric, not yet validated
In one sentence

A mature business workbook
is undocumented legacy code.

The product's first principle
Until the AI passes full-data regression tests,
none of its conclusions about the workbook are trustworthy.
Column names are variables, formulas are functions, cross-sheet references are module dependencies, and check columns are unit tests. Treat the workbook as software: understand it first, validate it next, and analyze it only after validation passes.

This deck answers four questions: what makes the work painful, how the method addresses it, where efficiency improves, and which gains were measured in recorded runs.

Chapter One · The Problem

The knowledge lives in people's heads.
When they leave, it disappears.

The next three statements came from a requirements meeting on 2026-07-29.
They show why monthly work remains difficult after the original owners leave.

Chapter One · The Problem

Three direct statements

“This work has been handed down countless times... At this point, not one person who has done it is still with the company.”
— Requirements meeting · Speaker 3
This is not simply poor documentation. Everyone who knew how to do the work has left, yet the work still has to be completed every month.
“Reconciliation is difficult and painful.”
— Requirements meeting · Speaker 3
When management and financial reports disagree, investigation depends on “I have seen this number before,” knowledge that cannot be transferred or independently reviewed.
“The links carry the formulas and data sources. If you break the chain, I cannot tell where the number came from.”
— Requirements meeting · Speaker 1
Workbooks from dozens or hundreds of business units are connected by external links. If someone saves a values-only copy, every downstream dependency breaks without raising an error.
Chapter One · The Problem

The pressure concentrates into three days

This is the monthly management-reporting cadence. The real window is the three days between receiving workbooks on the 6th and the meeting on the 9th, when reports from dozens or hundreds of business units must become one signable management report.

  3rd  Business units submit forecasts
  5th  Accounting closes the books
    6th  Business units submit actuals  ────┐
    7th  Company-level figures produced    │  These are the three days
    8th  Management-review materials     ────┘
  9th  Business performance review
During these three days, mistakes are most costly, staffing is thinnest, and the person who best understood the definitions has already left.
Source: operating calendar in skill/excel-ai-analyst/scripts/ws_init.sh · Dates follow the business team's working definition; the meeting did not distinguish calendar days from business days when referring to the 6th and 9th
Chapter One · The Problem

Four failure modes, none of which raise an error

① Broken links

The formula chain stops midway

Someone saves a values-only copy and the external links disappear. The workbook still opens and the numbers remain, but their provenance is gone.

② Manual overrides

A formula is replaced by hand

Eleven cells in a column contain formulas, while H7 is a constant. A manual review will almost certainly miss it.

③ Definition drift

One term acquires two meanings

Management basis / disclosure basis / YTD / BKD... If definitions are not recorded, everyone has to remember them independently.

④ Exceptions become policy

A one-off judgment becomes routine

The next person treats “this special handling” as a rule, turning an exception into policy.

What they share: none makes the program fail. The error reaches the business-review meeting before a person spots it, if anyone spots it at all.

Chapter Two · The Insight

Why handing the workbook directly to AI fails

Reason one

It reads values, not the algorithm

Converting Excel to text removes the formulas. The AI sees the result, not how the result was produced.

Reason two

It can turn uncertainty into false certainty

When it does not understand something, it may guess and make the guess sound convincing. A wrong answer with authority costs more than no answer.

Reason three

It does not retain last month's conclusions

Each run starts from zero. Business experts must explain this month what they already explained last month.

Someone opened the meeting with a reminder: “State up front that AI is not omnipotent.” That became a product discipline: every report separates Validated / Candidate / Undetermined instead of blurring them together.
Chapter Two · The Insight

Treat workbook elements as software components

In ExcelIt really isSo the method should
Column namesVariable namesBuild a field ontology first and define what every variable means
FormulasFunctionsReconstruct the formula chain and map the calls
Cross-sheet referencesModule dependenciesMap dependencies first and treat broken links like compilation errors
Check columns / total rowsUnit testsRun the test cases left by the original author
“I understand this workbook”A hypothesis to validateRun regression over the full dataset and accept it only if it passes

Software engineering already has methods for undocumented legacy code. This workflow applies those methods to workbook structure, formulas, dependencies, and checks.

Chapter Three · The Method

The six-step method: what AI does and what people do

STEP 0
Profile
Machine hypotheses for table type, structure, header depth, and primary key
SCRIPT
STEP 1
Convert to structured MD
Preserve multi-row headers, merged cells, and original formulas
SCRIPT
STEP 2·3
Field ontology + formula chain
Determine what each column and coefficient actually means
AI-AUTHORED
STEP 4 ★
Full validation
Row-level recalculation, 19 table checks, and cell-level provenance
SCRIPT
STEP 5
Deliver
Annotated results, business analysis, and What-If scenarios
SCRIPT
STEP 6
Dashboard
A single-file web page colleagues can open directly
SCRIPT
Steps 2 and 3 are intentionally not automated. A script can extract =D4+E4-G4, but it cannot infer “Column G is the employee social-insurance contribution, capped by local minimum and maximum bases, and not deducted in the month of departure.” Those definitions require AI interpretation and human review because business meaning is not fully stored in the workbook.
Chapter Three · The Method

If Step 4 fails, Step 5 is not allowed

This is the one non-negotiable rule in the method. The product enforces it by refusing to conclude when it cannot understand, instead of producing a clean-looking answer.

EXIT CODE 0

Everything passed

Every row reconciles and no table-level exception remains. Only then is the workbook understood.

EXIT CODE 1

The run completed with issues

The data or its definition has a problem; the report indicates where to investigate.

EXIT CODE 2

Run refused

The configuration is invalid, so validation did not occur. It is never downgraded to “mostly passed.”

A false green light is worse than an error message: a wrong column name → zero rows tested → a clean ✅. When an alias cannot be aligned, the tool therefore refuses to run and asks a person instead of guessing.
Chapter Three · The Method

Two gates: mathematically correct ≠ correct business meaning

GATE ONE · MACHINE

Does it calculate correctly?

The rule passes over the full dataset and leaves evidence.
A sample is insufficient because management-reporting problems usually hide in a few exceptional rows.

×

GATE TWO · HUMAN

Does it mean the right thing?

A business expert confirms each item and an accountable business owner is named.
A digital worker cannot sign off. AI cannot approve or publish its own output.

The two gates may proceed in parallel; either can pass first. If the data is not fixed by the review meeting, experts can still confirm the business meaning. Once the data is corrected and the rule passes the following month, the system automatically promotes it to publishable status without asking the expert twice. This is what makes repeated review meetings sustainable.
Chapter Three · The Method

Reuse confirmed knowledge next month

Here, “ontology” means a versioned inventory of business objects and rules. It records objects such as organizations, report items, and periods, along with their relationships and the rules that must hold. The “ontology registry” below is where that inventory is stored and versioned.
   ① EXTRACT           ② REVIEW             ③ RETAIN           ④ REUSE
AI reads the workbook  →  Experts confirm each item  →  Save to the ontology  →  Next month, analysts
and writes executable     “Does it mean this?”        registry and publish     write zero config lines
rules                     │                                                        │
                          └──────── Next extraction produces fewer questions ─────┘

Track whether the second review asks fewer questions

The second review should contain fewer items to confirm than the first. If the count does not fall, the knowledge grain or domain boundary is wrong; redesign it instead of scheduling more meetings.

Do not casually promote an exception into a rule

Only propose a general rule after something happens at least three times and is handled consistently each time. Otherwise, keep it in the exception register. Turning an accident into policy is more dangerous than manual work.

Chapter Four · Efficiency

Where efficiency improves

First define how the work is done today. Without a baseline, speed figures are meaningless.

ActivityCurrent stateTarget
End-to-end monthly consolidation Three people for two days, plus several investigation rounds One batch run in < 30 minutes
Excludes human confirmation
Variance localization Review each business unit manually At most three attribution candidates per variance
Top candidate hit rate ≥60%, tested after three months
New analyst onboarding Two-hour handoff recording plus months of trial and error Read the knowledge base and run the full workflow once
Rerun independently within one day
The entire right-hand column contains targets, not achievements. The left column is a PRD baseline recorded from the meeting description; the first task in phase one is to measure it directly. Until “how long it takes today” is measured, claiming a multiple of improvement would be misleading.
Source: goals G1–G3 in §2 and measures S1 / S2 / S6 in §2.5 of PRD/excel-ai-analyst-PRD-v2.0.md
Chapter Four · Efficiency

The largest gain: do not start over next month

BEFORE · START FROM ZERO EACH PERIOD

Explain everything again

Analysts hand-write every validation rule for every workbook;
business experts explain the same definitions again;
a new owner takes over → start from the beginning.

NOW · CONFIRMED RULES ACTIVATE AUTOMATICALLY

0hand-written validation config lines

Rules confirmed by business experts in the previous period are automatically added to this period's validation list
and aligned to the current workbook using the saved column mapping.

The self-test fixes this claim precisely: generate a draft directly from the original workbook, with zero analyst-authored table checks; the four injected rules execute, and the seeded error in the demo workbook is still detected. This shows that the rule was not aligned to the wrong column. MEASURED
Source: assertions #44–#47 in examples/mvp_selftest.sh (real path: generate a draft directly from the workbook) · Two additional checks are suggested automatically from the table pattern, so the analyst does not write those either
State the evidence strength accurately: “previous period / current period” in the self-test uses the same file and same sheet. The mechanism itself is period-independent because it reapplies the stored column mapping, but cross-file and cross-period behavior has not yet been demonstrated; that requires a second period of real data
Chapter Four · Efficiency

“Zero configuration” once failed because of one detail

A rule uses an alias such as DEPT, but only the original spec knows which column that alias means. A changed workbook next period cannot run if the alias no longer aligns.

Guess by name → miss or misalign

The rule expects DEPT, while the workbook column is named “Department”; the names do not match.
Worse, “Amount” could be matched to “Amount including tax,” causing the rule to test the wrong column silently and report a clean ✅.

Retain the mapping → align, or explain why not

Store “DEPT = Department” with the rule and reuse it next period.
If alignment fails, name the failure and never guess, then list every available column for a person to select.

These failures share one trait: nothing reports an error if they are ignored; the rule simply never takes effect. The tool therefore reports the problem at the first moment it can.
Source: ontology_registry.py payload.aliasColumns · excel_ai.py align_aliases_to_columns() · self-test #44–#49 MEASURED
Chapter Four · Efficiency

Scale: multiple workbooks across dozens or hundreds of business units

18
subcommands
Six-step method + logic layer + reporting domain pack + ontology awareness
19
table-level validation rules
Hierarchical totals, cross-sheet flattening, period variance, effective-dated foreign keys, sign direction, and version variance
9
table-pattern families
Payroll / finance / inventory / sales / budget / project / production / master data / management reporting
57
reporting-domain knowledge records
14 objects + 10 relationships + 15 rules + 6 terms + 12 questions awaiting confirmation

Check external links before analysis

Report broken links, circular references, and wide-range references together. If links are broken, request the original workbook first. Ten minutes here costs less than three hours of guessing later.

Batch processing continues

When dozens or hundreds of business units run together, one failure does not stop the batch; the summary names the failed items.

Every variance includes an investigation direction

Each variance identifies a routing layer and attribution candidates: check internal consistency before external reconciliation. Most mismatches originate in those internal steps.

All figures come from the code at the time of this deck: excel_ai.py · references/patterns/ · executed import --from domain MEASURED
Chapter Four · Efficiency

The engineering gain: make real scale runnable

Sanitized packaging once consumed disproportionate memory: a 17 MB workbook occupied 1.7 GB. With dozens or hundreds of business units in real reporting, the process would exhaust memory immediately.

BEFORE

1762 MB

73.5 seconds · about 104× amplification
The entire workbook remained in memory

AFTER

215 MB

32.6 seconds · one-eighth the memory and more than twice as fast
One streaming pass retains only bounded data

After the change, outputs remained byte-for-byte identical: headers, formula templates, external-link graph, shape statistics, synthetic samples, and even rejection reports were compared. Performance work must not change conclusions; that is a hard requirement. MEASURED
Source: “Unreleased · Performance” in CHANGELOG.md · Three root causes: avoid a whole-workbook object graph, process merged cells by region count, and locate raw-value provenance on demand
Chapter Four · Efficiency

Efficiency depends on preserving correctness

55
assertions · knowledge flywheel path
Extract → review → retain → reuse end to end
139
assertions · reporting domain pack
Six subcommands + six new checks + regression gate
24
assertions · logic layer
Reverse tracing / sensitivity / ontology merge
5/5
seeded demo detection paths hit
One seeded issue in each of 4 sheets · all 5 detection paths hit. The modified H7 was independently caught by row recalculation and formula-break detection
Nine additional unit tests cover evidence chains and snapshot immutability.
Every change runs a set of regression assertions that compare pass rates, record counts, and table-level exception sets item by item; specifically checking that new features do not change prior conclusions.
Executed counts MEASURED · examples/{mvp,m,e1}_selftest.sh · run_demo.sh · Note: assertion count is not test coverage
Chapter Four · Efficiency

What has been measured, and what remains a target

ClaimStatusEvidence
Confirmed rules activate next period with zero analyst-authored config lines MEASURED Self-test #44–#47 on demo data
19 checks / 18 subcommands / 9 table patterns / 57 imported knowledge records MEASURED Counted directly from the code at the time; 57 is the executed import result
Sanitized packaging uses one-eighth the memory with byte-identical output MEASURED 17 MB workbook, 1762 MB → 215 MB
One monthly consolidation batch in < 30 minutes TARGET Measure S1 · the current-state baseline has not itself been measured
At most 3 attribution candidates per variance; top-candidate hit rate ≥60% TARGET Measure S2 · requires a three-month retrospective test
A new analyst can rerun independently within one day TARGET Measure S6 · baseline is a two-hour recording plus months of exploration
Review throughput ≥30 items per hour TARGET Measure S4 · otherwise knowledge remains in draft
The second review has fewer items to confirm than the first TARGET Milestone MS6 · the real success criterion, requiring two periods of real data
The first three rows show that this implementation worked; the last five would show that the method compounds. None of the second group has been achieved yet. That is the evidence boundary of this deck.
Measure definitions: §2.5 of PRD/excel-ai-analyst-PRD-v2.0.md · Milestones: project development-plan documentation
Chapter Five · Security

Not one byte of the most confidential data leaves the internal network

The meeting was explicit: “Our most confidential data is all here” and “the entire computer can remain off the public internet.” The system was therefore designed as two tracks from the start.

        INTERNAL MACHINE (DATA STAYS IN-DOMAIN)       SAFE TO SEND TO AN EXTERNAL MODEL
        ┌────────────────────┐
        │  Original workbook  │  ← Always remains here
        │     │               │
        │     ▼               │
        │  Sanitized package  │ ──────▶  Multi-row headers and column shapes
        │                     │          Formula templates for each column
        │  Numeric validation│          Cross-workbook dependency graph
        │  (local, offline)    │          Row/column scale and orders of magnitude
        └────────────────────┘          Zero raw amounts / names / client names
This does not rely on personal discipline. Packaging performs two self-checks: scan memory before generation, then read the bytes back from disk and scan again. Detecting any raw value causes the package to be rejected, leaving no file in the directory that could be sent externally.
Chapter Five · Security

Three boundaries that cannot be crossed

BOUNDARY ONE

Do not send original workbooks to external models

Structural information is sufficient for AI to reconstruct formula chains, draw lineage, and formalize rules.
Numeric validation runs inside the internal network; it needs neither internet access nor AI.

BOUNDARY TWO

The anonymization “key” does not travel

The column-name restoration mapping is isolated separately and excluded from the sendable list.
Packaging uses a file allowlist instead of “zipping the directory.” The latter is easier, but also much riskier.

BOUNDARY THREE

Knowledge records cannot carry data either

The ontology stores structures and rules. If a verbatim statement containing a specific amount enters the knowledge base, it could propagate into cross-workspace reuse or external transfer.
Every import is scanned and the entire batch is rejected on detection.

Even rejection reports follow the boundary: report only digit count and order of magnitude, never echo the original value. Echoing it would simply copy the sensitive number into another log.
Chapter Six · Boundaries

What the system cannot do today

The table below records the current limits.

Unavailable / unfinishedWhat it means
Not yet run on real management-reporting data All figures come from constructed demo data. Real scale, broken links, and dirty data have not yet been encountered
Steps 2 / 3 cannot be automated AI must read the supporting material and author field meanings and formula chains; scripted guessing would harden misunderstandings
AI cannot approve or publish This is an intentional governance constraint, not a capability gap. A human must remain in the loop
Automatic exception promotion is not implemented The “promote only after ≥3 consistent occurrences” discipline is currently enforced by people, not the tool
Multi-user team spaces and permissions are not implemented The current local fallback treats each directory as a workspace. A unified entry point and account system depend on company infrastructure
Flywheel compounding is unvalidated Two periods of real data are required to learn whether the second review is genuinely easier
Source: “What this version does not include” in CHANGELOG.md · “Known boundaries” in runtime/registry/README.md
Closing

Three actions can begin in parallel

BUSINESS TEAM

Name one point of contact

The meeting stated: “Just name one person as the interface.”
That business-expert contact is the only role that can confirm definitions. Knowledge cannot enter the platform without an accountable business owner. this is a hard validation, not a process suggestion.

COLLEAGUE WITH DATA ACCESS

Provide originals with external links intact

One command creates the workspace; another completes external-link mapping, structural profiling, and sanitized packaging.
The workflow needs originals, not values-only copies.

BUSINESS EXPERT

Hold one review meeting

Double-click a web file, review one item per screen, and export when done.
This step does not need to wait for data preparation; review and data processing can proceed in parallel.

The first review should resolve prerequisite questions and blocking rules; the rest may wait. Its goal is to establish confirmed items that the next analysis can reuse.
Closing

Turn experienced judgment
into knowledge the company can retain.

No matter how fast a tool runs, it only finishes this month's work.
The largest efficiency gain is avoiding the same explanation next month.

Use plain language

Write “who owns each issue and how to fix it,” not a pile of metrics.

Evidence first

Do not conclude before validation runs. A 98% pass rate is not a pass. A profile always calls itself a “machine hypothesis.”

Do not exaggerate

Say clearly what cannot be calculated and never pretend an unavailable capability exists. That applies to this deck too.

LaurenceYang · Excel Intelligence & Controls · excel-ai-analyst · The operating-manual documentation contains role-specific step-by-step instructions · Every figure in this deck can be checked against the source artifacts
← → Navigate · F Full screen · Click the right half to advance
← Technical Evidence
1/24