Small Firm Discovery: Batch Extract Facts
No E-Discovery Required

Legal discovery — the pre-trial phase where parties exchange relevant documents — routinely produces thousands of pages of emails, contracts, deposition transcripts, medical records, and financial statements. For a solo practitioner or small firm, the volume alone creates a structural problem: you can't review 50,000 pages in 30 days if you're also handling hearings, client meetings, and the rest of your caseload. The American Bar Association's 2024 Legal Technology Survey found that only 27% of solo attorneys use litigation support software, compared to 73% at firms with 100+ lawyers. Only 29% use AI-assisted search. The gap isn't reluctance — it's that most e-discovery platforms were built for BigLaw budgets and BigLaw data volumes.

This article isn't about replacing e-discovery software. It's about adding a lightweight step before the review — batch-extracting key facts from discovery documents into a sortable spreadsheet, so you can decide what actually needs a close read without opening every file.

50,000 Pages, 30 Days: What Manual Review Actually Costs

The math is brutal. One gigabyte of documents contains roughly 5,000–7,500 files — and a single moderately sized commercial case can easily produce 20–50 GB. At a manual review pace of 50 documents per hour — a realistic rate once you account for dense contracts, handwritten margin notes, and deciphering whether two near-duplicate emails have different attachment chains — reviewing 20,000 documents takes approximately 400 attorney hours. That's 10 weeks of full-time work for one person. Your 30-day deadline gives you roughly 4 weeks.

The numbers that float around e-discovery pricing discussions are instructive. In 2012, RAND estimated the all-in cost of processing one gigabyte of data through full review at $18,000. Cloud platforms have driven processing costs down to roughly $25 per 100 GB in 2025, but that's just the hosting. The attorney review — the hours spent reading — still dominates the bill. The Association of Certified E-Discovery Specialists (ACEDS) estimates review consuming 64% of total e-discovery spend. Even when technology-assisted review (TAR) cuts the reviewable set by 70–80%, you still need to know what's in that set before you train the model.

This is where the batch extraction approach earns its place — not as a replacement for review, but as the step that tells you where review is actually needed.

What the FRCP Actually Requires — and What It Doesn't

A common hesitation from small-firm practitioners: "Am I allowed to run discovery documents through a third-party AI tool?" The Federal Rules of Civil Procedure provide a framework that's more flexible than most lawyers assume.

FRCP Rule 26(b)(1) limits discovery to matters "proportional to the needs of the case" — considering the amount in controversy, the parties' resources, and whether the burden or expense of the proposed discovery outweighs its likely benefit. The 2015 amendments elevated proportionality from a secondary consideration to the primary scope limitation. For a small firm handling a $75,000 dispute, a $30,000 e-discovery vendor bill is disproportionate by definition.

FRCP Rule 34 governs the production of documents — what the responding party must provide, in what form, and under what timeline. It does not prescribe what tools the receiving party may use to review those documents. Nothing in the Federal Rules prohibits a receiving party from using AI to extract structured data from produced documents, provided privilege is maintained and the tool's use doesn't alter or disseminate the underlying files. The same way a paralegal can create an index of documents by reading each one, AI can batch-read and extract fields into a spreadsheet.

The Electronic Discovery Reference Model (EDRM), the standard framework for the e-discovery lifecycle, places Review as its sixth of nine stages — after Information Governance, Identification, Preservation, Collection, and Processing. What the framework doesn't address is what you do when half of those stages are handled by the producing party and your real question is simpler: "Which of these 20,000 documents contain dollar figures, dates, and named parties relevant to my case theory?"

The answer, absent a full e-discovery platform, has historically been: start reading. But there's now a middle ground between full manual review and a $10,000 platform subscription.

The Batch Extraction Workflow: From Pile of Files to Sortable Fact Sheet

Here's the core idea: before you review any document for privilege or relevance, you do a fast, shallow extraction pass — pulling out the structural metadata and key data points that let you triage the entire production. You're not replacing review. You're building a map so you know which territory to explore first.

The mechanism that makes this possible is column-name extraction: instead of defining where data sits on each page — which would require a separate template for every document format in a heterogeneous production — you define what fields you want to extract. You type the column names once: "Date," "Sender," "Recipient," "Dollar Amount," "Document Type," "Key Parties Mentioned," "Case Reference Number." The AI reads each document, locates the relevant values wherever they appear, and populates a row in your spreadsheet. The output is one table where each row is a document and each column is a field you requested — regardless of whether the source was a PDF email, a scanned contract, or a screenshot of a handwritten deposition note.

The workflow has four steps, and the third is where most first-time users stumble:

Step 1 — Upload everything at once. Drag all produced files into the upload area. PDFs, JPGs, PNGs — all supported. A batch can handle dozens of files simultaneously. The tool doesn't care whether document #14 is a clean email PDF and document #15 is a crooked phone photo of a handwritten note — it reads each one independently and populates the same columns.

Step 2 — Define your column names in a single pass. This is the high-leverage decision. Too few columns and you miss data you'll need later. Too many and the extraction takes longer without adding triage value. Start with 6–8 fields that let you sort, filter, and prioritize: date, document type, parties, monetary figures, and 2–3 case-specific fields (case number, project name, relevant keywords). You can always run a second extraction pass with deeper columns on the subset of documents flagged for full review.

Step 3 — Download the merged spreadsheet and triage. The output is one Excel file where each row corresponds to one uploaded file. Sort by "Document Type" to separate contracts from emails. Filter "Amount" to find everything referencing dollar figures above your threshold of interest. Scan the "Parties Mentioned" column to identify which custodians appear most frequently — that tells you whose documents to prioritize in the deep-review pass. This is the step that replaces the first 100 hours of manual skimming.

Step 4 — Flag, annotate, and route. Add your own columns to the spreadsheet: "Review Priority (High/Medium/Low)," "Notes," "Assigned To." Now you have a single document index that the entire case team can reference, with each document traceable back to its source file by filename.

When Batch Extraction Works — and When You Need a Full E-Discovery Platform

This approach has a clear use case, and being honest about its limits is more useful than pretending it solves everything.

For small firms handling most civil litigation, the scenarios in the left column describe the majority of cases. A $50,000 breach of contract claim doesn't justify a $10,000 e-discovery platform subscription. A $500,000 employment dispute might. The batch extraction approach lets you allocate resources proportionally — using lightweight extraction for the cases that don't need heavyweight infrastructure, and knowing which cases actually warrant a platform before you spend the money.

Platforms like Nextpoint and GoldFynch have made full e-discovery accessible to smaller firms with flat-rate and per-case pricing models — GoldFynch starts at $27/month for a 3 GB case. These are excellent tools for cases that genuinely need the full EDRM pipeline. The point isn't that e-discovery platforms are bad. The point is that not every case needs the full pipeline — and having a lighter option for the ones that don't changes how you approach discovery intake across your entire caseload.

What This Looks Like on a Real Employment Discrimination Case

Consider a three-attorney plaintiffs' firm handling a wrongful termination case. The defendant — a mid-size company — produces 12,000 pages of discovery: HR emails, performance reviews, payroll records, internal Slack messages exported as PDFs, the plaintiff's personnel file, and deposition transcripts from three witnesses.

The firm's paralegal uploads all 12,000 pages as a batch — roughly 200 individual files. She defines eight column names: "Date," "Document Type," "Sender/Author," "Recipient/Audience," "Topic/Subject," "Key People Mentioned," "Performance Rating (if any)," and "Termination Reference (Yes/No)." The extraction completes while she moves on to another task. The output is a 200-row spreadsheet.

She sorts by "Termination Reference" and immediately isolates the 37 documents that directly mention the termination decision. She sorts by "Performance Rating" and finds that the plaintiff's last three reviews all rated "Meets Expectations" or above — contradicting the employer's stated reason for termination. She filters "Key People Mentioned" for the name of the supervisor who made the termination decision and finds 52 documents where that supervisor appears — including emails discussing the plaintiff's medical leave request, filed two weeks before termination.

The lead attorney now opens her document review with a curated priority list: the 37 termination-related documents first, the 52 supervisor-related documents second, and the performance reviews third. Instead of reading 12,000 pages cold, she's reading 200 pages of the highest-probability evidence — and the other 11,800 pages are indexed and retrievable if needed. The triage spreadsheet took 3 hours to create. It saved roughly 30 hours of attorney skimming.

Three Things Small Firms Get Wrong the First Time

The extraction part is straightforward. The batch logistics are where first attempts typically stumble.

1. File naming matters more than you think. The filename is the only field that connects a row in your spreadsheet back to the source document. If the produced files are named "0001.pdf," "0002.pdf," "0003.pdf" — a common Bates-stamped production convention — preserve those numbers. If they're named inconsistently ("Agreement_FINAL.pdf" vs. "Smith_contract_signed_2023.pdf"), spend 5 minutes applying a consistent convention before upload. Counterparty_DocType_Date.pdf is hard to beat for legal documents. A spreadsheet row that says "Document 0047 — Sender: John Smith, Date: March 2024 — Termination mentioned" is actionable. A row that says "IMG_4829.jpg" with the same data is a headache.

2. Over-specifying columns on the first pass. The instinct is to extract everything: every party, every date, every dollar amount, every clause, every reference. Resist it. A first-pass extraction with 6–8 columns covering the triage essentials (who, what, when, how much) runs in minutes and gives you a sortable index. A 25-column extraction covering every conceivable data point takes longer and produces a spreadsheet where the signal is buried in noise. Run the deep extraction on the documents you've already flagged as high-priority.

3. Treating the extraction output as a finished work product. The spreadsheet is a triage tool, not a filing. It tells you where to look and what you're likely to find. It doesn't replace the attorney's judgment about relevance, privilege, or strategy. Use it as the starting point for your review process — not the ending point. The most efficient small firms we've seen use the extraction spreadsheet as their case team's shared reference document: paralegals maintain it, attorneys annotate it, and everyone uses it to answer the question "what do we have and where is it?" without opening 200 PDFs.

Frequently Asked Questions

Does AI extraction work with scanned documents and handwritten notes?

Yes. The visual AI model reads scanned PDFs the way a human would — by interpreting the image, not by running text-layer OCR. This means it handles crooked scans, low-resolution faxes, and handwritten margin notes. The accuracy on handwriting is lower than on typed text (expect ~85–95% vs. 99% for clean print), which is why the batch extraction approach relies on sortability, not perfection. A handwritten note where you can read "Smith offered $45K settlement March 14" is useful even if the AI misreads the dollar amount by one digit — you now know which document to pull for manual verification.

Is this method defensible if challenged by opposing counsel?

The extraction spreadsheet is an internal work product — a case team's organizational tool, analogous to a paralegal-prepared index or a handwritten summary table. It is not a substitute for a privilege log, a production, or any filing made to the court. If the extraction is used solely for internal triage and the underlying documents are reviewed by an attorney before any substantive reliance, the method is no different in principle from a paralegal preparing a document index by hand — just significantly faster. For cases where discoverability of the extraction itself is a concern, consult your jurisdiction's work product doctrine under FRCP 26(b)(3) or the state equivalent.

How long does batch extraction take for a typical production?

Processing speed depends on document count and complexity, not page count. A batch of 200–300 files (each potentially multi-page) with 6–8 column names typically processes in 15–30 minutes. The bottleneck is not the AI; it's the upload speed for large files. For productions exceeding 500 files, consider splitting into multiple batches — each batch produces its own spreadsheet, and you can merge them in Excel with a simple copy-paste of rows.

Does this work with native file formats like .msg, .docx, or .xlsx?

Currently, the tool accepts PDF, JPG, PNG, WebP, and AVIF files. Native email formats (.msg, .eml), Word documents (.docx), and spreadsheets (.xlsx) must be converted to PDF before upload. Most e-discovery productions are already delivered as PDFs or TIFFs, so this is rarely a practical limitation for discovery review. If you receive native files, most practice management tools (Clio, MyCase, PracticePanther) can batch-convert them to PDF.

Can I use this for privilege review?

No. The extraction tells you what's in each document — dates, parties, amounts, topics — but it does not evaluate privilege. Attorney-client privilege and work product determinations require reading the substance of communications and applying legal judgment. What batch extraction does is tell you which documents are likely to contain privileged material (based on sender-recipient patterns, subject lines, or document type) so you can focus your privilege review on those documents, rather than reading 12,000 pages hunting for the 200 that might be privileged.

The Real Difference: You Decide What to Review, Not What to Skip

The manual review model forces a backwards decision: to avoid missing something, you read everything. When the production exceeds your capacity, you start skipping — sampling, skimming, making educated guesses about which documents don't matter. The risk isn't in what you read. It's in what you didn't read because you ran out of time.

Batch extraction inverts that model. The AI reads everything. It doesn't make relevance decisions — that's still your job. But it gives you a complete map of the production before you decide where to spend your time. You're not skipping documents you haven't seen. You're prioritizing the ones your triage system flagged as materially relevant, and you can prove — to yourself, to your client, to the court if necessary — that every document in the production was at minimum indexed and accounted for.

For a small firm where every hour of attorney time is a direct cost to the client or the practice, that inversion is not a convenience. It's a structural improvement to how discovery gets done.