Four surfaces, four jobs.
Most wasted effort on a legal AI platform is someone using the chat box for work that belongs in a review table. Here is the map, and the tell that you have reached for the wrong tool.
01 The chat-box default
Give somebody a powerful platform with five entry points and they will use the one that looks like the chat app they already know. This is not stupidity; it is how everybody behaves with new software. But in legal AI it has a specific cost, because the chat surface is the worst fit for the highest-value work.
If you find yourself pasting document after document into a chat window and asking the same question each time, stop. That is the signature of a job that belongs in a bulk-review surface. You are doing by hand the exact thing the expensive part of the platform exists to do.
02 The map
| If the job is… | Reach for | Because |
|---|---|---|
| Draft this clause / summarise this one contract / explain this provision | Assistant | Turn-by-turn steering, small document count, you want to argue with it |
| Find every change-of-control provision across 400 agreements | Vault | One question, many documents, and you need an auditable table rather than prose |
| What does the regulation actually say, with citations I can open | Knowledge | Grounded in licensed legal sources rather than in whatever you uploaded |
| Run our standard NDA intake, identically, forty times a month | Agents / workflows | The process is stable; the value is consistency, not creativity |
| Work a matter alongside outside counsel or the client | Spaces | Governed sharing beats emailing documents around, which is what will otherwise happen |
03 Review tables are the underused idea
The single most useful concept for transactional and litigation work is the review table: rows are documents, columns are questions, and each cell is an answer with a link back to the source text.
It is worth dwelling on why this shape matters more than a prose summary. A summary is one artefact that you either trust or do not. A table is checkable — you can spot the outlier row, sample five cells, and see immediately where the model got confused. It converts an act of faith into an act of review, which is what a supervising lawyer is actually trained to do.
The 2026 versions of these tables have grown up considerably: columns can reference other columns to build conditional logic, reviewers can comment on individual cells, and you can add manual columns so human judgement sits alongside machine output in the same grid.
You need to know which of 300 supplier contracts allow assignment without consent. You could (a) ask the chat assistant, uploading contracts in batches, or (b) build a three-column review table. Beyond speed, what does (b) give you that (a) cannot?
Think it through, then open.
An audit trail and a consistent question. In the chat approach, your phrasing drifts between batches, context from earlier documents can bleed into later answers, and when a partner asks how you reached the number you have a conversation log rather than a record.
The table asks all 300 documents the identical question in isolation, shows the source text behind every cell, and exports. Six months later, when someone asks why contract 214 was categorised the way it was, you can answer. That is worth more than the time saved.
04 Where the work actually happens
One practical note that changes habits more than any feature: these platforms increasingly meet you inside Word rather than asking you to come to them, with citations to source material visible in the document you are already editing.
This matters because the friction of switching applications is the quiet killer of adoption. A tool that requires leaving the document gets used for set-piece tasks. A tool inside the document gets used constantly. If your firm has the integration available and you are not using it, that is probably the highest-return ten minutes in this track.
This week's challenge
Take a task you did recently by pasting several documents into a chat box one at a time. Rebuild it as a review table with three columns. Time both. The point is not the time saved — it is noticing how differently you check the two outputs.