Where AI Fits In Discovery Preparation

A 90 day discovery window with millions of documents rarely fails because the team cannot read. It fails because the team cannot prioritize, measure, and explain its review decisions fast enough to meet production dates and defensibility expectations. In that setting, AI is most valuable when it produces concrete intermediate work product the litigation team can route, sample, and validate. Think AI issue tags on email and chat threads, AI clause or communication summaries for quick orientation, AI suggested custodians and entities for early case assessment, and AI timelines that stitch together what happened when across sources. The attorney remains accountable for what gets produced, what gets withheld, and how the process is documented.

Courts accept Technology-Assisted Review (TAR) as a court-accepted methodology for AI assisted eDiscovery when the workflow is defensible. In this course, we treat TAR as part of discovery preparation rather than a black box. The point is not to replace review. The point is to move attorney effort up the stack toward setting review objectives, training and iterating the model, and validating the results with sampling so production decisions can be defended. We will stay anchored in what the team relies on. That includes AI responsiveness predictions, AI privilege risk flags, and AI proposed review batches. Each of those outputs has a different verification depth depending on risk tier and where it will be relied on.

In case preparation outside document production, AI is most useful when it turns unstructured evidence into litigation ready scaffolding. That means an AI case timeline excerpt tied to Bates stamped documents, AI deposition outline building blocks grounded in specific exhibits, and AI gap lists that identify what the current record cannot yet prove. The team can then assign follow up tasks that are actually actionable. Request the missing documents, run targeted searches, or interview the right custodian. The skill you will build here is not asking for a generic summary. It is directing the model to produce review artifacts that map cleanly to your existing workflow and record keeping. Let’s look at a concrete example of the kind of AI timeline output teams use in early case prep.

Professional duties as workflow requirements

In litigation, professional duties become operational requirements the moment AI touches client data or influences a production decision. The duty of competence means you must understand what the AI output is and is not. If you cannot explain how a TAR workflow was trained and validated, you cannot defend it when challenged. The duty of confidentiality means the tool boundary matters. What leaves the firm environment, where it is stored, who can access it, and whether it is retained for model training are not technical footnotes. They determine whether privileged or confidential material was exposed. The duty to supervise means you cannot delegate judgment to the tool or the vendor. You direct the review objectives, approve the protocol, and ensure quality control checks are run and documented.

Professional responsibility checkpoint
Do not put privileged or confidential client material into non enterprise AI tools. Use enterprise approved review platforms with defined access controls, retention, and contractual limits on vendor use of client data. If data governance is unclear, pause and route the tool choice to the supervising attorney before upload or analysis.

A practical way to think about this is that discovery AI is either inside a controlled review system or outside it. Inside a review platform, TAR outputs can be tied to collections, custodian lists, audit trails, and coding decisions. Outside, a consumer AI tool may store prompts and outputs, use them to improve its models, or make retention and deletion hard to verify. That gap is where privilege exposure becomes case dispositive. You can still use AI aggressively, but you do it where the controls match the stakes. Next we will compare two tool setups the way a litigation team evaluates them before deploying AI on a matter.