Discovery is the part of litigation where AI either earns its keep or quietly burns the case. The sheer volume — tens of thousands of documents per matter — makes manual triage impossible, but the cost of a privileged document leaking to a third-party model is not theoretical.
Set the privilege boundary first
Before any AI touches a single document, decide:
- Which model you can send what to (most teams need a self-hosted or zero-retention deployment for anything potentially privileged).
- What metadata is safe vs. what content is.
- Where the audit log lives. If you can't show, in two months, exactly which documents were processed by which model, you're not running a defensible workflow.
Triage in passes, cheapest first
Run the cheapest classifier across everything. A simple keyword + date + custodian filter discards 70% of the corpus before any LLM gets involved. Then run a cheap, fast model for first-pass relevance scoring. Only the top quartile gets the expensive model and human review.
Tip
Build the triage as a queue, not a script. Each document has a state
(unprocessed → first-pass → reviewer → final). Adding a new pass later is a
column change, not a rewrite.
Surface, don't decide
The model's job is to surface candidates: "this document mentions the disputed product line and is dated within the responsive period." A human decides whether it's responsive. The model that tells you "this is privileged" without showing the basis is the model that loses you the case.
Sample, sample, sample
Pick 100 documents the model classified as "not responsive" and have a human review them. The false-negative rate is your honest accuracy metric — false positives waste reviewer time, false negatives lose evidence. Re-run the sample after any prompt change.
What to refuse to automate
Privileged-vs-not is a judgement call with stakes. Use AI to surface candidates and flag risk indicators (attorney name, "privileged and confidential" language). Use a human to make the call. Every time.