How law firms are using AI to handle discovery faster

Discovery is one of the few areas in litigation where the math has always been against the small firm. A mid-sized commercial dispute can produce tens of thousands of pages; document review at human reading speeds means weeks of associate or contract-attorney time. Most firms either eat the cost, push it to the client, or sample down to a manageable volume and accept the risk.

AI changes that math — but only when it's deployed correctly. Here's how we see it working in the firms we work with.

What AI is good at in discovery

• First-pass responsiveness coding: separating clearly responsive from clearly non-responsive documents at scale, with consistent reasoning across the set.
• Privilege flagging: catching attorney communications, work product, and common-interest-doctrine documents with very high recall — which is what matters most.
• Cluster identification: grouping near-duplicates and email threads so a reviewer reads the conversation once instead of fifteen times.
• Summary generation: producing reviewer-ready summaries of long documents so the human reviewer reads the summary first and only goes to the source when it matters.

What AI is bad at in discovery

• Borderline calls: the documents where reasonable lawyers disagree. AI will give you an answer, but the answer is no better than the prompt that produced it. These need a human.
• New legal theories: if you change the theory of the case midway through review, the AI doesn't "know" — you have to re-run the relevant coding pass.
• Anything where context outside the document set matters: AI can't weigh witness credibility or a story your client told you in confidence.

How to deploy it without a privilege incident

This is where firms get nervous, and they should. The right deployment posture for discovery looks like this:

• Use an enterprise platform with a contractual no-training commitment. Never a free consumer tool. Never a developer plan without zero-data-retention enabled.
• Deploy in your firm's tenant, not the vendor's shared environment. Most platforms support this for an enterprise tier; ask explicitly.
• Keep humans in the loop on every privilege determination. AI flags; an attorney signs off.
• Log everything. You want to be able to show, after the fact, exactly which documents the AI touched and what it said about them.

What this looks like in practice

A mid-sized firm we worked with last year cut discovery review on a commercial dispute from eight associate-weeks to under one, with no change in the volume reviewed and no privilege incidents. The associate time saved went into deposition prep — work that actually requires the associate. That is the trade we are looking for: not "AI does the lawyering" but "AI does the volume so the lawyer can do the law."