Beyond Data Extraction:
Getting Answers, Not Just Raw Data
An invoice shows Qty = 5 and Unit Price = $12. Most extraction tools give you those two numbers in two cells. But what you actually need is 60 — the line total — and that requires opening Excel, adding a formula, and dragging it down across every row. What if the extraction itself could produce the answer, not just the raw inputs?
Key Takeaways
- Dozens of extraction tools on the market give you raw data and call it done — but a 40-line invoice still means 40 formula cells to create in Excel before any number is actually usable.
- Even the rare tools that add computation during extraction simply repackage spreadsheet formulas in a different wrapper — which means they still break on ambiguous inputs, and cross-row calculations still require you to pre-define ranges per document.
- Describe "Multiply Qty by Unit Price" in plain English, and ImageToTable.ai reasons through the document's visual structure to produce answers — handling ambiguous inputs and deriving values the page itself never prints.
The Gap Between Extraction and Answers
Standard document extraction gives you what's printed on the page. A vendor name. A column of quantities. A column of unit prices. That is useful, but it is rarely where the work ends.
The thing you actually need is almost never just what's printed. You need the line total that the vendor did not print. You need to know whether the billed total matches what the numbers add up to. You need a food cost percentage that no document contains — because the menu price lives in your pricing spreadsheet, not on the supplier invoice.
Every extraction tool solves the first problem: getting text off a page. Almost none solve the second: turning that text into something actionable without opening Excel. The gap is a calculation step — and that step scales with volume. A 40-line invoice means 40 formula cells to create, verify, and maintain. A batch of 30 invoices means 1,200 cells. A weekly AP cycle means repeating the same formulas on different documents, indefinitely.
Computed columns close that gap. You tell the AI what you want calculated — in plain English, not in formula syntax — and it produces the result during extraction. The output you download already contains the answers.
How Column Extraction Already Works
Before explaining computed columns, it is worth understanding how ImageToTable.ai extracts data in the first place — because the computation layer extends directly from the extraction layer.
The tool uses column-name extraction. You type the fields you want — "Invoice Number," "Vendor Name," "Qty," "Unit Price" — and the AI locates each value anywhere in the document by understanding what it means, not where it sits on the page. There are no templates, no bounding boxes, no training on sample documents first. You define what you want, and the AI finds it. (How this differs from generic table extraction.)
There are two inputs you can provide:
1. Requirement Directives — the column names themselves. They tell the AI what fields to extract.
2. Directive Rules — optional natural-language instructions per column (available to logged-in users). They tell the AI how to format or transform each field.
Computed columns extend this same mechanism. A column name can contain not just a field label, but a computation instruction. And a Directive Rule can describe — in plain English — any multi-step calculation you want the AI to perform on the extracted values.
Natural Language, Not Formulas
This is where ImageToTable.ai's approach to computed columns diverges from everything else on the market — and the difference determines what you can compute.
Other tools that offer calculation during extraction (and there are very few) use formula syntax. You write something equivalent to a spreadsheet formula — like @EVALUATE_FORMULA(qty * unit_price) — and the tool executes it arithmetically against extracted values. It is a formula in a different wrapper.
ImageToTable.ai uses natural language reasoning. You describe what you want: "Multiply Qty by Unit Price for this line item" or "Sum all Line Amount values within the same section." The AI vision model reads your instruction, reads the document, understands the relationships between fields, and figures out how to produce the result. It is not executing a formula — it is reasoning about what you asked for.
This distinction produces three capabilities that formula-based approaches cannot match:
Cross-row awareness. A formula operates on cell values. The AI operates on document context. When you ask it to "sum all Line Total values under the same section heading," it identifies section boundaries by visual structure — headings, spacing, grouping — and aggregates only the right rows. No range selection, no cell references, and it works across any document layout.
Ambiguity handling. A formula breaks when data does not match expected patterns. The AI reasons about edge cases: a quantity written as "Approx. 5 units," a unit price embedded in a paragraph, a subtotal labeled "Net Due" instead of "Subtotal." It uses the full document context to resolve ambiguity — the same way a person would.
Derived values from unprinted data. When a document omits a value you need — a missing line total where Qty and Unit Price are both present — a formula has nothing to compute. The AI notices the gap and derives the value from what is available.
This is AI reasoning applied at extraction time, not formula execution. And it is why the output is reliably usable without post-processing — even across messy, inconsistent, or incomplete documents.
Four Types of Computed Columns
Here are the four categories of computation you can define — each with both a column-name example (works immediately, no login) and a Rule Format example (login required, cleaner output, more control).
1. Row-level arithmetic
Multiply, divide, add, or subtract values on the same row. The most common case — and the one that replaces the most Excel formulas.
Column name (no login required)
Rule Format (login required)
2. Cross-row aggregation
Sum or average values across rows that belong to the same group or section. This is where formula-based approaches break — they require you to manually define ranges, and ranges change per document.
Column name (no login required)
Rule Format (login required)
3. Conditional logic
Ask the AI to compare values and output a result that depends on the comparison. Turns manual reconciliation into a single output column.
Column name (no login required)
Rule Format (login required)
4. Fixed-parameter references
Embed values in the instruction that the document does not contain — tax rates, menu prices, standard markup percentages. The AI treats them as given constants during computation.
Column name (no login required)
Rule Format (login required)
These four patterns cover most real-world scenarios: invoice verification, cost analysis, payroll reconciliation, estimate building, and tax computation. Each example above works today — copy the column name into the demo below and try it.
When Precision Matters: Precision+
For some computations, basic extraction accuracy is enough. Multiplying two clearly labeled numbers on the same row rarely fails. But when you ask the AI to sum values across sections, compare totals against each other, or derive a figure that requires multiple reasoning steps, a deeper level of analysis matters.
ImageToTable.ai includes a Precision+ toggle — available in the interface, enabled manually when needed. It gives the AI additional reasoning steps during extraction: the model can pause, verify relationships between fields, cross-check calculations, and confirm internal consistency before returning results.
Enable Precision+ when:
- Computing cross-row or cross-section aggregates (section subtotals, category averages)
- Running conditional logic that depends on multiple fields (comparing calculated vs. billed totals)
- Processing documents where key values are not explicitly labeled (handwritten job sheets, scanned forms with irregular layouts)
Skip it when:
- Performing simple row-level arithmetic (Qty × Unit Price on clearly labeled columns)
- Applying single-field formatting rules ("remove currency symbols")
Precision+ adds a few seconds per page. For straightforward extractions, leave it off. For anything involving cross-row logic, ambiguous formatting, or multi-step derivation, it is the difference between "mostly right" and "reliably right."
Two Paths: Column Names for Quick Tests, Rule Format for Production
ImageToTable.ai gives you two ways to define computed columns, and they serve different stages of your workflow.
| Column Name Method | Rule Format | |
|---|---|---|
| Login required | No — works in demo | Yes |
| How it works | Computation inside the column name itself | Computation in a separate JSON rule |
| Best for | Quick tests, one-off extractions, trying the concept | Production workflows, batch processing, team templates |
| Trade-off | Column headers get long; complex logic is harder to express | Requires an account; set-up takes an extra minute |
Both approaches produce the same result — computed columns in the output table. Start with column names to understand what is possible, then move to Rule Format when you are ready to build a repeatable workflow.
The system is designed so the column-name method works in the guest demo immediately. This means you can try every example from this article right now — no sign-up, no configuration. Open the demo below, paste in the column names from any section above, and upload a document.
Files are processed securely and not stored. Try adding a computed column name like Line Total (Qty × Unit Price).
What Computed Columns Can and Cannot Do
They can handle:
- Row-level arithmetic (multiply, divide, add, subtract)
- Cross-row aggregation (sum by section, average by category)
- Conditional logic (if/then comparisons, match verification)
- Fixed-parameter references (tax rates, standard prices embedded in the rule)
- Derived values not printed in the document (compute line totals when only Qty and Unit Price appear)
They cannot handle:
- Cross-document references (each extraction processes one document independently)
- Multi-system workflows (e.g., "extract data, send to accounting system, trigger GL update") — these belong in an automation layer outside the extraction tool
- Fully accurate computation when the source document is illegible (if the AI cannot read Qty or Unit Price, no computation can fix that)
The boundary is simple: if a person looking at the document could compute the answer using the information on that page — plus any fixed parameters you provide — the AI can too.
See also: invoice line item verification · payslip net pay computation · subcontractor quote line amounts · food cost percentage from invoice photos