E-discovery, knowledge management, and lateral analytics at scale

A single litigation matter can drop a million documents on your desk. Emails, contracts, spreadsheets, Slack exports, scanned PDFs. Reviewing all of it by hand at, say, 50 documents an hour would take one attorney roughly 20,000 hours. That is ten years of full-time work for one case.

This is why data tooling stopped being optional in law firms. The same corpus that must be reviewed for a production can also be mined to build a precedent bank and to score a lateral partner's book of business. Same data, three very different jobs. This lesson walks through all three.

The three jobs, one corpus

Picture the document set from a large commercial dispute. Here is what a firm does with it:

1. E-discovery: Find the responsive, non-privileged documents you must hand over. This is defensive and deadline-driven.
2. Knowledge management (KM): Extract reusable work product (clause language, briefs, expert reports) so the firm does not reinvent it next time.
3. Lateral analytics: Use client and matter data to evaluate whether a partner joining from another firm brings real, portable revenue.

Each job has different rules about who can see what. That constraint, not the technology, is usually the hard part.

Job 1: Document review at scale with TAR

E-discovery is the process of identifying electronically stored information (ESI) relevant to a legal matter. In US litigation, parties must produce documents responsive to the other side's requests, minus anything privileged (protected attorney-client communications and work product).

TAR (Technology Assisted Review), sometimes called predictive coding, uses machine learning to prioritize or classify documents so humans do not read all of them. It has been accepted by US courts since the 2012 decision in *Da Silva Moore v. Publicis Groupe*.

Here is the workflow in plain terms:

• A senior attorney reviews a seed set (a sample of documents) and codes each as responsive or not.
• The model learns from those decisions.
• It scores every remaining document by likelihood of being responsive.
• Reviewers focus on the high-scoring pile and validate the low-scoring pile with statistical sampling.

The two flavors you will hear about:

• TAR 1.0: Train the model once on a fixed seed set, then apply it. Simpler, but the seed set has to be representative.
• TAR 2.0 (Continuous Active Learning, or CAL): The model keeps retraining as reviewers work, constantly resurfacing the most likely responsive documents. Most modern platforms use this.

The payoff is measured with two metrics borrowed from information retrieval:

• Recall: Of all the truly responsive documents, what fraction did you find? Courts care about this most. Missing responsive documents is a legal risk.
• Precision: Of the documents you flagged, what fraction were actually responsive? This drives cost. Low precision means humans waste time on junk.

A defensible review often targets recall around 75 to 80 percent, though the right number is negotiated and matter-specific. For a plain-language primer, see The Sedona Conference publications, which courts frequently cite on ESI standards.

🎬 [VIDEO: "What is Technology Assisted Review (TAR)?" — youtube.com — a short explainer on predictive coding in e-discovery]

Why review still needs humans

TAR does not decide privilege well. A model can flag that a document mentions a lawyer, but judging whether a communication is genuinely privileged requires legal judgment. Firms use TAR to shrink the pile, then run targeted human review and privilege logging on what remains.

Job 2: Knowledge management, turning matters into reusable assets

The same million documents contain gold that has nothing to do with this case. Every brief your firm has ever filed, every negotiated indemnification clause, every deposition outline is a reusable asset. Knowledge management is the practice of capturing and retrieving that work product so lawyers stop drafting from scratch.

The modern version leans on retrieval-augmented generation (RAG): a system that searches your firm's own documents and feeds the relevant passages to a large language modellarge language modelEin Large Language Model ist ein KI-System, das auf riesigen Textmengen trainiert wurde, um Sprache vorherzusagen und zu erzeugen. Damit sind Aufgaben wie Schreiben, Zusammenfassen und Beantworten von Fragen möglich.Vollständige Definition ansehen → to draft or answer. Crucially, the model answers from your vetted corpus, not from the open internet, which reduces the risk of fabricated citations (a real and sanctioned problem when lawyers have used consumer chatbots).

A simplified retrieval flow looks like this:

# Pseudocode: retrieve firm precedents relevant to a drafting task
query = "limitation of liability clause, SaaS, mutual cap"

# 1. Convert query and documents to vectors (embeddings)
q_vec = embed(query)

# 2. Search only documents this user is cleared to see
results = vector_db.search(
    q_vec,
    top_k=5,
    filter={"ethical_wall": {"$not_in": user.blocked_matters}}
)

# 3. Feed retrieved clauses to the model as grounding context
draft = llm.generate(prompt=task, context=results)

Note line 2: ethical_wall. An ethical wall (or information barrier) is a mandatory internal restriction that blocks certain lawyers from accessing certain matters, usually to prevent conflicts of interest. Your KM system must respect these at the retrieval layer. A precedent bank that surfaces a walled-off client's confidential settlement terms to the wrong partner is a serious ethics breach.

The KM payoff

Concrete wins from a working precedent bank:

• A junior associate finds three prior versions of a data processing addendum in minutes instead of emailing partners.
• Pitch teams pull relevant matter descriptions without exposing client identities that are confidential.
• The firm identifies its own strongest arguments across past cases before a hearing.

The barrier is rarely technical. It is getting lawyers to tag and contribute work product when they are billing by the hour and have no time to file things.

Job 3: Lateral analytics and pipelinepipelineAlle aktiven Verkaufschancen über die Phasen des Vertriebsprozesses hinweg, zusammen mit ihrem gesamten potenziellen Wert und ihrer Abschlusswahrscheinlichkeit.Vollständige Definition ansehen → scoring

Law firms grow largely by hiring lateral partners (partners who move from another firm), betting that they bring their clients with them. The bet is often wrong. A partner's stated book of business (the revenue attributed to their client relationships) may not be portable.

Data helps pressure-test the bet. Firms analyze:

• Origination vs. working attorney credit: Did the partner originate the client relationship, or just work the matters? Only originated relationships tend to move.
• Revenue concentration: Is 80 percent of the book from one client? That is fragile.
• Client overlap and conflicts: Would the incoming clients conflict with existing ones, forcing the firm to decline work?
• Realization rate: The percentage of billed fees actually collected. A big book with poor realization is worth less than it looks.

A simple pipelinepipelineAlle aktiven Verkaufschancen über die Phasen des Vertriebsprozesses hinweg, zusammen mit ihrem gesamten potenziellen Wert und ihrer Abschlusswahrscheinlichkeit.Vollständige Definition ansehen → score might weight portable, diversified, high-realization revenue and discount concentrated or non-originated revenue. The point is to replace the partner's self-reported number with an evidence-based range.

The confidentiality trap

Here is where it gets delicate. A candidate partner cannot hand you their current firm's confidential client data during recruiting. Much of the analysis relies on the partner's representations plus public sources (court dockets, deal databases, reported matters). The firm builds an estimate, not a verified ledger, and documents its assumptions.

The constraint that ties it all together: client-mandated reporting

Large clients increasingly dictate how their data is handled through outside counsel guidelines (OCG): contractual rules a client imposes on its law firms. These commonly require:

• Specific data retention and destruction timelines for matter documents.
• Restrictions on using the client's data to train AI models.
• Reporting on matter budgets, staffing, and diversity in a mandated format.
• Segregation of the client's data from other clients.

This reshapes every job above. Your TAR platform must be able to purge a client's corpus on schedule. Your KM system must exclude clients who prohibit reuse of their documents, even anonymized. Your lateral analytics cannot quietly ingest a client's confidential pricing.

Good practice is to encode OCG rules as machine-readable policies (retention flags, no-train flags, reuse-permitted flags) attached to each matter, so the constraints are enforced automatically rather than remembered by an overworked associate.

Bringing it together

The strategic insight: the marginal cost of a second use of legal data is low, but the marginal legal and ethical risk can be high. A firm that mines its corpus aggressively without policy controls will eventually breach a confidentiality obligation. A firm that locks everything down captures none of the KM or analytics upside. The winners build the plumbing (access controls, retention policies, ethical walls) once, then reuse data safely many times.

Key Takeaways

• One corpus, three jobs: the same documents drive e-discovery, knowledge management, and lateral analytics, but each has different access rules.
• TAR shrinks review, humans still judge: measure success with recall (find responsive docs) and precision (avoid wasted review), and keep privilege calls with lawyers.
• KM value depends on enforced access controls: RAG on your own precedents only works if ethical walls and reuse permissions are enforced at retrieval time.
• Lateral scoring is an estimate, not a ledger: weight portable, diversified, high-realization revenue, and never ingest a candidate's confidential client data.
• Encode client rules as policy, not memory: outside counsel guidelines on retention, no-train, and segregation should be machine-enforced across every workflow.