A client sends over a portfolio of forty leases and asks four questions that turn into eight once you write them down properly. Break-right mechanics, rent review pattern, alienation, repair obligation, service charge cap, insurance, guarantor position, and whether the Landlord and Tenant Act 1954 security has been excluded. Forty rows, eight columns, three hundred and twenty answers. What decides whether that grid is usable is not the model. It is what sits behind one cell.
Why the arithmetic favours the grid
Reading forty leases end to end is roughly two working weeks for a paralegal, and produces a summary written from memory of the last fifteen documents. The comprehension is real, but it is spent evenly across forty documents when the advice will turn on four or five of them.
The grid inverts that. It asks the same eight questions of all forty, then lets the outliers announce themselves, and the human attention goes to the documents that differ. That is the same amount of reading, redistributed. Nobody reads less overall; the reading happens in the places where it changes the answer.
What has to be inside a cell
A cell containing only an answer is a claim, and a claim cannot be supervised at scale. Three hundred and twenty claims from a system that is right most of the time is not a work product, because you cannot tell which of them are the exceptions without redoing the work.
So a cell has to carry its provenance: the answer, the passage it came from, and a route back to that passage in the document. Verified against source, at cell level, means the quoted text was located in the firm's own uploaded file. It says nothing about whether the clause is enforceable or good for your client, and blurring those two claims would be the easiest way for a product like this to mislead a reviewer.
The two checks that must not be confused
Grounding and citation verification are separate mechanisms doing separate jobs. Grounding confirms that a quoted clause appears in your lease. Citation verification runs on public authorities: where a cell's answer cites the Landlord and Tenant Act 1954, that authority is checked against the fixed list of 74 approved public sources, and carries a verdict of verified, needs a check, or not found.
One tells you the words are in the document in front of you. The other tells you the statute says what the answer claims. A reviewer who understands which check produced which mark can move quickly. A reviewer who thinks there is one general notion of correctness will either trust too much or check everything, and both are expensive.
A sampling rule for the supervising solicitor
Three hundred and twenty cells cannot each be checked by hand, and pretending otherwise produces a rule that is ignored on the first afternoon. A workable rule has three parts. Check every cell that drives a number the client will rely on: every break date, every rent review date, every cap. Those are the cells where an error becomes advice.
Then check every blank. A blank cell means the question was not answered, and the reason matters: the lease is silent, which is a finding, or the clause was missed, which is a defect. Finally, sample one in ten of everything else, chosen at random rather than by eye, because the cells that look worth checking are not the ones that catch a systematic problem.
Playbooks, and when to cut columns
Reviews run against a playbook, which is the set of questions written for a document type. The value of a playbook is not that it saves you writing questions. It is that it stops the grid being shaped by whatever the last reviewer happened to remember at the time.
Cutting columns is usually the better decision. Eight questions across forty documents is readable. Twenty questions across forty documents is a spreadsheet nobody opens twice, and the eleven columns nobody reads cost real money to produce. A credit is roughly a penny of model usage, so an over-wide grid is a bill for reading that will not happen.
Where humans still read
Three places, reliably. The outliers the grid identifies, which get read properly, end to end, because they are where the risk sits. Anything the grid could not answer, since silence in a document is a legal question and not a gap in the data. And the document the deal actually turns on, which a supervising solicitor should always read personally regardless of what any cell says about it.
The output is a memo you can work with rather than a chat log: Summary, Files, Action items, Authorities, Open questions and How this was made, with a Word download. Read the open questions first. They are what the run could not settle, which is a better guide to where your afternoon should go than the summary is.
The honest edge
A grid answers the questions you asked and only those. Ask the wrong eight and you get a confident, well-evidenced grid full of the wrong answers, and nothing in it will tell you that nobody asked about the guarantor's covenant strength. Choosing the columns is legal work, and it is the part of the exercise that changes the outcome most and involves the least software.
Documents come in by upload. There are no live connectors to iManage, NetDocuments or SharePoint; those are catalogue tiles. There is no redlining and no Word add-in, so the drafting happens in Word after the download. And a privilege call still belongs to a person: a flag is a prompt to look, never a determination you can rely on in a disclosure exercise.
