Main AI Analyze a document
TRIAGE

Twenty documents. Which one is on fire?

Triage is a different problem from analysis. Analysis asks what a document says; triage asks which document to open first, and it has to answer that before anyone has read any of them. Here is what that looks like when the reading is genuinely sequential and nobody pretends otherwise.

The honest shape of the queue

Select several files and they are queued. They are then read one at a time — the next document does not start until the previous one has finished or failed. Each document is allowed up to 90 seconds. Nothing about this runs in parallel, and no amount of copy will change that, so here is the arithmetic instead:

Documents queuedWorst case, start to finish
5up to about 7 ½ minutes
10up to about 15 minutes
20up to about 30 minutes

Those are ceilings, not averages. Twenty is the cap on a single batch, and the reason is the arithmetic above rather than a round number: past half an hour a queue that lives in a browser tab and resumes nothing after a reload stops being a batch you wait through.

What happens when one of them fails

It fails on its own row, with its own reason, and the queue continues. That sounds obvious and it is not: the common failure mode in batch tools is a stall that reports a timeout, so a document refused for a reason the system already knew — no analyses left this month, for instance — costs you the full ceiling and then names the wrong cause. Every distinguishable outcome here gets a distinguishable message.

WHY THERE IS NO “3 FINDINGS” BADGE ON A QUEUE ROW

Because it could disagree with the number on the document itself. The list that is rendered and the list that is stored are, today, two different computations in this product, and a triage badge showing one while the document shows the other would be worse than no badge. When those two lists derive from each other, the count can appear here. Not before.

Triage still needs a person

A queue can tell you which documents finished, which failed, and let you open each one. It cannot rank a caseload by urgency, because urgency is a fact about the person, not about the paper: the lease with the worst clause is not automatically the one with a hearing on Thursday. What the queue removes is the part where you open fourteen files to find out which four matter.

Every flag inside each document quotes the sentence it came from and names the rule it turns on, so the ranking you do afterwards is done on evidence you can check rather than on a score you have to trust. How to check one is a page of its own, because at volume the checking method matters more than any individual finding.

Try it with a batch you would otherwise read tonight

Queue the stack, do something else for a quarter of an hour, and see what the report looks like against what you expected.

Read a document →
Free to start · no card · who it’s for