Gains:
- Ability to understand the EDRM workflow and TAR/predictive coding concept and accelerate the review phase with artificial intelligence
- Ability to understand the balance of recall and precision and ensure not to miss relevant documents by prioritizing high recall in a forensic context
- Understanding that legal responsibility for privilege and personal data decisions rests with the human lawyer and that methodology transparency is required
In a civil lawsuit or regulatory investigation, one party requests electronic documents held by the other: emails, contracts, chat logs, files. The process of collecting these documents in accordance with the law, examining them, sorting out the relevant ones and delivering them to the other party/court is called e-discovery - finding, reviewing and presenting electronic evidence within the scope of a legal dispute. E-discovery is the area where computer forensics is most intertwined with law and is perhaps the most mature and widespread use of AI. In this unit, you will learn the EDRM workflow, the concept of TAR/predictive coding and how AI transforms this process.
EDRM: standard workflow of e-discovery
E-discovery in the world largely operates within the framework of EDRM (Electronic Discovery Reference Model - the reference model that defines the standard stages of the e-discovery process). Main stages:
- Information management and identification: Determining what data is where.
- Preservation & collection: Preventing the deletion of relevant data (legal hold — the obligation to stop deleting data in anticipation of litigation) and collecting it with integrity.
- Processing: Making data searchable, removing duplicates, format conversion.
- Review: Deciding whether documents are sensitive or privileged — covered by attorney-client privilege and will not be handed over.
- Production: Delivering the relevant documents to the other party/court.
The most expensive and manpower-requiring stage is the review. Having lawyers read hundreds of thousands of documents one by one is both very expensive and very slow. This is exactly where AI creates a revolution.
TAR and predictive coding
The key technology that speeds up review is TAR (Technology Assisted Review). The most common form is predictive coding: An expert attorney labels sample documents as "relevant/irrelevant"; The system learns from these examples and automatically classifies the remaining hundreds of thousands of documents. Thus one reads only the most likely relevant documents and checks the rest by sampling.
Modern AI (large language models) takes this a step further: it can explain why the document is relevant, topic summarize, privilege flag, not just "relevant/irrelevant". But the basic principle does not change: the AI suggests, the legal responsibility for the decision of relevance and privilege lies with the lawyer.
The defensibility of TAR depends on the process being auditable. The training examples labeled by the expert lawyer, the system learning from these examples and classifying the remaining documents are repeated in a cycle; Accuracy is measured by sampling in each round. This cycle provides documentable answers to the questions "what did the machine look at, what did it learn, how accurate was it?" The other side often questions this methodology; Therefore, training set size, decision thresholds and validation sample rates are recorded from the beginning. The more transparent and repeatable the process, the more likely the outcome will be accepted in court.
Tip: To make your TAR process defensible, document your “training” samples, validation metrics (recall/precision), and decision thresholds. The other party asks "what did the machine miss?" he may ask; Transparent methodology ensures trust.
Two metrics: recall and precision
Two concepts are vital in e-discovery. Recall (sensitivity — how much of all relevant documents were you actually able to find) measures the risk of missing; It is legally critical because failure to submit a relevant document may result in sanctions. Precision—how much of what you find is actually relevant—measures unnecessary reading load; affects the cost. There is a balance between the two: if you relax the threshold, recall increases, but unnecessary documents also increase. In a forensic/legal context, high recall is often prioritized — not missing is more important than over-reading.
Privilege and privacy: the red line
The most dangerous mistake in e-discovery is accidentally handing over a privileged document to the other party (privilege waiver). Once attorney-client correspondence is delivered, it is very difficult to retrieve it. AI helps flag privilege, but every privilege decision must be confirmed by a human attorney. Additionally, redaction may be required for documents containing personal data (KVKK/GDPR scope); AI flags masking candidates, final control is with the human.
Caution: Observe data privacy and jurisdiction (which country's data, where does it go) rules when loading raw corporate data into AI. Privileged and personal data must be processed on vehicles whose data does not go to model training, under contract and in the appropriate jurisdiction.
three mini cases
Case 1 — TAR reduced cost. In a commercial case of 1.3 million documents, the classic linear review estimate was 14 months. With predictive coding, lawyers tagged 8,000 documents; the system classified the remainder and human review was down to 90,000 documents. The process was reduced to 6 weeks, 92% confirmed by recall sampling.
Case 2 — Privilege saved. Pre-production AI flagged 240 documents in the delivery set as “possibly privileged.” The lawyer team examined; 210 were actually attorney-client correspondence and were not delivered. Without AI, these documents could have accidentally gone to the other party.
Case 3 — The risk of overconfidence. One team set the TAR threshold tightly for high precision; The system eliminated those it deemed "irrelevant" without sampling. Subsequent audit found that several critical emails were missed due to the strict threshold. The error was noticed late because call sampling was not performed.
Four copyable templates
1) Draft review protocol:
Your role: e-discovery consultant. Case subject: [breach of contract]. Draft a review protocol for me: responsive criteria, privilege flags, personal data masking rules and TAR verification steps (recall/precision sampling). State that every decision is subject to attorney confirmation.
2) Relevance preliminary classification:
Subject matter and relevance criteria: [here]. I will give you document summaries. For each document: relevant/irrelevant/uncertain and give a BRIEF justification. Highlight the "unclear" ones; these will go to attorney review. Those you call irrelevant will be checked by sampling, it is not a definitive elimination.
3) Privilege scanning:
Tick the following documents that POSSIBLY pose attorney-client privilege: attorney names, "confidential/privileged" language, legal opinion content. These are not FINAL decisions, they are candidates to be confirmed by a lawyer. Give justification and the relevant line for each sign.
4) Candidates for personal data masking:
List the personal data candidates that need to be masked in these documents to be delivered before production: TR, IBAN, health, third party private data. Mark as candidate only; I will make the final masking decision. Change text.
Weak prompt / Strong prompt
Weak prompt:
From these documents, select those that are relevant to the case.
No relevance criteria, no exclusivity and no verification; AI makes subjective decisions, may miss critical document or leak privileged.
Powerful prompt:
Your role: e-discovery review assistant. Case: Allegation that company X violated the supply contract with Y. Relevance criteria: delivery dates, price changes, quality complaints, termination negotiations. I will give you document summaries. Give relevant/irrelevant/uncertain + justification for each document. ALSO mark as "possibly privileged" those containing lawyer's name or legal opinion. Do not consider any decision as final; All are subject to attorney confirmation.
The substantive criteria, privilege screening, and “lawyer verification” constraint make the process defensible.
EDRM stages and AI table
Stage
AI contribution
human responsibility
collection
template, plan
Legal hold, integrity
Processing
De-dup, format, directory
quality control
review
TAR, relevance recommendation
Relevance decision
privilege
Marking candidate
final decision, confirmation
Production
masking candidate
Final check, delivery
Common mistakes
- Leaving TAR unattended. Saying "the machine said it is not interested" without performing a call sample will cause hijacking.
- Leaving the privilege decision to the AI. A misdelivery risks irrevocable waiver.
- Focusing only on precision instead of recall. Tight threshold misses the relevant document; In law, recall has priority.
- Not documenting the methodology. The other party may question the TAR process; transparency is a must.
- Uploading personal/privileged data to an inappropriate tool. Observe jurisdiction and confidentiality rules.
In summary
E-discovery is the process of finding and delivering relevant evidence from corporate electronic data in a legal manner and operates within the framework of EDRM. AI transforms the most expensive phase, reviewing, with TAR/predictive coding: the human reads only the most likely relevant and ambiguous documents. But the legal responsibility for the decision of relevance and especially privilege lies with the lawyer; Recall sampling, methodology transparency and personal data protection are ensured by human discipline.
Application task
Define a fictitious case and 4-5 items of relevance criteria. Prepare 20-25 document summaries (a few privileged, a few obscure). Apply the “Relevance pre-classification” and “Privilege screening” templates; Then manually check the recall by sampling from the "irrelevant" set and confirm the privilege candidates the AI has flagged.
checklist
- [ ] I observed the EDRM stages and the legal hold obligation.
- [ ] I checked the TAR/pre-classification result with recall sampling.
- [ ] I have tied every privilege decision to human attorney confirmation.
- [ ] I have approved the personal data masking candidates with the final check.
- [ ] I documented the methodology and thresholds I used.