AI document review in a UK law firm works best as a grid. The documents are the rows. The questions are the columns. Each cell holds a short answer, the passage it was taken from, and a link that opens the document at that point. You read across a column to compare forty agreements on one point, rather than reading forty agreements from first page to last.
Reading down a document and reading across a column
Reading a contract end to end is an act of comprehension: you are building a picture of one deal. Reading across a column is an act of comparison: you are trying to find the one agreement that differs from the others. A bundle of forty documents usually contains both jobs. Two or three of them decide the advice; the other thirty-seven exist to tell you which two or three those are.
The grid answers the same twelve questions of every document and lets the outliers announce themselves. A blank cell carries information too: if the column is 'cap on liability' and eight rows read not addressed, that is eight agreements with no cap, which is a finding you would want on the first page of a report.
What sits in a cell
Each cell has four parts: the answer in a sentence or two, the quoted passage it was taken from, a link that opens the document at that passage, and a 'Verified against source' chip when the quoted text was located in the file. The chip is the reason a supervising solicitor can work with the grid rather than around it.
An answer with no passage behind it is a claim. An answer with the clause attached, and a link that lands you on the clause, is something you can check in four seconds and either accept or overturn. The grid is not asking to be believed. It is offering a fast route back to the words in the firm's own file.
The playbooks and the scope picker
Five review playbooks ship with the product: NDA, MSA, M&A due diligence, Lease abstraction and Employment. Each is a set of questions written for that document type. Before the review runs, a scope picker lets you choose which documents go in as rows and which questions run as columns. Cutting the columns to the ones the deal turns on makes the grid shorter, cheaper and easier to read across.
When a cell fails, you retry that cell rather than the whole grid, and the finished grid exports, so the comparison can leave the product and go into the report or the email to the partner. Saving your own playbook is not available today.
Two separate checks, not one
Blurring the two checks would be the easiest way to mislead you. 'Verified against source' is a grounding check: it means the quoted text in that cell was found in the firm's own uploaded document. It answers whether that passage appears in the file, and says nothing about whether the clause is enforceable or good for your client.
Citation verification is a different mechanism that runs on public authorities. Where an answer cites legislation or a judgment, the authority is checked against the closed world of 74 approved domains, paywalled sources are refused, four deterministic gates run, and each authority carries a verdict of verified, needs a check or not found. One confirms that a quoted clause is in your NDA. The other confirms that a cited section exists in the statute.
A worked example
A matter holds a services agreement the firm is reviewing on the customer side and ten NDAs signed with prospective suppliers. The MSA playbook is selected and the scope cut to five questions, so eleven rows by five columns is 55 cells. Read across, not down.
On governing law, ten cells read England and Wales and one NDA reads Singapore, with the courts of Singapore given exclusive jurisdiction: that single cell is the outlier, and it took about as long to find as it took to read this paragraph. On cap on liability, nine NDA rows read not addressed and one carries an uncapped indemnity for breach of confidence. The grid told you which two documents were worth an hour, and that the other nine could be summarised in a line.
What this does not do
A grid answers the questions you asked, and only those. If the playbook is wrong for the deal, you get a confident, well-evidenced grid full of the wrong questions, and nothing in the product will tell you that you forgot to ask about anti-bribery. It does not replace reading the documents that matter; it tells you which ones those are.
Documents from connected sources cannot be used in a review today: files have to be in the matter, and the DMS tiles in the marketplace are catalogue entries, not live integrations. There is no redlining and no Word add-in. A privileged-document call still needs a human: the product runs a privilege review step and flags what it finds, but a firm cannot rely on it to decide what is privileged in a disclosure exercise.