A custom column is any column whose value the AI produces for you. That value can come from one of three places: read straight from the document, calculated from what the document shows, or inferred from context when nothing is printed. All three are set up the same way — a column name plus an optional Format Requirement — with no separate mode to switch on.
The three ways a column gets its value
Every column produces a value. Which kind of column it is depends on what you ask for, not on a setting.
| Kind | What the AI does | Example |
|---|---|---|
| Direct | Reads the value that is printed on the document and returns it as it appears. | Vendor Name |
| Computed | Calculates a value instead of copying one — for example an amount the document does not print, or a check against a figure that is printed. | Line Total (Qty × Unit Price) |
| Inferred | Judges a value that is not written anywhere, using the document as a whole. | Category (options: Meals, Transport, Office, Other) |
Direct is the default: if your instruction only names a value, the AI reads it. A calculation or an inference happens only because your instruction asks for it. All three kinds are written in the same two fields, so there is nothing to toggle between them.
Direct extraction
The AI finds the value on the document and returns it. Unless the Format Requirement asks otherwise, it does not calculate, clean, or rewrite the text, so what you get is what is printed. If the value cannot be found, the cell is left empty rather than guessed.
Computed columns
Ask for a calculation in the instruction, and the AI performs it while it is already reading that file — in the same pass as the extraction, not as a separate calculation step afterward. The result arrives as an ordinary cell in the table, so there is no formula column to add later.
Examples of what you can ask for:
- A line total that is not printed:
Line Total (Qty × Unit Price). - A comparison against a printed figure:
Verification (OK if Calculated Total = Billed Total, else the difference). - Deducting listed charges, applying a rate you supply in the instruction, or a difference between two dates.
The calculation uses what the AI can see on the file it is processing, so it is a snapshot of that file rather than something recalculated later.
Inferred columns
An inferred column asks the AI to decide something that is not written on the document, using the document as a whole. You give the column a name and describe how to decide, including the allowed answers when you want the result constrained. For example, Category (options: Meals, Transport, Office, Other) on a receipt fills in the right category even though no category line is printed. The inference happens in the same pass as extraction, so it works across a whole batch, not just one file.
Where you write the instruction
Columns are created and edited in the Data Columns panel. Opening a column's edit action shows the Edit Rule dialog with two inputs:
| Field | What it controls |
|---|---|
| Data Field Name (Extraction Instruction) | The column's label, and the instruction for what to find or produce. Keep it short when the detail lives in the Format Requirement. |
| Format Requirement (Optional) | How the value should come back, or how to produce it — the calculation or the decision rule. It applies to this column only. |
There is no separate field that says “computed” or “inferred”. The kind of column is expressed entirely by what the instruction asks for, which is why all three kinds share the same two inputs. For example: column name Line Total with the Format Requirement Quantity multiplied by Unit Price.
You can also put the calculation in the column name itself — Line Total (Qty × Unit Price) — and leave the Format Requirement empty. This is the usual approach in the no-login demo, which has a single Columns to Extract box and no Format Requirement field. On the dashboard, keeping the calculation in the Format Requirement and the name short is easier to reuse.
See Templates & Columns for how columns are created, edited, reordered, and saved into a template.
Current limitations
- Combining many values can sometimes be wrong. The hardest calculations are the ones that need several values combined — a section total, a multi-week sum, or several operands in one expression. These can occasionally return an incorrect number.
- A calculation cannot reach across files or pages. It is done on the one file being processed and only sees the values on that file. If the numbers you need are split across two pages or two files, the column cannot combine them by itself; extract each input as its own column and combine them after export instead.
- You have to ask for the calculation. Without an instruction that requests it, the AI treats the column as a direct read and does not compute or clean the value on its own.
Calculation-heavy documents: the Precision+ toggle in the toolbar asks the model to spend more time reasoning before it answers (its tooltip reads “Best quality, slower speed”). Turn it on when a column needs several values combined — for example, summing a column of line amounts — and accept a slower run.
Common questions
Does a computed or inferred column cost more credits?
No. A run is charged per processed file, at the rate of your account's Model Tier — Standard, Advanced, or Premium. The number of columns you ask for does not change that: adding a computed or inferred column costs nothing extra on top.
Do I have to extract the source values as their own columns too?
No. A calculation can use values that are visible on the document even if you did not create separate columns for them. Extracting the inputs as their own columns is still useful when you want to check the calculated result against them.
What if the AI cannot determine the value?
The cell is left empty rather than filled with a guess, so a blank is not mistaken for real data. If an inferred column comes back empty often, add more detail or the allowed options to the instruction.