Published 10 October 2026
Every litigation team has felt the same squeeze: document volumes keep growing, review deadlines don’t move, and the billable-hour model makes it expensive to throw more associate time at a problem that’s fundamentally about volume, not legal complexity. E-discovery and legal research are two of the most natural places to apply AI in a law firm, and they’re also two of the easiest places to apply it badly, because the cost of a missed privileged document or a hallucinated case citation isn’t a minor inconvenience — it’s a sanctions motion or a malpractice exposure. The honest answer to “what can AI do here” is narrower than the marketing around it, but the narrower version is still genuinely useful.
Why E-Discovery Became an AI Target First
E-discovery was an obvious automation candidate well before generative AI existed, because the underlying task — sorting a large, messy document set into responsive, non-responsive, and privileged buckets — is pattern recognition at scale. Technology-assisted review (TAR) using predictive coding has been defensible in court for years, built on a workflow where a human reviewer trains a model on a sample set, the model ranks the remaining documents by likely relevance, and human reviewers validate the results on a statistically sound sample. What’s changed recently isn’t the core workflow; it’s that large language models can now do a better first pass at some of the harder sub-tasks — identifying near-duplicate threads, summarizing long email chains, and flagging documents that look privileged even when they weren’t marked that way — reducing the volume a human has to touch before validation.
That’s a meaningful efficiency gain, but it’s an evolution of an already-validated process, not a replacement for it. Firms that try to skip the TAR validation step because “the AI already filtered it” are the ones that end up explaining a production gap to opposing counsel.
What Automation Actually Handles Well Right Now
Within that validated framework, there’s a lot AI genuinely speeds up: early case assessment, where a model scans an initial document pull to give counsel a rough sense of what’s in the data set before review even starts; deduplication and thread consolidation, so reviewers aren’t reading the same email forty times across forty custodians; chronology building, pulling dated events out of a document set into a timeline a litigator can actually use; and first-pass privilege screening, flagging communications involving known counsel or legal-sounding language for closer human review rather than claiming privilege determinations on its own. In each case, the pattern is the same: AI does the triage, a lawyer makes the call.
Where Attorney Judgment Still Has to Stay in the Loop
Privilege calls, responsiveness determinations on close or ambiguous documents, and anything that becomes a representation to a court all need to stay with a lawyer, full stop. The risk isn’t that AI gets these wrong more often than a tired associate at 11 PM — it’s that AI errors don’t come with an obvious tell. A model that misclassifies a document does so with the same confident formatting as one it gets right, which means review protocols need to be designed assuming the output needs checking, not assuming it’s generally trustworthy and spot-checked occasionally. This is less about AI being unreliable and more about matching the review rigor to what’s actually at stake if it’s wrong.
There’s also a documentation benefit worth calling out separately from the review itself: a defensible AI-assisted process needs a clear audit trail showing what the model flagged, what a human changed, and why. Firms that treat this as a compliance afterthought tend to regret it the first time a production decision gets challenged, because “the tool said so” is not an answer that holds up, while “here’s the validation sample and the override log” is. Building that logging in from the start costs very little compared to reconstructing it after the fact under deadline pressure.
Legal Research Automation Is a Different Problem Than E-Discovery
Legal research gets lumped in with e-discovery in a lot of vendor pitches, but it’s a different risk profile. E-discovery works with a closed, known document set — the risk is misclassification within bounded data. Legal research with a general-purpose LLM works against the model’s training data and whatever it can retrieve, and the well-documented failure mode is citation fabrication: a model generating a case name and holding that sounds plausible and doesn’t exist, or exists but doesn’t say what the model claims. This has already produced sanctioned filings, which should tell you everything about how this gets used safely.
The safer architecture is retrieval-based rather than generative-from-memory: a system that searches a real, current legal database and summarizes what it actually finds, with citations a human can click through and verify, rather than one that generates an answer from parametric memory and hopes it’s right. This is the same underlying pattern behind well-built retrieval-augmented generation systems in other domains — the model’s job is to find and synthesize real source material, not to recall facts from training. Any legal research tool that can’t show you the exact source passage behind a claim is one to be skeptical of, regardless of how fluent its output reads.
The Compliance and Privilege Risk Nobody Should Skip
Before any of this touches real client matters, there’s a data governance question that has nothing to do with AI capability and everything to do with where documents end up. Confidential and privileged material processed through a third-party AI tool needs clear answers on data retention, training-data usage, and jurisdiction, in writing, before a single document is uploaded. This is the same due diligence firms already apply to cloud document management and e-discovery vendors; AI tools don’t get a pass on it just because the underwriting conversation feels newer. Ethical obligations around competence and confidentiality attach to the lawyer regardless of which vendor’s model is doing the first pass.
Integration Reality: Most Firms Aren’t Starting From Zero
Few firms are building e-discovery or research tooling from scratch. Most already run a document management system, a review platform, and some combination of legal research subscriptions, and the realistic project is connecting AI capability into that existing stack rather than replacing it — automating the handoff between systems, standardizing how chronologies and privilege logs get produced, and reducing the manual re-entry that happens when review platform output has to be reformatted for court filings. That kind of legal workflow integration work tends to deliver more practical value per dollar than a wholesale platform swap, mostly because it doesn’t ask reviewers to abandon tools they already trust.
Wo Sie anfangen sollten
A reasonable first step is narrow by design: pick one matter type with predictable document volume, apply AI-assisted triage and chronology building within the existing TAR validation framework, and measure the actual review-hour reduction against a traditional pass before expanding further. That gives the team a real, defensible data point instead of a vendor’s benchmark, and it surfaces where the firm’s own review protocols need tightening before the stakes get higher on a bigger matter.
If your firm is weighing where AI genuinely reduces review burden versus where it just adds another tool to manage, start a project conversation with us, or get in touch to talk through your current review stack and where a scoped pilot would actually move the needle.
Verwandt