Litigation Document Coding With AI: Objective vs Subjective

The Short Answer

Litigation document coding records structured information about every document in a collection, and it splits into two kinds: objective coding (what the document says about itself: date, author, recipients, document type, Bates range) and subjective coding (what a reviewer concludes about it: relevance, responsiveness, privilege, issue tags). Claude does dependable first-pass objective coding into a table, proposes subjective codes with a stated reason, and writes NOT FOUND rather than guessing when a field is not on the face of the document. Privilege calls stay with a lawyer. A QC sample catches the rest.

This post is written for solo and small-firm litigators handling document sets in the hundreds or low thousands, the range where a review platform is overkill but a spreadsheet plus a capable model is not. If you run large-scale e-discovery, your platform's tools and protocols govern and this post is background.

Objective vs. Subjective Coding: The Definitions

The distinction is standard in the e-discovery field, not something we invented. The Sedona Conference Glossary defines coding as capturing information from documents into a database and notes that coding may be objective (such as sender name or date) or subjective (an evaluation of relevance or probative value). EDRM's glossary carries separate entries for bibliographic or objective coding, which captures information visible on the document such as date, author, recipients, and document type, and for subjective coding, which records legal interpretations linked to individual documents.

The reason the line matters for AI use is simple. Objective fields have a right answer that can be checked against the document in seconds. Subjective fields are judgments that depend on the case, the issues, and the reviewer's legal training. A model is a strong fit for the first and a useful assistant for the second.

Objective coding fields

Field	What goes in it	Notes
Bates begin / end	Production numbers on the first and last page	Include family range if the document has attachments
Document date	Date on the face of the document	Use ISO format (2026-03-14); NOT FOUND if undated
Document type	Email, letter, memo, contract, invoice, spreadsheet, text message, photo, other	Use a fixed list
Author / from	Person or entity who created or sent it	Name as written; do not normalize without a lookup table
Recipients / to, cc, bcc	All addressees	Separate fields for to, cc, bcc
Title / subject	Subject line or document title verbatim
Page count	Number of pages
Attachments / family	Bates of parent or children
Custodian	Whose files it came from	Usually from the collection metadata, not the document

Subjective coding fields

Field	Values	Who decides
Responsiveness	Responsive / Not responsive / Needs review	AI proposes with reason; reviewer confirms
Relevance to issues	Issue tags from a fixed list (e.g. Training, Lighting, Notice, Damages)	AI proposes with reason; reviewer confirms
Privilege flag	Possible A-C / Possible work product / None apparent	AI flags only; lawyer decides
Confidentiality	Per protective order tiers	Reviewer
Hot / key document	Yes / No	Reviewer
Reviewer notes	Free text	Reviewer

How Claude Does First-Pass Coding

The workflow is the same whether you have forty documents or four hundred. Work in batches sized to what you can QC in one sitting.

1. Fix the field list and the value lists before you start. A document type field with a free-text answer produces "email," "e-mail," "Email message," and "correspondence" for the same thing. Give Claude the exact permitted values.
2. Give Claude the issue list and one sentence per issue. Issue tagging is only as good as the definitions. "Notice: any document showing Meridian knew of the lighting problem before the incident" is a workable definition; "Notice" alone is not.
3. Attach the batch and paste the extraction prompt. Ask for a table with one row per document and the Bates range as the key. For batches too large to attach at once, run in sub-batches and keep the same prompt text so the coding is consistent.
4. Enforce NOT FOUND discipline. The single most important instruction: if a field is not on the face of the document, the cell says NOT FOUND, never a guess. An inferred date looks identical to a real one in a spreadsheet and is far more dangerous.
5. Require a reason for every subjective code. "Responsive: discusses bay 6 lighting repair request, Issue: Notice" can be checked in five seconds. A bare "Responsive" cannot.
6. Route privilege flags to a lawyer's queue. Every row flagged Possible A-C or Possible work product goes to lawyer review before anything else happens to it.
7. QC sample, then load. Described below. Only after QC does the table go into your working index or review tool.

Model choice matters here more than for most tasks. Objective coding is a high-volume, low-difficulty task suited to a faster model; issue tagging across a nuanced issue list benefits from a stronger one. Our breakdown of which Claude model to use for legal work covers the trade-off. For a broader set of litigation workflows, including chronology building that consumes coded documents, see Claude for litigation practices.

Why Privilege Calls Stay Human

Under Federal Rule of Civil Procedure 26(b)(5)(A), a party withholding otherwise discoverable material as privileged or as trial-preparation material must expressly make the claim and describe what was withheld in a manner that lets other parties assess the claim without revealing the protected content. That description is a lawyer's representation to the court and opposing counsel. A model is useful for surfacing candidates (attorney names in the header, phrases like "legal advice" or "litigation hold," communications with outside counsel) and for drafting the neutral description once a lawyer has made the call. The determination itself, and the consequences of getting it wrong (waiver, clawback fights, sanctions motions), belong to the lawyer. State rules of civil procedure impose parallel requirements, and the professional-responsibility overlay is addressed in our guide to the ABA guidance on lawyer use of AI.

A practical rule that small firms can actually follow: the AI never sets the privilege field to a final value. It sets a flag. A lawyer sets the value.

QC Sampling

You do not need a statistician to QC a first-pass coding run, but you need a routine you follow every time:

• Full review, no sampling: every document flagged for possible privilege; every document with NOT FOUND in a key field (date, author, type); every document tagged Hot.
• Random sample of the rest: pick a fixed fraction you can sustain (one in ten is a common starting point for a small set) and check every field against the document. Record errors by field.
• Act on the error pattern, not the count. If date errors cluster on scanned letters, the fix is better OCR or a rule about which date to use, not more sampling. If issue tags are wrong on a specific issue, the fix is a better issue definition. Re-instruct and re-run the affected batch.
• Escalate thresholds. Decide in advance the error rate at which you stop trusting a batch and re-run it entirely. Write the number down before you see the results.
• Keep the QC log with the coding table. If the coding is ever challenged, the log is your evidence of a defensible process.

The Extraction Prompt

You are assisting a litigation team with first-pass document coding. Attached are [N] documents from a production in [MATTER]. We represent [PARTY].

Produce a table with one row per document and these columns, in this order:
Bates begin | Bates end | Document date (YYYY-MM-DD) | Document type | Author/From | To | CC | BCC | Subject/Title | Page count | Attachments (Bates) | Responsiveness | Issue tags | Privilege flag | Reason

Permitted values:
- Document type: Email, Letter, Memo, Contract, Invoice, Spreadsheet, Text message, Photo, Report, Other.
- Responsiveness: Responsive, Not responsive, Needs review.
- Issue tags (choose all that apply): [ISSUE 1: one-sentence definition]; [ISSUE 2: definition]; [ISSUE 3: definition].
- Privilege flag: Possible A-C, Possible work product, None apparent.

Rules:
1. Objective columns come only from the face of the document. If a value is not present, write NOT FOUND. Never infer a date, author, or recipient from context.
2. Copy names and subject lines exactly as written. Do not normalize spelling.
3. For Responsiveness and Issue tags, the Reason column must cite the specific passage or feature that supports the code, in under 25 words.
4. Privilege flag is a flag for lawyer review, not a determination. Flag if any of: an attorney or law firm appears as author or recipient; the text requests or conveys legal advice; the document references litigation hold, counsel, or attorney work product. State the trigger in the Reason column.
5. If a document is unreadable in whole or part, code what is legible and add UNREADABLE to the Reason column.
6. After the table, list any documents that appear to be duplicates of each other by Bates range.

If your matter also requires pulling dates into a calendar, the same discipline applies and we cover it in extract and calendar deadlines. Coded documents feed directly into a chronology; see build a case chronology.