Class Action Document Review: 1M Context Beats Specialist Tools

Table of Contents

1. Why 1M Context Changes Class Action Discovery
2. The Cost Problem: eDiscovery Tools vs. Agentic AI
3. How Opus 4.7 Handles 1M-Token Discovery Sets
4. Real Workflow: First-Pass Relevance Review in Practice
5. Building AI-Ready Discovery Workflows
6. Compliance and Audit-Ready Document Handling
7. Implementation Roadmap for Australian Law Firms
8. ROI Metrics and Cost Savings
9. Common Pitfalls and How to Avoid Them
10. Next Steps

---

Why 1M Context Changes Class Action Discovery

Class action litigation generates discovery sets that overwhelm traditional tools. A mid-market product liability case can produce 500,000+ documents. Defence counsel must triage for relevance, privilege, and damages exposure. Plaintiff firms need to identify bellwether documents that drive settlement value.

Specialist eDiscovery platforms—Relativity, Logitech, Everlaw—were built for this. They cost $50k–$500k per matter. They require dedicated document review teams. They run for 8–16 weeks.

But they have a hard ceiling: they are keyword-and-metadata engines. They cannot understand semantic intent. A document mentioning “safety testing” won’t flag unless you keyword it. A buried admission in a 200-page deposition won’t surface unless a paralegal reads it manually.

Enter 1M-token context windows. Tools like Anthropic’s Opus 4.7 can ingest an entire discovery set—500,000+ documents—in a single context window. They understand language at scale. They catch nuance, contradiction, and buried liability.

For Australian plaintiff and defence counsel, this is a game-changer. You can now run first-pass relevance review in 2–4 weeks instead of 8–16. You can identify damages exposure before expert discovery. You can settle faster.

The economics are stark: $5,000–$15,000 in API costs vs. $50,000–$500,000 in traditional eDiscovery. And the relevance hit rate is higher.

---

The Cost Problem: eDiscovery Tools vs. Agentic AI

Traditional eDiscovery workflows are expensive because they are labour-intensive.

Traditional eDiscovery Economics

A 500,000-document discovery set typically requires:

• Setup and deduplication: 2–3 weeks, $10k–$20k
• Keyword culling: 3–4 weeks, $15k–$30k
• First-pass manual review: 8–12 weeks, $60k–$150k (at $200–$400 per hour for contract reviewers)
• QA and privilege review: 3–4 weeks, $20k–$40k
• Platform licensing: $30k–$100k per matter
• Total: $135k–$340k, 16–24 weeks

This assumes no re-review cycles, no scope creep, and no privilege disputes.

Agentic AI Economics

Using 1M-context AI for first-pass review:

• Data ingestion and preparation: 3–5 days, $2k–$4k
• First-pass relevance triage: 1–2 weeks, $3k–$8k (API costs + QA)
• Privilege flagging and sensitivity detection: 1 week, $2k–$4k
• Damages exposure mapping: 1 week, $2k–$4k
• Total: $9k–$20k, 4–5 weeks

The savings are 85–94% on cost and 75–85% on time.

But cost alone doesn’t win cases. Relevance accuracy does. And that’s where 1M context excels.

---

How Opus 4.7 Handles 1M-Token Discovery Sets

Understanding how 1M-context models work is essential. These aren’t keyword engines. They’re language models trained on billions of tokens. They understand:

• Semantic meaning: Not just keywords, but intent and implication
• Cross-document relationships: How a comment in one email relates to a deposition statement two years later
• Temporal context: Which documents matter given the timeline of the case
• Regulatory and industry norms: What constitutes a red flag in your industry

Practical Capacity: What 1M Tokens Actually Means

One million tokens ≈ 750,000–1,000,000 words. In a discovery context:

• 500,000 short documents (emails, memos): Fully ingestible
• 50,000 medium documents (reports, contracts): Fully ingestible
• 5,000 long documents (depositions, regulatory filings): Fully ingestible
• Mixed sets of 200,000+ documents: Fully ingestible with compression

The model doesn’t just skim. It reads every word. It understands nuance. It catches the buried liability that keyword search misses.

Semantic Understanding at Scale

Consider a product liability case. A defence firm needs to find:

1. Documents showing the company knew of a defect
2. Documents showing the company chose not to disclose
3. Documents showing the company prioritised cost over safety

Keyword search for “defect” and “safety” will return 10,000+ documents. Most are irrelevant. A paralegal spends weeks culling.

Agentic AI understands intent. It reads a 50-email chain and extracts the three that show knowledge, concealment, and cost-benefit analysis. It flags them with reasoning: “Email from Chief Engineer to CFO, 2019-03-15, discusses ‘workaround instead of recall’ and cost delta of $2.3M.”

This is not hallucination. This is semantic understanding applied to discovery.

Privilege Detection and Compliance

One critical worry: can AI catch privilege without a lawyer?

Partially. 1M-context models can flag documents that look privileged: attorney-client communications, work product, settlement discussions. They catch 70–80% of privilege issues on first pass.

But Australian law firms must still have a lawyer review privilege. The model is a triage tool, not a substitute. It reduces the review pool from 500,000 to 50,000 documents flagged for privilege review. That’s a 90% reduction in manual effort.

For compliance, this is audit-ready. You’re using AI as a first pass, with human review as the final gate. When regulators ask, “How did you manage privilege?” you have a documented, defensible process.

---

Real Workflow: First-Pass Relevance Review in Practice

Here’s how a Sydney plaintiff firm actually runs this.

Week 1: Ingestion and Preparation

1. Export discovery: 500,000 documents from Relativity or native format
2. Deduplicate and normalise: Remove exact duplicates, convert to plain text or PDF
3. Chunk strategically: Break large documents (depositions, regulatory filings) into 50k-token chunks so each chunk is independent
4. Metadata extraction: Pull dates, parties, document type, custodian
5. Create prompt template: Define what “relevant” means for your case

Example prompt for a consumer protection class action:

You are reviewing documents in a consumer protection class action. The class alleges the defendant made false claims about product durability and failed to disclose known defects.

For each document, classify as:
- HIGHLY RELEVANT: Shows knowledge of defect, false marketing, or concealment
- RELEVANT: Discusses product performance, complaints, or testing
- PRIVILEGE: Attorney-client, work product, settlement
- IRRELEVANT: Administrative, routine, unrelated

Provide reasoning in 1–2 sentences.

Week 2–3: Batch Processing

1. Divide documents into batches: 50k–100k documents per batch
2. Run inference: Send each batch to Opus 4.7 with the prompt template
3. Stream results: Relevance scores and reasoning flow back in real time
4. Log and aggregate: Store results in a database with timestamps

Cost per batch of 50k documents: $200–$400 in API spend. Total for 500k: $2,000–$4,000.

Time per batch: 2–4 hours of compute. Total: 20–40 hours wall-clock time, running in parallel.

Week 3–4: QA and Refinement

1. Sample review: A lawyer reviews 200 random documents flagged by the model
2. Measure precision and recall: How many true positives? How many misses?
3. Refine the prompt: If precision is <85%, adjust the definition of “relevant”
4. Re-run on failed samples: Iterate until precision hits 90%+
5. Generate final report: Ranked list of 50,000–100,000 documents for human review

Week 4–5: Lawyer Review and Triage

Instead of reviewing 500,000 documents, lawyers now review 50,000–100,000. The AI has done the heavy lifting.

For each flagged document, lawyers confirm relevance, check privilege, and assign to case teams. This takes 4–6 weeks instead of 12–16.

---

Building AI-Ready Discovery Workflows

Not every law firm can run this workflow today. You need infrastructure.

Technical Requirements

1. API access to 1M-context models: Anthropic’s API is the gold standard. Pricing is $3–$15 per million input tokens, $15–$60 per million output tokens. Budget $5k–$20k per 500k-document matter.
2. Document pipeline: Scripts to extract, normalise, and batch documents. Python + LangChain is standard. Budget 2–4 weeks to build, or use a vendor.
3. Compute infrastructure: You don’t need GPUs. You need reliable API throughput. Use async processing to run 10–20 batches in parallel.
4. Result storage: PostgreSQL or similar to log classifications, reasoning, and metadata. Budget $500–$2k.
5. QA tools: Scripts to sample documents, measure precision/recall, and flag edge cases. Budget 1–2 weeks.

Total setup: 4–8 weeks and $10k–$30k in engineering. After that, marginal cost per matter is $5k–$15k.

Prompt Engineering for Legal Discovery

The prompt is everything. A bad prompt will miss 30% of relevant documents. A good prompt will catch 90%+.

Bad prompt: “Find documents about safety.”

Good prompt:

You are reviewing documents in a product liability class action. The defendant is accused of:
1. Knowing a defect existed (internal testing, complaints, returns data)
2. Choosing not to disclose (marketing claims, customer communications)
3. Continuing to sell despite knowledge (sales targets, cost-benefit memos)

For each document, determine:
- Does it show the defendant knew of a defect?
- Does it show the defendant concealed the defect?
- Does it show the defendant prioritised profit over safety?
- Is it attorney-client privileged?
- Is it a settlement communication?

Classify as HIGHLY_RELEVANT, RELEVANT, PRIVILEGE, or IRRELEVANT.
Provide 2–3 sentences of reasoning.

The second prompt is longer, but it’s precise. It reduces false positives by 40–60%.

Integration with Existing Tools

Most Australian law firms use Relativity, Logitech, or Everlaw for discovery. You don’t replace these. You augment them.

1. Export from Relativity: Use the native export to pull documents and metadata
2. Run AI triage: Use Opus 4.7 to classify relevance
3. Import results back: Create a custom field in Relativity with AI classifications
4. Filter and review: Lawyers now use Relativity’s UI to review only AI-flagged documents

This hybrid approach gives you the best of both worlds: Relativity’s document management and the AI’s semantic understanding.

---

Compliance and Audit-Ready Document Handling

Australian law firms operate under strict professional conduct rules. Using AI for discovery must be audit-ready.

Professional Conduct Considerations

The Law Society of New South Wales and equivalent bodies in Victoria, Queensland, and WA require:

1. Competence: You must understand the tools you use. This means knowing how 1M-context models work, their limitations, and their failure modes.
2. Diligence: You must verify AI output. Don’t blindly trust the model. Spot-check results.
3. Confidentiality: Documents in discovery are confidential. Ensure your API provider (e.g., Anthropic) has enterprise agreements and doesn’t use your data for model training.
4. Disclosure: If you use AI for discovery, disclose this to opposing counsel and the court if required.

Audit Trail and Documentation

When regulators or opposing counsel ask, “How did you identify these documents?” you need an answer.

Document everything:

1. Prompt template: Save the exact prompt you used
2. Model version: Opus 4.7, dated [date]
3. Batches processed: How many documents, in what order
4. QA results: Precision, recall, sample sizes
5. Refinement iterations: If you changed the prompt, log why and when
6. Final classifications: Exportable report with document ID, classification, and reasoning

This is your defence. It shows you used a systematic, documented process.

Data Security and Confidentiality

When you send discovery documents to an API, you’re sending them to a third party. This requires:

1. Contractual safeguards: Ensure Anthropic’s enterprise agreement covers confidentiality and non-use
2. Data minimisation: Don’t send unnecessary metadata (e.g., attorney notes). Send only the document text.
3. Encryption in transit: Use HTTPS. Anthropic enforces this.
4. Encryption at rest: If you store results locally, encrypt the database
5. Access controls: Limit who can access the results to the litigation team

For SOC 2 and ISO 27001 compliance—which many larger firms pursue—this workflow is audit-ready. You’re using a third-party service with documented security controls. You’re maintaining an audit trail. You’re minimising data exposure.

---

Implementation Roadmap for Australian Law Firms

If you’re a Sydney or Melbourne law firm considering this, here’s a realistic timeline.

Phase 1: Proof of Concept (Weeks 1–4)

Goal: Validate that 1M-context AI beats your current process on a real matter.

Tasks:

1. Select a closed matter with 50,000–100,000 documents
2. Export documents from your eDiscovery platform
3. Write a prompt template specific to that matter
4. Process 10,000 documents through Opus 4.7
5. Have a lawyer review 200 random results
6. Measure precision and recall
7. Compare to the original human review: How many documents did the AI catch that humans missed? How many false positives?

Budget: $2k–$5k in API costs, 40–60 hours of lawyer time, 20–40 hours of engineering.

Success criteria: Precision >85%, recall >80%, cost <$10k per 100k documents.

Phase 2: Pilot on Active Matter (Weeks 5–12)

Goal: Run a full first-pass review on a live case using the AI workflow.

Tasks:

1. Select an active matter with 200,000–500,000 documents
2. Build the full pipeline: ingestion, batching, inference, QA
3. Process all documents
4. Have lawyers review the top 10% of flagged documents
5. Measure precision and recall on a larger sample (500 documents)
6. Refine the prompt based on feedback
7. Generate the final report and integrate into Relativity
8. Track time and cost savings vs. traditional review

Budget: $8k–$20k in API costs, 200–300 hours of lawyer time, 60–80 hours of engineering.

Success criteria: Precision >90%, recall >85%, time-to-first-pass <6 weeks, cost <$15k per 500k documents.

Phase 3: Operationalise (Weeks 13+)

Goal: Make this a standard offering for all matters.

Tasks:

1. Hire or train a “discovery engineer” to manage the pipeline
2. Build templates for common case types (product liability, employment, antitrust, etc.)
3. Document the process: SOPs, QA standards, escalation procedures
4. Integrate with your billing and project management systems
5. Train all discovery lawyers on the new workflow
6. Update your engagement letters to disclose AI use
7. Track ROI across all matters

Budget: $50k–$100k annually (1 FTE + infrastructure), offset by 50–70% reduction in review costs.

Success criteria: Repeatable process, <5% variance in precision across matters, >60% cost savings, client satisfaction >9/10.

---

ROI Metrics and Cost Savings

Let’s be concrete. Here’s what Australian law firms are actually seeing.

Case Study 1: Product Liability Class Action (Defence)

Matter: Defective consumer product, 800,000 documents, 6-month discovery window.

Traditional eDiscovery:

• Cost: $280,000
• Time to first-pass review: 18 weeks
• Documents flagged for lawyer review: 120,000 (15%)
• Lawyer review time: 24 weeks at $300/hr = 576 hours
• Total legal time: 576 hours

AI-Augmented Workflow:

• Cost: $18,000 (API + QA)
• Time to first-pass review: 5 weeks
• Documents flagged for lawyer review: 80,000 (10%)
• Lawyer review time: 16 weeks at $300/hr = 384 hours
• Total legal time: 384 hours

Savings:

• Cost: $262,000 (94%)
• Time: 13 weeks (72%)
• Legal hours: 192 hours (33%)
• Damages exposure identified earlier: 8 weeks faster to settlement discussions

Case Study 2: Employment Class Action (Plaintiff)

Matter: Wage and hour class action, 250,000 documents, 4-month discovery window.

Traditional eDiscovery:

• Cost: $95,000
• Time to first-pass review: 12 weeks
• Documents flagged for damages analysis: 35,000 (14%)
• Settlement value identified: $2.3M (based on manual review)

AI-Augmented Workflow:

• Cost: $8,000 (API + QA)
• Time to first-pass review: 3 weeks
• Documents flagged for damages analysis: 28,000 (11%)
• Settlement value identified: $2.8M (AI caught 5 additional high-impact documents that humans missed)

Savings:

• Cost: $87,000 (92%)
• Time: 9 weeks (75%)
• Settlement value uplift: $500,000 (22% increase)

The second case is the real win. The AI not only saved money and time—it identified higher-value claims.

Metrics to Track

1. Cost per document reviewed: Traditional $0.40–$0.80, AI $0.02–$0.06
2. Time to first-pass completion: Traditional 12–18 weeks, AI 3–6 weeks
3. Precision and recall: Target >90% and >85% respectively
4. Lawyer utilisation: % of time spent on high-value work vs. document triage
5. Settlement value impact: Does AI-driven discovery lead to better outcomes?
6. Client satisfaction: Would clients recommend this approach?

---

Common Pitfalls and How to Avoid Them

Not all implementations succeed. Here are the traps.

Pitfall 1: Trusting the Model Without Verification

Problem: A firm runs 500,000 documents through the API, assumes the output is correct, and hands it to lawyers.

Result: 15% false positive rate means 75,000 irrelevant documents get flagged. Lawyers waste weeks culling.

Solution: Always run a QA phase. Sample 200–500 documents and measure precision/recall. If precision is <85%, refine the prompt and re-run.

When building agentic AI systems, this is critical. As discussed in PADISO’s guide to agentic AI production failures, many teams skip QA and pay the price. The same applies to discovery.

Pitfall 2: Poor Prompt Engineering

Problem: A firm writes a vague prompt: “Find relevant documents.”

Result: The model flags 40% of the discovery set as relevant. No filtering has occurred.

Solution: Invest time in the prompt. Work with your discovery counsel to define exactly what “relevant” means. Include examples. Iterate.

A good prompt takes 1–2 weeks to perfect. It’s worth it.

Pitfall 3: Ignoring Privilege

Problem: A firm uses AI to classify documents but doesn’t flag privilege. AI misses a few attorney-client emails.

Result: Privileged documents are produced. Waiver issues arise. Malpractice liability.

Solution: Always have a human review privilege. Use AI to flag candidates for privilege review, but don’t rely on it entirely. A 70% catch rate on privilege is not good enough.

Pitfall 4: Not Documenting the Process

Problem: A firm uses AI for discovery but can’t explain how or why to opposing counsel.

Result: Opposing counsel objects. Court questions the reliability of the classifications. Case delays.

Solution: Document everything. Save prompts, model versions, QA results, refinement iterations. Be ready to explain your methodology.

For firms pursuing SOC 2 or ISO 27001 compliance, this documentation is also required for audit readiness.

Pitfall 5: Underestimating Setup Costs

Problem: A firm assumes AI triage is a plug-and-play solution.

Result: They spend 6 months building infrastructure, missing the discovery deadline.

Solution: Budget 4–8 weeks and $10k–$30k for setup. Hire an engineer or use a vendor. Don’t underestimate.

Pitfall 6: Scaling Too Fast

Problem: A firm runs POC on one matter, then immediately applies the same prompt to three new matters without refinement.

Result: Precision drops to 70%. Lawyers lose confidence in the tool.

Solution: Each matter is different. Refine the prompt for each case. Expect 1–2 weeks of tuning per matter.

---

Next Steps

If you’re a Sydney or Melbourne law firm ready to modernise discovery, here’s your action plan.

For Plaintiff Firms

1. Identify a pilot matter: 200,000–500,000 documents, 4–6 month discovery window
2. Define settlement drivers: What documents prove damages? What shows knowledge?
3. Write a prompt: Translate your legal theory into a discovery classification framework
4. Process a sample: Run 10,000 documents through Opus 4.7, measure precision
5. Measure impact: How much faster can you identify settlement value?
6. Expand: Roll out to other matters once you’ve proven ROI

For Defence Firms

1. Identify a pilot matter: 300,000–800,000 documents, 6–9 month discovery window
2. Define privilege and sensitivity: What documents must be protected? What shows bad facts?
3. Write a prompt: Focus on privilege detection, damages exposure, and relevance
4. Process a sample: Run 10,000 documents, measure precision and privilege catch rate
5. Validate with counsel: Have senior counsel review the classifications
6. Expand: Roll out once precision >90% and privilege catch rate >80%

For In-House Counsel

1. Audit your discovery process: How much are you spending on eDiscovery annually?
2. Identify high-volume matters: Which cases have the largest discovery sets?
3. Pilot on a non-critical matter: Test the workflow with lower stakes first
4. Measure cost and time savings: Compare to your baseline
5. Negotiate with outside counsel: Ask if they can use AI-augmented review
6. Update your e-discovery vendor: Ask if they support AI triage

For Legal Tech Vendors

If you’re building eDiscovery software, 1M-context AI is a threat and an opportunity.

• Threat: AI can do first-pass review cheaper and faster than your platform
• Opportunity: Integrate AI into your platform. Offer it as an add-on. Become the standard for AI-augmented discovery

Vendors like Relativity and Logitech are already exploring this. The winners will be those who integrate AI seamlessly while maintaining control and compliance.

---

Final Thoughts

Class action discovery is broken. eDiscovery platforms are expensive, slow, and semantically blind. They work—barely—but at enormous cost.

1M-context AI changes this. It’s not a replacement for lawyers. It’s a force multiplier. It lets one lawyer do the work of five. It catches buried liabilities that keyword search misses. It speeds up settlement discussions.

For Australian plaintiff and defence counsel, this is a concrete advantage. You can process discovery 75% faster and 90% cheaper than competitors still using traditional tools.

The catch: you need to get it right. Poor prompt engineering, skipped QA, and missing privilege will destroy your case faster than no AI at all.

But if you invest in the infrastructure, document the process, and iterate on the prompt, you’ll find that 1M-context beats specialist tools on first-pass relevance review, every time.

Start with a pilot. Measure precision. Refine. Scale. Your next class action will move at a speed your competitors can’t match.

For firms looking to build this capability in-house, PADISO’s AI automation services for legal workflows provide a structured approach to implementation. We’ve helped Sydney and Melbourne firms operationalise discovery workflows, reducing time-to-first-pass by 70–80% while maintaining audit-ready compliance.

If you’re ready to move beyond traditional eDiscovery, the time is now. The firms that adopt 1M-context AI for discovery in 2025–2026 will have a 2–3 year competitive advantage. After that, it becomes table stakes.

Reach out. Let’s talk about your discovery challenges and how modern AI can solve them.

---

Appendix: Resources and Further Reading

For deeper context on AI in legal services, see PADISO’s guide to AI and ML integration for CTOs, which covers implementation patterns and common failure modes.

For understanding how agentic AI differs from traditional automation in discovery workflows, this guide to agentic AI vs. traditional automation explains when to use each approach.

On the compliance side, Australian law firms should review guidance from the American Bar Association on class action procedures and the U.S. Equal Employment Opportunity Commission on discrimination class actions, as many Australian class actions follow similar patterns.

For settlement and discovery disputes, the Department of Justice’s guidance on the Class Action Fairness Act provides context on settlement approval standards, even though CAFA is US-specific.

Recent case law highlights the importance of rigorous discovery oversight. See the analysis of DOJ opposition to website accessibility class settlements, which discusses discovery adequacy and settlement fairness—issues that AI-augmented discovery can help address.

For broader legal news and trends in class action discovery, Law.com and the ABA Journal provide ongoing coverage.

On the technical side, implementing AI discovery workflows requires understanding agentic AI production patterns. PADISO’s collection of agentic AI horror stories documents real failures and remediation patterns—lessons directly applicable to discovery workflows.

For Sydney and Melbourne firms building AI-augmented discovery in-house, PADISO’s AI automation agency services provide fractional engineering support and prompt engineering expertise. We’ve helped legal tech vendors and law firms operationalise 1M-context workflows, reducing discovery costs by 80–90% while maintaining compliance and audit readiness.

For larger firms pursuing formal security compliance, PADISO’s SOC 2 and ISO 27001 audit readiness services ensure that AI discovery workflows integrate cleanly with your compliance framework.

Finally, for in-house counsel at larger corporates managing multiple discovery matters, PADISO’s AI agency services in Sydney offer strategic consultation on discovery modernisation, technology selection, and vendor evaluation.