Legal Tech11 min read

AI E-Discovery in Singapore: A 5-Step Claude Workflow

Run document review on thousands of files using Claude. Five-step workflow with PDPA redaction, privilege flagging, and audit trail. For Singapore litigators and disputes teams.

Haojun See

Founder & Director, On The Ground

Updated 1 May 2026

The five steps

Step 1 — Define the issues clearly. Before any review, write a one-page issue list: what facts are in dispute, what claims/defences depend on those facts, what kinds of documents are likely relevant. Claude needs this as a system prompt. Step 2 — First-pass classification. Pass each document through Claude with a structured classification prompt. Output: relevant (yes/no), issue tags, privilege (yes/no/borderline), urgency (1–5). Step 3 — Sample-test the classifications. Manually review a random 5–10% sample. Calculate Claude's precision and recall on relevance and privilege. If accuracy is below threshold (typically 90% for relevance, 99% for privilege), refine the prompt and re-run. Step 4 — Second-pass for the relevant subset. For documents Claude marked relevant, run a deeper analysis prompt extracting key facts, dates, and quotations. Step 5 — Senior verification of privileged and high-stakes documents. Human partner-level review of every document Claude flagged as privileged or borderline. Document the methodology in your matter file as required by the MinLaw GenAI Guide.

The classification prompt

Use this as the workhorse prompt for step 2: *"You are reviewing a document in [matter name]. The issues are: [paste one-page issue list]. For this document, return a structured JSON with: - relevant: yes / no - issue_tags: array from the issue list - privilege_flag: privileged / not privileged / borderline - privilege_basis: if flagged, the basis (lawyer-client, litigation, common interest, none) - key_facts: bullets, max 5 - key_dates: array of dates referenced - urgency: 1–5 - notes: anything unusual Document follows. [paste document text]"* Run programmatically across your document set. Save outputs to a CSV or database. The full pipeline can be wired up in Claude Code in a day — see our Internal Tools guide.

Privilege flagging — the critical step

Privilege errors create real ethical and litigation exposure. Two rules. Rule 1: false negatives are the failure mode that matters. A false positive (Claude says privileged when it isn't) just means a document gets a second review. A false negative (Claude says not privileged when it is) means a privileged document gets produced — that's a waiver. Rule 2: always do a 100% human review of borderline and a sampled review of "not privileged" documents. Set Claude's prompt to over-call privilege ("when in doubt, mark as borderline"). The cost is more human review on the borderline pile; the benefit is no privileged documents slip through. The MinLaw guide is explicit: human accountability sits with the lawyer. AI assists, the lawyer decides. Document the methodology — sample size, accuracy metrics, second-pass coverage — in the matter file.

Sample-test methodology

After step 2, before relying on Claude's classifications, sample-test: 1. Random sample 5–10% of documents Claude marked relevant. Manually review. Count precision (how many were actually relevant). 2. Random sample 5–10% of documents Claude marked not relevant. Manually review. Count recall (how many should have been marked relevant). 3. Calculate F1 score. Target: ≥0.85 for relevance, ≥0.95 for privilege. 4. If below threshold, refine the prompt (add examples of what's borderline, sharpen the issue list) and re-run on the affected subset. This is standard TAR methodology applied to LLM-assisted review. Defensible if challenged.

When to use Claude vs a managed-review platform

Use Claude-based workflow when: matter is under 5,000 documents, you control the document set, the issues are clearly defined, and budget matters. Use a managed-review platform (Relativity, Reveal, Everlaw) when: matter is over 10,000 documents, multiple parties need access, the matter is multi-jurisdiction with chain-of-custody requirements, or the court has specified a platform. Hybrid approach: load documents into Relativity for chain-of-custody, then use Relativity's Claude integration (or pull a subset for Claude-based first-pass) and push results back. Best of both — defensibility plus speed. For prompt patterns useful across all of these scenarios, see Claude for Legal Research and The Singapore Prompt Library.

What this looks like with OTG

OTG has built Claude-assisted document review pipelines for Singapore litigation and disputes teams. Typical engagement: - 1-day workshop covering the methodology and tools - Custom pipeline build (Functional App Sprint or Process Automation tier) - Senior associate trained on prompt iteration and sample-testing - Documentation suitable for inclusion in matter files For sensitive matters, deployment via OTG Legal Box keeps everything on-firm hardware. Book a free 30-minute call to discuss your matter type. For PDPA framework and broader guidance, see PDPA-Safe Claude Prompts for Lawyers and PDPA Prompting Checklist.

Frequently asked questions

Can Claude replace a document-review platform like Relativity?

Not entirely. For Singapore matters under 5,000 documents and where you control the document set, a Claude-based workflow can replace 80% of the platform's value at a fraction of the cost. For large-scale, multi-jurisdiction, or court-mandated review platforms, integrate Claude into Relativity / Reveal rather than replacing them.

Is it admissible? Will the court accept Claude-assisted review?

Singapore courts have not specifically blessed AI-assisted review, but TAR (technology-assisted review) is well-established and AI-assisted review falls within the same principles. The key is documenting your methodology — sample sets, accuracy metrics, second-pass verification — so that the process is defensible. See [MinLaw GenAI Guide](https://www.mlaw.gov.sg/files/Guide_for_Using_Generative_AI_in_the_Legal_Sector.pdf).

What about privileged documents?

Privilege flagging is the highest-stakes part of any review. Always have a human verify Claude's privilege classifications. Set the false-negative rate target very low (every false negative is a privilege waiver risk). Sample 20–30% manually.

How much does it cost vs. traditional review?

Manual review of 10,000 documents at S$300/hour averages S$30,000–S$80,000. A Claude-based workflow with senior verification typically lands at 20–35% of that for first-pass classification, with quality matching for relevance and superior consistency for issue tagging.

Is this PDPA-safe?

Yes when run on enterprise-tier Claude with no training on input + redaction of any third-party personal data not relevant to the matter. For criminal-defence-grade confidentiality, run on-device via [OTG Legal Box](/apps/legal-box). Full framework at [PDPA-Safe Claude Prompts for Lawyers](/resources/pdpa-safe-claude-prompts-lawyers).

PreviousClaude for Legal Research: 12 Prompts That Beat Manual Search Next Claude for Singapore SMEs: 10 Prompts That Save 10 Hours a Week

Want to Apply This to Your Business?

We're a Singapore AI development and automation agency. Let's discuss how we can help solve your specific challenges.

Get in Touch View our services