In a FELLOW pilot for a mid-sized services company, the estate held about 5,900 Excel workbooks. One year's unique files contained roughly 16,000 sheets. The job was to connect commission statements, revenue schedules, payroll requests and plan terms into records finance could check.

The first year passed automated testing in seven days. Getting there took decisions about what each sheet meant, which version to use and what the evidence could support.

The file mix

One year's folder held about 1,800 standard Excel workbooks and 300 macro-enabled workbooks, alongside roughly 380 Word documents and 230 PDFs. Spreadsheets carried most of the operating detail. Word and PDF files supplied the plans and approvals around it.

The estate contained:

  • Multi-tab workbooks with monthly statements and accumulated history.
  • Visible and hidden sheets, old models and empty templates.
  • Formula columns interrupted by typed numbers, shifted headings and mixed units.
  • Legacy spreadsheets that needed conversion, and a corrupt workbook that needed replacement.
  • Word plans, signed PDFs and a supporting PowerPoint overview.

We recorded the overview at file level. That did not show every deck had been extracted. The same discipline applied throughout: inventory, capture and usable structured data are separate states.

The folder is a delivery record

Files in one year's folder could describe a different year. Sorting by location alone would mix reporting periods.

FELLOW compared file contents to find identical copies and set data periods from the contents. Hidden sheets entered the inventory too. Some held relevant inputs, others were stale models. Each received an explicit outcome.

One corrupt workbook could not be read as delivered. We recorded the missing source and the need for a clean copy. Discovering a file never counted as loading its data.

Tabs need their own interpretation

A workbook can hold a statement, an invoice listing and a rate table. Those sheets need different interpretations even when their names look alike.

FELLOW recognized layouts from headings and other content. We checked for sheets claimed by two recognition rules and refined the rules until those clashes reached zero. A tab holding two tables needed separate handling for the second table.

Capture needed an audit too. An early row limit cut off longer grids. Later work found limits in two statement layouts. We corrected the limits and reloaded the affected data. The lesson: measure the output against the source, including the places where capture stops.

Version history changes totals

Cumulative workbooks carry earlier months forward. A statement can recur unchanged, return with revised figures or move from estimate to actual. An identical file copy is only one form of duplication.

Across 203 tables with version tracking, summing all versions produced about 2.5 times the current-version total. That ratio compares database sums across versions of the same records. It says nothing about amounts paid.

FELLOW kept the history and labeled repeated rows by what happened to them. Analysts could then start from current-version views. We separated detail rows from subtotals and identified overlapping tables, so the same invoice would not be counted through several routes.

Some decisions still needed the owner to pick an authoritative source. Version tracking exposed that choice. It did not remove it.

Headings and formulas need checks

A percentage heading could sit above currency values. One column could mix fractional percentages with whole points. A shifted layout could put an identifier where an amount belonged.

FELLOW assigned explicit units, tested values against them and kept the source context. Uncertain currency assumptions stayed visible. We also classified row types, so a quota or subtotal never became an ordinary line item.

The formula review explained 11,113 distinct patterns. It flagged overwritten formulas, broken references, external links and calculations that changed within a column. A flag marks a place for a reviewer to investigate.

Reported values stayed as reported. Recalculations were stored separately, so a reviewer could compare the workbook's result with our interpretation of what it should have produced.

Names need rules

Statements used nicknames, initials and surnames alone. Some employees shared a name. Some apparent names were totals or placeholders.

FELLOW matched records using employee identifiers, known aliases and reviewable rules. Ambiguous matches stayed unresolved, with the printed text retained. We linked plans to the periods they covered and labeled inferred assignments.

A statement, a request, an approval and a ledger entry describe different events, and the database kept them apart. Where the ledger was absent, we recorded missing evidence and left the request's status open.

A pass needs a defined denominator

Across both years, FELLOW captured about 20 million numbers and recorded outcomes for about 37,000 sheets. The database held about 17 million rows in named-column tables, including repeated workbook versions. These figures measure different things: numbers captured from the workbooks, sheets with a recorded outcome, and rows loaded into tables. A sheet could be loaded, summarized, retained for reference or excluded with a reason.

For the first year, 99.95% of captured numbers reached named columns. Counting numbers retained separately with labels, 100% were accounted for. The denominator is captured numbers. It does not cover every possible value in every delivered file.

Both years passed automated testing. Open business questions remained visible.

Retrieval ends at evidence

The database answers practical questions: which plan covers this period, does the statement agree with the request, and why is a figure missing?

FELLOW kept source links on those records down to the workbook, tab and cell. In a check against original workbooks, 239 of 240 sampled values matched their source cells. That is a sample result. It does not mean every figure was independently checked.

For finance, the deliverable combines three things: current records to query, exceptions to review and evidence to inspect. Thousands of spreadsheets become manageable when each has a defined place.