AI legal document review is useful when it helps a lawyer find and verify relevant material across a defined document set. A fluent summary without the source passage, document identity and review status creates another assertion for the lawyer to investigate.
This guide proposes a review workflow for commercial law firms. It is not legal advice and does not claim Binarify client results. The responsible lawyers must define the document population, legal question, jurisdiction, privilege and confidentiality controls, review standard and final use. See our AI consultancy for law firms page for the wider implementation approach.
Define the review product
“Review these documents” is too broad for a reliable specification. The expected product may be:
- a clause table across agreements;
- a chronology from correspondence and records;
- a list of change-of-control provisions;
- a comparison against an approved playbook;
- a factual evidence index for a dispute;
- a list of documents responsive to an approved question; or
- a preparation memo for lawyer review.
Define the question, eligible sources, required fields, treatment of missing information and reviewer before processing begins.
A bounded outcome could be:
Produce a review matrix for an approved document set and question list, with each proposed answer linked to the exact source passage and every unsupported, conflicting or ambiguous result sent to a lawyer.
Inspect native legal-AI capabilities first
Legal document and research platforms already provide substantial review functions. Ask iManage supports document questions, extraction, chronologies, multi-document tables and comparisons inside iManage Work. Its current documentation describes evidence links and repository-search controls. Thomson Reuters CoCounsel Review Documents returns an answer for each document and question, with supporting analysis and source navigation.
Use those capabilities when they meet the firm’s workflow, security, jurisdiction and licensing requirements. Integration work may still be valuable for matter setup, approved question lists, exception routing, reviewer decisions and writing validated results back to the correct matter. Custom document AI should not reproduce a licensed tool without a clear operational reason.
A source-linked document-review workflow
1. Approve the matter and source population
Identify the client, matter, review purpose, jurisdiction, custodians, repositories, date range, document types and exclusions. Preserve the original collection and its identifiers.
Confirm that the team may process the material with the selected product and in the selected location. Apply matter permissions, ethical walls, legal hold requirements and client instructions before AI access begins.
Do not let semantic search quietly widen the population beyond the approved workspace. If the product can search attached documents, the matter repository or wider firm knowledge, record which mode was used for each result.
2. Prepare the files without losing provenance
Record each document’s stable ID, filename, version, location, hash where appropriate, author, date and parent-child relationship. Extract text with page or paragraph coordinates so an answer can return to the source.
Handle scans, handwritten material, spreadsheets, tracked changes, email threads, attachments, duplicates and password-protected files explicitly. An unreadable page is an exception, not an empty answer.
Use deterministic processing for dates, file identity, counts and duplicates where possible. AI interpretation should not replace reliable metadata.
3. Convert the review protocol into questions
Create a question list with the responsible lawyer. Each question should define:
- the fact, clause or issue sought;
- acceptable answer formats;
- relevant context and defined terms;
- whether absence can be reported;
- the evidence required;
- escalation rules; and
- the intended downstream use.
“Is this agreement risky?” is not a stable review question. “What liability cap does the agreement state, what exclusions apply, and where is each term located?” is more reviewable. The lawyer still decides significance against the transaction, advice and applicable law.
4. Produce one answer per document and question
Return structured fields such as present, absent, unclear, conflicting or not readable. Attach the exact passage, page and document version to every substantive answer.
Keep extraction separate from interpretation. For example:
- Extraction: the agreement states a 12-month fee cap in clause 14.2.
- Comparison: the approved playbook position is a 24-month fee cap.
- Legal or commercial assessment: counsel decides the effect, negotiation position and advice.
If the answer relies on more than one passage, show each passage. If the source does not support the conclusion, mark it unsupported rather than filling the gap from general model knowledge.
5. Build chronologies from events, not prose alone
For chronology work, create separate records for date, event, participants, source, date certainty and reviewer status. Distinguish:
- the date an event occurred;
- the date a document was created;
- the date it was sent or filed; and
- a date mentioned retrospectively in the text.
Keep conflicting accounts side by side. Do not merge them into one confident narrative. A lawyer determines relevance, credibility and the meaning of silence or inconsistency.
6. Route exceptions by consequence
Send results to lawyer review when:
- a source is unreadable, incomplete or outside scope;
- defined terms change the apparent meaning;
- several provisions interact;
- documents conflict;
- the answer depends on law or authority outside the matter set;
- privilege, confidentiality or sensitivity needs a decision;
- the model supplies an uncited statement; or
- the output could materially affect advice, negotiation, filing or disclosure.
Priority should follow the firm’s approved protocol, not model confidence alone. Confidence scores are not substitutes for evidence or legal judgement.
7. Record the lawyer’s disposition
The reviewer should be able to accept, correct, reject, request more evidence or mark the question inapplicable. Capture the correction reason without forcing the lawyer to rewrite the entire row.
Common correction categories include wrong source, missed clause, incomplete context, incorrect entity, incorrect date, unsupported inference, outdated authority, privilege concern and question design. These categories show whether to improve the source population, extraction, question list or user training.
8. Export only reviewed work
Write the approved matrix, chronology or summary to the matter workspace with its source links, reviewer, time and version. Label working material clearly. Prevent an unreviewed AI draft from being mistaken for approved advice or filed work.
When results populate another document, retain the route back to the review record. A copied sentence without its evidence and status loses the main control that made it useful.
Binarify source-linked review matrix
This proposed artifact connects every answer to its evidence and final disposition.
| Review question | Proposed answer | Source evidence | Automated status | Lawyer disposition | Final use |
|---|---|---|---|---|---|
| Liability cap | Extracted amount and basis | Document, version, clause and passage | Supported, unclear or conflicting | Accept, correct or escalate | Due-diligence table |
| Term and renewal | Dates, notice and renewal rule | Relevant clauses and definitions | Complete or missing context | Confirm interpretation | Contract summary |
| Event date | Date, event and participants | Email, record or exhibit passage | Certain, approximate or conflicting | Confirm chronology entry | Case chronology |
| Governing law | Extracted jurisdiction | Exact clause | Supported or absent | Confirm scope relevance | Matter index |
| Playbook variance | Source position versus approved standard | Source clause and playbook version | Match, variance or unassessed | Decide response | Negotiation preparation |
Keep the question-list version with the output. A later change to the protocol should not silently alter an earlier review.
Legal research is a separate evidence population
Matter documents establish what the selected records say. Legal authorities establish the law. Keep these populations separate even when one interface can search both.
For legal research, require licensed or otherwise approved authoritative sources, jurisdiction and date limits, links to the full text, treatment history where relevant and lawyer verification. Never turn an uncited model answer into a case, quotation or proposition.
The SRA’s warning notice on misuse of AI highlights inaccurate or fabricated cases, legislation, quotations and facts. The South African Legal Practice Council’s 2026 ICT and cyber-law guide says lawyers using AI for legal research must check accuracy against authoritative sources before using it professionally.
Measure quality before speed
Create a lawyer-reviewed reference set by workflow and document type. Track:
- review minutes per document or question, including setup and verification;
- material facts or clauses correctly identified;
- material omissions and unsupported answers;
- wrong-document, wrong-version and wrong-entity errors;
- lawyer correction and rejection rates by reason;
- source links that resolve to the supporting passage;
- documents sent to exception review;
- downstream corrections after the review product is used; and
- access, confidentiality, privilege and retention incidents.
Report performance by document type, language, scan quality, matter type and question. One aggregate accuracy percentage can hide a serious failure in the documents that matter most.
Pilot one repeatable review product
Choose a repeated task with a stable source population and question list, such as a bounded contract clause table or chronology preparation. Use synthetic, public or explicitly approved material during setup, then run shadow review on a representative controlled sample.
Agree the reference answers, double-review sample, materiality threshold, stop conditions and rollback owner. Compare the complete workflow with the existing process. A pilot has failed if it shortens extraction but increases lawyer verification, misses material items or breaks the matter evidence trail.
Begin earlier with AI client intake and conflict-check preparation. Use the law-firm AI ROI guide to connect measured workflow changes to the firm’s economics, or book a 30-minute conversation about one repeatable review product.