Classification and Scoring Columns

A project does not have to stop at extracting facts. The same grid can run a review: classify every contract against categories you define, check whether the controls you care about are actually in the document, and turn those answers into a single rating you can report on.

That takes three column types beyond the plain extraction ones: a List with your own options, a List shown as a checklist, and a Computed status that scores the others. This guide covers all three, plus the fourth derived type, Compare to reference.

Five types ask the AI a question about the document. Two are derived: they are calculated from other columns or from a reference document, and one of those uses no AI at all.

You pick the type in the column editor, next to the Question. A Computed status column has no Question: it does no reading, so there is nothing to ask.

Choose List and add the answers you will accept under Options. The AI must answer with one of them, or leave the cell blank when none apply. That is the difference between a classification and free text: the column stays sortable, countable, and comparable across every contract.

• Add option adds a row. Type the label exactly as you want it to read in the grid.
• Allow multiple selections lets one contract land on more than one option.
• The swatch beside each option sets its pill colour. Leave it on Auto for distinct colours automatically, or set them deliberately: green for the benign options, amber for the ones worth a look, red for the ones that need review. A traffic-light column is readable in one pass down the grid.

A common pattern is a pair of columns rather than one. A single-select column holds the primary category, so every contract has exactly one answer you can count. A multi-select column beside it holds the secondary categories that also apply. That keeps the reporting clean without losing the nuance.

When the options are not categories but a list of things you want the contract to have, turn on Allow multiple selections and set Show as to Checklist (X of N).

The cell then shows a coverage badge instead of pills, for example 3 of 5, and hovering lists which items were met and which were not. Use it for questions like “which of these protections does this contract actually contain?”: confidentiality obligations, restrictions on onward use, consent requirements, audit rights.

The question still matters. Ask one clear thing and let the options carry the detail:

Which of the following contractual controls does this agreement contain?

A Computed status column takes columns you already have and calculates a rating from them. It is deterministic arithmetic, not a second AI pass, so the same inputs always give the same answer and the result cannot drift between runs.

Under Checks, tick the columns that should feed the score. Only two kinds of column are eligible:

• Yes / No columns. Use Pass when: Yes or Pass when: No to say which answer counts as a pass. An empty or not-found cell counts as No.
• Multi-select list columns with options, that is, a checklist. Set Passes at ≥ N of M for how many items must be met to count as a pass.

Single-select classifications are deliberately not eligible: any non-blank answer would count as met, which would make the check meaningless.

For each check you can set:

The score is the weighted percentage of checks that pass, banded into four tiers:

The cell reads like Mostly compliant (80%), and its reasoning spells out the arithmetic: “Passed 4 of 5 checks (80% weighted). Failed: Audit rights.” A failed must-pass check shows the tier with no percentage, because a high score next to a hard failure reads as a contradiction.

A computed cell is only calculated once every check it depends on has settled. While inputs are still running it shows Waiting for checks…. If an input errored, was never analyzed, or went outdated because you edited its question, the computed cell names the check that needs attention rather than scoring around the gap.

You can also ask the AI for a rating directly: a single-select List column with options High risk, Medium risk, Low risk, and a Question that spells out your matrix. That is a legitimate approach and often the quicker one to prototype.

There is a subtlety worth understanding first. Every column a contract needs is answered in a single pass, left to right, so by the time the AI answers your rating column it has just written the earlier answers and can take them into account. On a first full run, a rating column that refers to your other columns usually behaves the way you would hope.

It is not a dependency the project actually models, though, and it stops holding in the cases you will meet soonest:

• Re-running one column. Editing a question and re-running touches only that column, so the rating is answered from the document alone, with none of the other answers in front of it.
• A column added later. Running it across contracts that are already analyzed does the same.
• Corrections. A value someone edited or verified by hand is never fed back into a later extraction.

So an AI rating column is fine for prototyping, as long as you re-run the whole row rather than the single column whenever you change it. When the rating has to be reproducible, has to respect corrections, or has to stand up in an audit, use Computed status: it reads the stored values of the columns you nominate, including any a person has corrected, and calculates the result rather than inferring it.

Compare to reference classifies each contract against one or more reference documents, your template or your master agreement, rather than answering a question. Add the documents under Reference documents. Each cell comes back as On template, Template modified, or Counterparty paper, with up to five material deviations listed on hover.

Classification and scoring give you a first draft, not a finding. Every cell still carries its reasoning and source quote, and any cell can be confirmed or corrected by hand: double-click it, then save to stamp it with a tick, your name and the date. See Checking and Trusting Results.

A review that asks “which of these agreements need a closer look, and why” usually lands on four columns:

1. Primary category. List, single select, options coloured green through red. “Which single category best describes the subject matter of this agreement?”
2. Other categories. List, multiple selections, same options. “Which other categories does this agreement also touch?”
3. Contractual controls. List, multiple selections, shown as Checklist (X of N). “Which of the following controls does this agreement contain?”
4. Rating. Computed status, with the controls checklist as a check (Passes at ≥ 3 of 5) and any deal-breaker Yes / No column marked Must pass.

Build it on five or ten contracts first. Read the reasoning on each cell, and tune the option labels and questions before running it across the portfolio. Editing a question after a run marks that column’s cells outdated and they need re-running, so it is cheaper to iterate small.

• Writing Column Questions
• Checking and Trusting Results
• Analyze Contracts
• Export Project Results
• Projects Overview