How to Use Custom Column Extraction:
Get Only the Data You Need, Not the Whole Page
Most document extraction tools work like a photocopier for text: they scan the page and dump everything they find into a spreadsheet. If you need three fields from a purchase order — PO number, supplier name, and total — you get all 47 text strings on the page, and then you have to manually pick out the three you wanted. Custom column extraction reverses this logic. You tell the AI what you want first, and it brings back only that.
Why "Extract Everything" Is the Wrong Starting Point
Traditional OCR answers one question: "what text is on this page?" It converts an image of a document into a stream of machine-readable characters — every word, every number, every stray watermark and page number. For a one-page invoice, that might be 50 to 80 separate text strings, dumped into a spreadsheet with no structure and no labeling. Your job then becomes: read through the output, identify which text string is the invoice number, which is the date, which is the total, and copy them into your actual working spreadsheet.
This is why "OCR all text" is rarely the right answer for business document processing. You almost never need every piece of text on a document. You need five to ten specific fields. The rest is noise — and the time you spend filtering out that noise is the same time you were trying to save by automating in the first place.
Custom column extraction starts from the other direction. Instead of extracting everything and filtering afterward, you define your columns upfront. The AI reads the document with your target fields as a lens — searching for the PO number, the supplier name, and the total, and ignoring everything else. The output is exactly the spreadsheet you wanted, with exactly the columns you named.
The key insight: Extraction accuracy improves when you narrow the target. An AI asked to "find and extract all text" has no prioritization mechanism — every text string on the page gets equal treatment. An AI asked to "find the PO Number, Supplier Name, and Order Total" can focus its attention on locating those specific values, using semantic cues ("what does a PO number look like?") and context ("where does a total usually appear?") to improve precision. Specificity is accuracy.
Column Names Are Instructions: A Naming Strategy
The column names you type are the primary communication channel between you and the AI. A well-written column name is a precise instruction — it tells the AI exactly what kind of value to look for and how to format it. A vague column name ("Amount") gives the AI too little guidance. A good column name ("Order Total Amount" or "Invoice Total") narrows the search and increases the odds of getting the right number on the first try.
Here are the naming principles that produce the most accurate extraction results:
| Principle | Good | Better | Why |
|---|---|---|---|
| Be specific about the entity | Number | PO Number | "Number" could be anything — invoice number, line number, reference number. "PO Number" tells the AI exactly which number to find. |
| Include context about position or role | Date | Invoice Date | A document may have multiple dates (issue date, due date, delivery date). Specifying which one avoids ambiguity. |
| Use descriptive qualifiers | Address | Shipping Address | Invoices and POs often have both billing and shipping addresses. Qualifying the type eliminates mix-ups. |
| State the format you want | Amount | Total Amount (Number) | Appending "(Number)" signals the AI to extract just the numeric value and strip currency symbols. |
| Match common document terminology | Customer | Buyer Company Name | Use the language that actually appears in your documents. "Buyer" is standard in POs; "Customer" is standard in invoices. |
You don't need to follow a rigid naming convention — the AI understands natural language. "Due Date" and "Payment Due Date" will both work. But the more specific you are, the less ambiguity the AI has to resolve, and the higher your extraction accuracy will be.
Three Real-World Extraction Scenarios
The same custom column mechanism adapts to very different document types and workflows. Here are three common scenarios, each with a different column setup strategy.
Scenario 1: Invoice Processing
What you need: Invoice Number, Invoice Date, Due Date, Vendor Name, Subtotal, Tax Amount, Total Amount, Currency. These eight fields appear on almost every invoice, regardless of format or layout.
Column setup: Enter the field names exactly as listed above. For multi-page invoices with line items, add columns for "Item Description," "Quantity," "Unit Price," and "Line Total." Each line item becomes one row in the output, with the invoice header fields (number, date, vendor) repeated in every row for that invoice — making it easy to filter and sum by vendor.
Batch tip: If you're processing invoices from multiple vendors at once (e.g., end-of-month AP run), you can upload all invoices in a single batch. The AI extracts from each independently and merges the results into one spreadsheet with consistent columns across all rows.
Scenario 2: Purchase Order Processing
What you need: PO Number, Issue Date, Supplier Name, Item Code, Item Description, Quantity Ordered, Unit Price, Line Total, Requested Delivery Date. These fields cover both the PO header and the line-item detail in one pass.
Column setup strategy: POs are more table-heavy than invoices. Include all the line-item columns you need — Item Code, Description, Quantity, Unit Price, Line Total — in the same column list as the header fields. The AI understands that header fields (PO Number, Supplier Name) apply to the whole document and repeats them for each line-item row, while line-item fields vary per row within the table.
Multi-page handling: For POs that span multiple pages with repeated column headers on each page, the AI recognizes the repeated headers and excludes them from the output. A 10-page PO with 200 line items produces 200 data rows, not 10 separate tables that need manual stitching.
Scenario 3: Vendor Quote Comparison
What you need: Supplier Name, Item Description, Quantity, Unit Price, Line Total, Lead Time (Days), Payment Terms, Currency. Upload quotes from five different suppliers in one batch, and the AI extracts each using the same column definitions — regardless of how differently each supplier formats their quote.
Key advantage: Different suppliers may use different names for the same information. One calls it "SKU," another "Part Number," a third "Item Code." The AI's semantic understanding maps all three to your "Item Code" column without you needing to configure synonyms or per-vendor mappings.
Files are processed securely and not stored.
Save as a Template: Set Up Once, Use Forever
Once you've defined a column set that works well for a particular document type — say, your standard invoice extraction columns — you can save it as a template. Next time you process invoices, load the template with one click instead of re-typing all eight column names.
This is where the one-time setup cost pays off continuously. Setting up your invoice columns takes two minutes the first time. Every subsequent batch of invoices — whether it's five or fifty — starts with those columns pre-loaded. You upload the files, the AI processes them against your saved template, and you get a consistently structured spreadsheet every time.
Templates also help with team consistency. If multiple people on your team process documents, sharing the same template ensures that everyone's output has identical columns in identical order — no variation in field names, no missing columns, no formatting drift over time.
Frequently Asked Questions
How many columns should I define — is there a limit?
There's no hard limit, but there's a practical one. Up to about 15-20 columns, extraction is consistently reliable. Beyond that, you start covering edge-case fields that may not appear in every document. A focused column set of 5-12 well-named fields almost always produces better results than a 30-column exhaustive list, because the AI can give each field more focused attention. If you need many fields, consider splitting into two passes: core fields on the first pass, supplementary fields on a second.
What happens if a field doesn't exist in a particular document?
The cell is left blank. For example, if your column set includes "Tax Amount" but one of your documents doesn't have tax (common with inter-state B2B invoices or export documents), that row in your output simply has an empty Tax Amount cell. The spreadsheet stays structurally consistent — all rows have the same columns — but values are only populated where the AI found them in the document.
Can I mix different document types in one batch?
Yes, but only if your column set makes sense for all document types in the batch. If you upload invoices, POs, and receipts together with columns like "Invoice Number / PO Number / Receipt Number," the AI will try to find each field in each document. Documents that don't contain a given field will have blank cells for that column. For best results, batch documents of the same type together — all invoices in one batch, all POs in another. This keeps your column definitions focused and your output clean.
How accurate is it compared to typing the data manually?
For printed, well-formatted documents, field-level accuracy exceeds 90% — meaning 9 out of 10 values are extracted correctly on the first pass. This compares favorably to manual data entry, which has an inherent error rate of 1-3% per keystroke according to data quality research. The difference is that AI errors tend to be predictable (confusing similar field names, misreading low-resolution text) while human errors are random (typos, transpositions, skipped rows). A quick scan of the AI output catches most issues; catching your own manual entry errors requires line-by-line verification against the source document — which takes as long as the original entry.
Can the AI handle checkboxes, stamps, and signatures?
Yes — you can define columns for these. Use descriptive names like "Approval Checkbox (Checked/Unchecked)," "Company Stamp Present (Yes/No)," or "Authorized Signature Present (Yes/No)." The AI identifies these visual elements and returns the appropriate status. For signatures, the output is typically "Present" or "Not Present" — it verifies existence, not identity.