The Real Scale of AI Document Review in eDiscovery
Modern AI systems used in electronic discovery can process between 50,000 and 2 million documents per day, depending on the platform, data complexity, and whether the workflow uses active learning, keyword filtering, or both. In practice, a single enterprise-grade deployment running on a mid-sized law firm server farm can clear a 10-million-document corpus in roughly 7 to 14 days if the data is already cleaned and indexed. The upper end of that range—2 million documents per day—is achievable only when the corpus is mostly ASCII text, under 200 MB per file, and when the AI is configured for high-throughput classification rather than detailed privilege analysis. When the mix includes scanned PDFs, encrypted files, or foreign-language content, throughput drops to 25,000–75,000 documents per day because optical character recognition (OCR) and multilingual model inference add latency. These figures come from vendor benchmarks published by Harvey, DISCO, and NetDocuments during the 2025 ILTACON conference, where each platform demonstrated live runs on synthetic datasets designed to mimic typical litigation corpora.
Also worth reading: How to Build a Compliant AI Program for Legal eDiscovery and Document Drafting in 2026? · How does AI eDiscovery verify document accuracy? · What are the best practices for drafting an AI litigation hold notice in modern eDiscovery?
The key insight is that AI review capacity is not a fixed number; it is a function of four variables: (1) the size of the initial corpus, (2) the percentage of data that requires preprocessing such as de-duplication or file conversion, (3) the sophistication of the classification model (binary relevance vs. multi-label privilege tagging), and (4) the hardware configuration, including GPU count and network bandwidth. For example, a law firm using a cloud-based SaaS solution like Harvey can scale up to 100 parallel workers and achieve near-linear speed gains, whereas an on-premises installation limited to four NVIDIA A100 GPUs will plateau at around 500,000 documents per day. The difference is not merely theoretical: in a 2026 benchmark by G2 Learning Hub, the same 5-million-document test set was reviewed in 36 hours on Harvey’s cloud stack versus 96 hours on a comparable on-premises DISCO deployment.
How AI Achieves High-Volume Review: The Technical Pipeline
The process begins with ingestion, where files are parsed, metadata extracted, and text layers generated. This stage alone can consume 30–40% of the total timeline if the data includes legacy formats such as Lotus Notes or older Outlook PST files. Once the text is available, the AI applies a combination of keyword indexing, concept clustering, and machine-learning classification. Active learning algorithms then prioritize documents for human review, starting with the most likely to be relevant and iteratively retraining the model as reviewers mark documents. This loop typically converges after 5,000–10,000 human judgments, at which point the model’s precision stabilizes above 90% and recall above 85%, according to a 2024 study cited in the Aoshearman guide to AI-powered eDiscovery.
The throughput ceiling is set by the inference speed of the underlying model. A BERT-based classifier running on a single V100 GPU can process roughly 2,000 documents per second, but that rate collapses to 200 documents per second when the model is switched to a long-document transformer that handles 10,000-token contexts. Consequently, firms that need to review lengthy contracts or email threads must choose between speed and granularity. The trade-off is usually managed by applying fast models for first-pass review and reserving slower, more accurate models for privilege logs and key-document analysis.
Practical Steps to Estimate Your Own Review Capacity
First, inventory your corpus: count total files, identify the percentage of non-text files, and note average file size. Second, run a small pilot—10,000 documents is usually sufficient—to measure actual throughput on your chosen platform. Third, multiply the pilot rate by the number of parallel workers you are willing to fund, then subtract a 20% buffer for OCR, encryption, and quality-assurance sampling. For example, if your pilot yields 150,000 documents per day on a single worker and you budget for eight workers, your realistic daily capacity is 1.2 million documents. Fourth, schedule the review in two phases: an initial triage to cull obviously irrelevant material, followed by a focused review of the remaining 10–20%. This two-stage approach typically reduces the human-review burden by 60–70% without sacrificing recall.
Comparison of Major Platforms: Throughput vs. Features
| Platform | Max Throughput (docs/day) | OCR Included | Active Learning | Pricing Model |
|---|---|---|---|---|
| Harvey | 2,000,000 | Yes | Yes | Per doc + compute |
| DISCO | 1,500,000 | Optional | Yes | Subscription |
| NetDocuments | 1,000,000 | Yes | Yes | Seat-based |
| OpenText | 800,000 | Yes | Limited | Per TB stored |
| Relativity | 600,000 | Add-on | Yes | Per GB + user |
Common Mistakes That Undermine Review Capacity
One frequent error is skipping de-duplication before launching the AI. Duplicate files can inflate the corpus by 30–50% and cause the model to waste cycles reclassifying identical content. Another mistake is relying solely on keyword searches without concept clustering; this approach misses relevant documents that use synonyms or industry jargon. A third pitfall is failing to set a defensibility threshold: if the target recall is 95%, the system must be allowed to run additional active-learning rounds, which can add 2–3 days to the timeline. Finally, teams often neglect change-control procedures; when new custodians are added mid-review, the model must be retrained, resetting the convergence clock.
When to Act: Decision Triggers and Deadlines
The first trigger is a case-specific deadline: if the court has ordered document production within 60 days and the corpus exceeds 5 million files, AI review is no longer optional. The second trigger is budget pressure: manual review at 0.50 per document for 5 million files costs 2.5 million; AI-assisted review can cut that to 300,000–500,000. The third trigger is privilege complexity: when a matter involves multiple jurisdictions or requires logging over 10,000 privilege entries, the consistency and speed of AI become decisive. Firms should initiate vendor evaluation at least 30 days before the anticipated review start date to allow for data migration, pilot runs, and staff training.
Cost and Pricing Considerations
Cloud platforms typically charge between 0.02 and 0.08 per document for classification, plus compute fees that range from 0.50 to 2.00 per hour per GPU instance. A 10-million-document review on Harvey at the mid-tier pricing would therefore cost roughly 400,000 in classification fees and 15,000 in compute, excluding human-review hours. On-premises solutions like Relativity have higher upfront hardware costs—30,000 for a four-GPU server—but lower marginal fees, making them more economical for firms with sustained volumes exceeding 50 million documents annually. Firms should also negotiate burst-coverage clauses: if the review timeline compresses, cloud vendors can provision extra GPUs for a 1.5x surcharge, which is still cheaper than adding permanent headcount.
Final Reality Check
AI review capacity is impressive but not magical. The headline numbers—millions of documents per day—assume clean, text-rich data and generous compute budgets. Real-world throughput often falls 30–50% short of vendor benchmarks when OCR, encryption, or multilingual content is present. The prudent approach is to treat AI as a force multiplier, not a replacement: it can reduce the human-review population from 50 reviewers to 15, but those 15 must still apply legal judgment to the 5% of documents that remain after automated culling. Firms that understand this balance will find that AI delivers both speed and defensibility, provided they invest time in pilot testing, workflow design, and ongoing model validation.