Workflow Automation · Read + Write

OCR documents → Excel

Automate Rep · High Reason · Med

The problem

Finance / ops receive PDFs and scans full of tables — bills, statements, certificates, lab reports. People copy-paste rows into Excel for hours. Errors, OT, audit risk.

The AI approach

Multi-page table IDP. Detect tables, extract rows with vision-LLM, normalise columns into a single sheet schema, validate types, output a clean Excel.

The outcome

~30 sec/doc

200-page bundle goes from ~1 day → ~10 min. Output is one tidy spreadsheet, not a paste-from-N-PDFs mess.

Try itInput → Process → Output

Input — multi-page PDF bundle

p.1 / 8

Lab Report — Plant A

Sample	pH	NTU
S-101	7.2	1.4
S-102	7.4	1.1

p.4 / 8

Lab Report — Plant B

ID	pH val	turbid.
B-220	7.0	1.8
B-221	6.9	2.1

p.7 / 8

Lab Report — Plant C (scanned)

sample	pH	turb
C-008	7.3	1.5
C-009	7.2	1.6

Process — AI pipeline

1Detect tables on each pageReadPredictive

2Extract rows (vision-LLM)ReadGenerative

3Normalise columns to schemaRulesSymbolic

4Type-check + write .xlsxWriteSymbolic

Output — consolidated Excel

Click Run demo to flatten the multi-page bundle into one spreadsheet.

lab_results_consolidated.xlsx1 sheet · 6 rows · normalised

.xlsx

plant	sample_id	pH	turbidity_NTU	source page
A	S-101	7.2	1.4	p.1
A	S-102	7.4	1.1	p.1
B	B-220	7.0	1.8	p.4
B	B-221	6.9	2.1	p.4
C	C-008	7.3	1.5	p.7
C	C-009	7.2	1.6	p.7

schema-validated · pH ∈ [0,14] ✓extracted in 12s

Three AI types in this use case

SymbolicColumn-alias dictionary; schema validator (type checks, ranges); .xlsx writer.

PredictiveTable detection on each page; layout classifier picks the extractor strategy.

GenerativeVision-LLM extracts rows from heterogeneous tables and reconciles column names across pages.

The stack

Table detection · Textract / DocAI / Camelot
Row extraction · Vision-LLM (Claude / GPT-4o)
Schema mapper · Pydantic + column-alias dict
Writer · openpyxl / xlsxwriter

When this works

Inputs share a core schema even if column names vary
Tables are visually clear (printed, not heavily handwritten)
Volume is high enough — 100+ docs / week
Downstream consumer is Excel-shaped (analyst, audit)

When it fails

Merged cells + multi-row headers — extraction misaligns
Stamped/handwritten overlays change values
Documents in multiple languages with no glossary
Schema drift — new column appears, mapper silently drops it